Let’s dive into how machine learning methods can be used for the classification and forecasting of time series problems with Python.
But first let’s go back and appreciate the classics, where we will delve into a suite of classical methods for time series forecasting that you can test on your forecasting problem prior to exploring machine learning methods.
While traditional methods have an emphasis on linear relationships, they are adept and effective for many scenarios. As long as the data is well-prepped and the technique is set up, you are good.
This article has been put together to be your perfect cheat sheet, offering bite-sized insights into each method alongside a ready-to-use Python code snippet, aswell as further pointers on where to dive deeper.
All our code examples harness the power of the Statsmodels library in Python. The APIs for this library can be tricky for beginners (trust me!), so having a working code example as a starting point will greatly accelerate your progress.
This is a large post; you may want to bookmark it.
Want a deeper dive? Check out: Time Series Forecasting With Python, which is packed with step-by-step tutorials and all the Python source code you will need.
Let’s get started.
Kick-start your project with my new book Time Series Forecasting With Python, including step-by-step tutorials and the Python source code files for all examples.
Let’s get started.
Updated Apr/2020: Changed AR to AutoReg due to API change.
Updated Dec/2020: Updated ARIMA API to the latest version of statsmodels.
Updated Nov/2023: Updated for readability and clarification.
11 Classical Time Series Forecasting Methods in Python (Cheat Sheet) Photo by Ron Reiring, some rights reserved.
Overview
This cheat sheet demonstrates 11 different classical time series forecasting methods; these are:
Seasonal Autoregressive Integrated Moving-Average with Exogenous Regressors (SARIMAX)
Vector Autoregression (VAR)
Vector Autoregression Moving-Average (VARMA)
Vector Autoregression Moving-Average with Exogenous Regressors (VARMAX)
Simple Exponential Smoothing (SES)
Holt Winter’s Exponential Smoothing (HWES)
Did I miss your favourite classical time series forecasting method?
Let me know in the comments below.
Each method is presented in a consistent manner.
This includes:
Description. A short and precise description of the technique.
Python Code. A short working example of fitting the model and making a prediction in Python.
More Information. References for the API and the algorithm.
For each code example provided, we utilise a basic illustrative dataset. We understand that it may not always be suitable for the method in focus, therefore we advise you to replace the contrived dataset with your data in order to test the method.
Remember: Tailoring each method will require tuning to your specific problem. In many cases, I have examples of how to configure and even grid search parameters on the blog already, try the search function.
If this guide proves beneficial, please share your thoughts in the comments section below.
Autoregression (AR)
The autoregression (AR) method predicts the subsequent value in a sequence using a linear combination of previous observations.
The notation for the model involves specifying the order of the model p as a parameter to the AR function, e.g. AR(p). For example, AR(1) is a first-order autoregression model.
The method is best suited for single-variable time series that lack trend and seasonal components.
The Moving Average (MA) method models predict the next step in the sequence as a linear function of the residual errors from a mean process at prior time steps.
It’s important to note that a Moving Average model is different from calculating the moving average of the time series.
The notation for the model involves specifying the order of the model q as a parameter to the MA function, e.g. MA(q). For example, MA(1) is a first-order moving average model.
The method is suitable for univariate time series without trend and seasonal components.
We can use the ARIMA class to create an MA model and set a zeroth-order AR model. We must specify the order of the MA model in the order argument.
Python Code
We can use the ARIMA class to create an MA model and setting a zeroth-order AR model. We must specify the order of the MA model in the order argument.
The Autoregressive Moving Average (ARMA) method model predicts the next step in the sequence based on a linear combination of both past observations and past residual errors.
The method combines both Autoregression (AR) and Moving Average (MA) models.
To represent the model, the notation involves specifying the order for the AR(p) and MA(q) models as parameters to an ARMA function, e.g. ARMA(p, q). An ARIMA model can be used to develop AR or MA models.
The method is suitable for univariate time series without trend and seasonal components.
The Autoregressive Integrated Moving Average (ARIMA) method model predicts the next step in the sequence as a linear function of the differenced observations and residual errors at prior time steps.
The method integrates the principles of Autoregression (AR) and Moving Average (MA) models as well as a differencing pre-processing step of the sequence to make the sequence stationary, called integration (I).
The notation for the model involves specifying the order for the AR(p), I(d), and MA(q) models as parameters to an ARIMA function, e.g. ARIMA(p, d, q). An ARIMA model can also be used to develop AR, MA, and ARMA models.
The ARIMA approach is optimal for single-variable time series that exhibit a trend but lack seasonal variations.
The Seasonal Autoregressive Integrated Moving Average (SARIMA) method models the next step in the sequence based on a linear blend of differenced observations, errors, differenced seasonal observations, and seasonal errors at prior time steps.
SARIMA enhances the ARIMA model with the ability to perform the same autoregression, differencing, and moving average modeling at the seasonal level.
The notation for the model involves specifying the order for the AR(p), I(d), and MA(q) models as parameters to an ARIMA function and AR(P), I(D), MA(Q) and m parameters at the seasonal level, e.g. SARIMA(p, d, q)(P, D, Q)m where “m” is the number of time steps in each season (the seasonal period). A SARIMA model can be used to develop AR, MA, ARMA and ARIMA models.
The method is suitable for univariate time series with trend and/or seasonal components.
Python Code
1
2
3
4
5
6
7
8
9
10
11
# SARIMA example
from statsmodels.tsa.statespace.sarimax import SARIMAX
Seasonal Autoregressive Integrated Moving-Average with Exogenous Regressors (SARIMAX)
The Seasonal Autoregressive Integrated Moving-Average with Exogenous Regressors (SARIMAX) is an extension of the SARIMA model that also includes the modeling of exogenous variables.
Exogenous variables are also called covariates and can be thought of as parallel input sequences that have observations at the same time steps as the original series. The primary series may be referred to as endogenous data to contrast it from the exogenous sequence(s). The observations for exogenous variables are included in the model directly at each time step and are not modeled in the same way as the primary endogenous sequence (e.g. as an AR, MA, etc. process).
The SARIMAX method can also be used to model the subsumed models with exogenous variables, such as ARX, MAX, ARMAX, and ARIMAX.
The method is suitable for univariate time series with trend and/or seasonal components and exogenous variables.
Python Code
1
2
3
4
5
6
7
8
9
10
11
12
13
# SARIMAX example
from statsmodels.tsa.statespace.sarimax import SARIMAX
The Vector Autoregression (VAR) method models the next step in each time series using an AR model approach. Essentially, it extends the AR model to cater to multiple parallel time series, e.g. multivariate time series.
The notation for the model involves specifying the order for the AR(p) model as parameters to a VAR function, e.g. VAR(p).
The method is suitable for multivariate time series without trend and seasonal components.
Python Code
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# VAR example
from statsmodels.tsa.vector_ar.var_model import VAR
The Vector Autoregression Moving-Average (VARMA) method models the upcoming value in multiple time series by utilising the ARMA model approach. It is the generalization of ARMA to multiple parallel time series, e.g. multivariate time series.
The notation for the model involves specifying the order for the AR(p) and MA(q) models as parameters to a VARMA function, e.g. VARMA(p, q). A VARMA model can also be used to develop VAR or VMA models.
The method is suitable for multivariate time series without trend and seasonal components.
Python Code
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# VARMA example
from statsmodels.tsa.statespace.varmax import VARMAX
Vector Autoregression Moving-Average with Exogenous Regressors (VARMAX)
The Vector Autoregression Moving-Average with Exogenous Regressors (VARMAX) extends the capabilities of the VARMA model which also includes the modelling of exogenous variables. It is a multivariate version of the ARMAX method.
Exogenous variables, also called covariates and can be thought of as parallel input sequences that align with the time steps as the original series. The primary series(es) are referred to as endogenous data to contrast it from the exogenous sequence(s). The observations for exogenous variables are included in the model directly at each time step and are not modeled in the same way as the primary endogenous sequence (e.g. as an AR, MA, etc. process).
The VARMAX method can also be used to model the subsumed models with exogenous variables, such as VARX and VMAX.
The method is suitable for multivariate time series without trend and seasonal components with exogenous variables.
Python Code
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
# VARMAX example
from statsmodels.tsa.statespace.varmax import VARMAX
The Simple Exponential Smoothing (SES) method models the next time step as an exponentially weighted linear function of observations at prior time steps.
The method is suitable for univariate time series without trend and seasonal components.
Python Code
1
2
3
4
5
6
7
8
9
10
11
# SES example
from statsmodels.tsa.holtwinters import SimpleExpSmoothing
The Holt Winter’s Exponential Smoothing (HWES) also called the Triple Exponential Smoothing method models the next time step as an exponentially weighted linear function of observations at prior time steps, taking trends and seasonality into account.
The method is suitable for univariate time series with trend and/or seasonal components.
Python Code
1
2
3
4
5
6
7
8
9
10
11
# HWES example
from statsmodels.tsa.holtwinters import ExponentialSmoothing
In this post, you discovered a suite of classical time series forecasting methods that you can test and tune on your time series dataset. These methods are designed for a wide range of time series datasets, allowing you to implement them across various scenarios and industries. Whether you are dealing with stock market trends, weather forecasts, or sales projections, these methods are equipped to offer valuable predictions.
Perhaps there’s a classical time series forecasting method that you’ve found particularly effective and it wasn’t mentioned in this post. If so, I would love to hear about it. Let me know in the comments below.
I’m interested in hearing about your hands-on experiences with these methods. Have you applied any of the techniques discussed in the post to your own datasets? Let me know about your findings in the comments.
Diving into time series forecasting can come with its fair share of questions and uncertainties. If you have any questions, ask them in the comments below and I will do my best to answer.
Remember, your participation and input are what make this community thrive. So, don’t hesitate to share your thoughts, experiences, and questions. Let’s continue to learn and grow together in the fascinating world of time series forecasting!
Want to Develop Time Series Forecasts with Python?
Hi Jason, thanks for such an excellent and comprehensive post on time series. I sincerely appreciate your effort. As you ask for the further topic, just wondering if I can request you for a specific topic I have been struggling to get an output. It’s about Structural Dynamic Factor model ( SDFM) by Barigozzi, M., Conti, A., and Luciani, M. (Do euro area countries respond asymmetrically to the common monetary policy) and Mario Forni Luca Gambetti (The Dynamic Effects of Monetary Policy: A Structural Factor Model Approach). Would it be possible for you to go over and estimate these two models using Python or R? It’s just a request from me and sorry if it doesn’t go with your interest.
Hi Jason,
Great technical insights. I work in the oil and gas sector and one of our day to day job is to to rate vs time oil rate forecasting. We typically use empirical equations and use regression to fit those empirical input parameters. I was wondering if we could use sophisticated ML techniques to accomplish this?
Thanks!
#Abhishek I work in oil&gas company too, and I tried to generate forecasts using ML. the problem is that you can’t relay only on prodaction rates, some parameters should be included in the model such as WHFP, BHFP, SEP P, WC… for all wells to accurately estimate the total production. An other problem is that even you came somehow to collect those data you cannot relay on them because generally those data are daily averages and dont reflect the variations in real time. Also, prodaction rates for each well are genarelly calculated using allocation method.
I am working on Time series or Prediction with neural network and SVR, I want to this in matlab by scratch can you give me the references of this materials
Thank you in advance
Hi Jason! From which editor do you import the python code into the webpage of your article? Or what kind of container it that windowed control used to display the python code?
I recently stumbled over some tasks where the classic algorithms like linear regression or decision trees outperformed even sophisticated NNs. Especially when boosted or averaged out with each other.
Maybe its time to try the same with time series forecasting as I’m not getting good results for some tasks with an LSTM.
The focus is deep learning (MLP, CNN and LSTM) with tutorials on how to get the most from classical methods (Naive, SARIMA, ETS) before jumping into deep learning methods. I hope to have it done by the end of the month.
Well, although it is a really helpful and useful book as it is usually made by Jason, this book does not cover multivariate time series problems, in fact Jason explicitely says “This is not a treatment of complex time series problems. It does
not provide tutorials on advanced topics like multi-step sequence forecasts, multivariate time series problems or spatial-temporal prediction problems.”
Hi Arun,
Can you let me know how you worked with fbprophet. I am struggling with the installation of fbprophet module. Since it’s asking for c++ complier. Can you please share how you installed the c++ complier. I tried all ways to resolve it.
Alberto Garcia GalindoAugust 11, 2018 at 12:14 am#
Hi Jason!
Firstly I congratulate you for your blog. It is helping me a lot in my final work on my bachelor’s degree in Statistics!
What are the assumptions for make forecasting on time series using Machine Learning algorithms? For example, it must to be stationary? Thanks!
Hi Jason
I’m interested in forecasting the temperatures
I’m provided with the previous data of the temperature
Can you suggest me the procedure I should follow in order to solve this problem
Cool stuff as always. Kudos to you for making me a ML genius!
Real quick:
How would you combine VARMAX with an SVR in python?
Elaboration.
Right now I am trying to predict a y-value, and have x1…xn variables.
The tricky part is, the rows are grouped.
So, for example.
If the goal is to predict the price of a certain car in the 8th year, and I have data for 1200 cars, and for each car I have x11_xnm –> y1_xm data (meaning that let’s say car_X has data until m=10 years and car_X2 has data until m=3 years, for example).
First I divide the data with the 80/20 split, trainset/testset, here the first challenge arises. How to make the split?? I chose to split the data based on the car name, then for each car I gathered the data for year 1 to m. (If this approach is wrong, please tell me) The motivation behind this, is that the 80/20 could otherwise end up with data of all the cars of which some would have all the years and others would have none of the years. aka a very skewed distribution.
Then I create a model using an SVR, with some parameters.
And then I try to predict the y-values of a certain car. (value in year m)
However, I do not feel as if I am using the time in my prediction. Therefore, I turned to VARMAX.
Final question(s).
How do you make a time series prediction if you have multiple groups [in this case 1200 cars, each of which have a variable number of years(rows)] to make the model from?
Am I doing right by using the VARMAX or could you tell me a better approach?
Sorry for the long question and thank you for your patience!
Excellent post! I also would like to invite you to know the Fuzzy Time Series, which are data driven, scalable and interpretable methods to analyze and forecast time series data. I have recently published a python library for that on http://petroniocandido.github.io/pyFTS/ .
Hi Jason,
I have multivariate binary time series. If a time series is 0 it shows us the specific transmission line in power system isn’t failed, as soon as it fails it turns to 1 (and never get back to zero). The problem is to understand the influences of line failures on each other.
My teacher told me that to try this by time series analysis tools, but I’m not sure about it.
Is it reasonable to use this methods for such problems? If yes which tools do you suggest?
Thank you so much
Hello Sir, Can you please share an example code using your Fuzzy Logic timeseries library..
I want to implement Fuzzy Logic time series, and i am just a student, so that’s why it will be a great help from you if you will help me in this.
I just need a sample code that is written in python.
Thank you so much for the many code examples on your site. I am wondering if you can help an amatur like me on something.
When I pull data from our database, I generally do it for multiple SKU’s at the same time into a large table. Considering that there are thousands of unique SKU’s in the table, is there a methodology you would recommend for generating a forecast for each individual SKU? My initial thought is to run a loop and say something to the effect of: For each in SKU run…Then the VAR Code or the SARIMA code.
Ideally I’d love to use SARIMA, as I think this works the best for the data I am looking to forecast, but if that is only available to one SKU at a time and VAR is not constrained by this, it will work as well. If there is a better methodology that you know of for these, I would gladly take this advice as well!
Great post. I’m currently investigating a state space approach to forecasting. Dynamic Linear Modeling using a Kálmán Filter algorithm (West, Hamilton). There is a python package, pyDLM, that looks promising, but it would be great to hear your thoughts on this package and this approach.
Thanks for your great tutorial posts. This one was very helpful. I am wondering if there is any method that is suitable for multivariate time series with a trend or/and seasonal components?
Thanks for your respond. I also have another question I would appreciate if you help me.
I have a dataset which includes multiple time series variables which are not stationary and seems that these variables are not dependent on each other. I tried ARIMA for each variable column, also VAR for the pair of variables, I expected to get better result with ARIMA model (for non-stationarity of time series) but VAR provides much better prediction. Do you have any thought why?
In the (S/V)ARIMAX procedure, should I check to see if my exogenous regressors are stationary and difference if them if necessary before fitting?
Y = data2 = [x + random() for x in range(101, 200)]
X = data1 = [x + random() for x in range(1, 100)]
If I don’t, then I can’t tell if a change in X is related to a change in Y, or if they are both just trending with time. The time trend dominates as 0 <= random() <= 1
In R, Hyndman recommends "[differencing] all variables first as estimation of a model with non-stationary errors is not consistent and can lead to “spurious regression”".
I do have a question regarding data that isn’t continuous, for example, data that can only be measured during daylight hours. How would you approach a time series analysis (forecasting) with data that has this behavior? Fill non-daylight hour data with 0’s or nan’s?
Hey..
Kindly Help us in making hybrid forecasting techniques.
Using two forecasting technique and make a hybrid technique from them.
Like you may use any two techniques mentioned above and make a hybrid technique form them.
Thanks.
I wondered which is the best way to forecast the next second Packet Error Rate in DSRC network for safety messages exchange between vehicles to decide the best distribution over Access Categories of EDCA.
I hesitated to choose between LSTM or ARMA methodology.
Could you please guide me to the better method of them ?
Kindly, note that I’m beginner in both methods and want to decide the best one to go deep with it because I don’t have enouph time to learn both methods especially they are as I think from different backgrounds.
Hi Jason,
Thanks for great post. I have 2 questions. First, is there a way to calculate confidence intervals in HWES, because i could not find any way in the documentation. And second, do we have something like ‘nnetar’ R’s neural network package for time series forecasting available in python.
Regards
Thanks for your reply Jason. “nnetar” is a function in R, https://www.rdocumentation.org/packages/forecast/versions/8.4/topics/nnetar
it is used for time series forecasting. I could not find anything similar in Python.
but now i am using your tutorial of LSTM for time series forecasting.
And i am facing an issue that my data points are 750. and when i do prediction the way you have mentioned i.e. feed the one step forecast back to the new forecast step. So, the plot of my forecasting is just the repetition of my data. Forecast look just like the cyclic repetition of the training data. I don’t know what am i missing.
Hi Jason,
Thank you for this great post!
In VARMAX section, at the end you wrote:
“The method is suitable for univariate time series without trend and seasonal components and exogenous variables.”
I understand from the description of VARMAX that it takes as input, multivariate time series and exogenous variables. No?
Another question, can we use the seasonal_decompose (https://www.statsmodels.org/dev/generated/statsmodels.tsa.seasonal.seasonal_decompose.html) function in python to remove the seasonality and transform our time series to stationary time series? If so, is the result residual (output of seasonal_decompose) is what are we looking for?
The seasonal_decompose function implemented in python gives us 4 resutls: the original data, the seasonal component, the trend component and the residual component. Which component should we use to forecast this curve? the residual or the trend component?
For the AR code, is there any modification I can make so that model predicts multiple periods as opposed to the next one? For example, if am using a monthly time series, and have data up until August 2018, the AR predicts September 2018. Can it predict September 2018, October, 2018, and November 2018 based on the same model and give me these results?
I have dynamic demand forecasting problem, i.e. simple time series but with DaysBeforeDeparture complexity added.
Historical data looks like:
daysbeforedeparture – DepartureDate Time – Bookings
175. 21 09 2018. 16 00 00 – 10
Etc.
So taking into account the dynamics (daysbeforedeparture) is vital.
I am trying to use fbProphet model, aggregating time series by hour, but no idea how to deal with DBDeparture feature.
I think the date is redundant as you already have days before departure.
I think you may have 2 inputs, the days before departure, bookings, and one output, the bookings. This would be a multivariate input and univariate output. Or univariate forecasting with an exog variable.
In fact, if you have sequential data for days before departure, it is also redundant, and you can model bookings directly as a univariate time series.
Hi, thank you so much for your post.
I have a question, have you used or have you any guidelines for the use of neural networks in forescating time series, using CNN and LSTMboth together?
All methods have common problems. In real life, we do not need to predict the sample data. The sample data already contains the values of the next moment. The so-called prediction is only based on a difference, time lag. That is to say, the best prediction is performance delay. If we want to predict the future, we don’t know the value of the current moment. How do we predict? Or maybe we have collected the present and past values, trained for a long time, and actually the next moment has passed. What need do we have to predict?
You can frame the problem any way you wish, e.g. carefully define what inputs you have and what output you want to predict, then fit a model to achieve that.
Dear Jason : your post and book look interesting , I am interested in forecasting a daily close price for a stock market or any other symbol, data collected is very huge and contain each price ( let’s say one price for each second) , can you briefly tell how we can predict this in general and if your book and example codes if applied will yield to future data.
can we after inputting our data and producing the plot for the past data , can we extend the time series and get the predicted priced for next day/month /year , please explain
Thank you for this great post.
I have a requirement of predicting receipt values for open invoices of various customers. I am taking closed invoices – whose receipt amount is used to create training data and open invoices as test data.
Below is the list of columns I will be getting as raw data
For Test Data – RECEIPT_AMOUNT, RECEIPT_DATE will be blank, depicting Open Invoices
For Training Data – Closed Invoices will have receipt amount and receipt date
Hi Jason, STAR models are Space-Time Autoregression models. I have the same question. I have a multivariate time-series with additional spatial dimension: latitude and longitude. So we need to account not only for the time lags but also for the spacial interactions. I’m trying to find a clear example in Python with no luck so far…
I have a dataset on weekly detection rates of different bird species in different counties in the US. I want to build forecasting models to predict the detection rate of each species in each county, week by week.
I am thinking of using longitude, latitude as regressors for the model, so that I don’t have to run a model for each county. But I am not sure if I should treat spatial data different from numeric features. And if so, how should I do that?
Thanks for this.
I want to forecast whether an event would happen or not. Would that SARMAR actually work work if we have a binary column in it?
How would I accomplish something like this including the time?
Thank you very much this paper. I have a time series problem but i can’t find any technique for applying. My dataset include multiple input and one output like multiple linear regression but also it has timestamp. Which algorithm is the best solution for my problem?
I have a problem about time series data.
My dataset include multiple input and one output.
Normally it is like multiple linear regression but as additional has timestamp 🙁
So i can’t find any solution or algorithm.
For example: AR, MA, ARIMA, ARIMAX, VAR, SARIMAX or etc.
Which one is the best for my problem?
First of all, I have read two of your books(Basics_for_Linear_Algebra_for_Machine_Learning and deep_learning_time_series_forecasting) and the simplicity with which you explain difficult concepts is brilliant. I’m using the second one to face the problem hat I present below.
I’m facing a predicting problem for food alerts. The goal is to predict the variables of the most probable alert in the next x days (also any information I could get about future alerts is really useful for me).Alerts are recorded over time (so it’s a time series problem).
The problem is that observations are not uniform over time (not separated by equal time lapses), i.e: since alerts are only recorded when they happen, there can be one day without alerts and another with 50 alerts. As you indicate in your book, it is a discontiguous time series.
The entry for the possible model could be the alerts (each alert correctly coded as they are categorical variables) of the last x days, but this entry must have a fixed size/format. Since the time windows don’t have the same number of alerts, I don’t know what is the correct way to deal with this problem.
Any data formatting suggestion to make the observations uniform over time?
Or should I just face the problem in a different way (different inputs)?
There are many ways to frame and model the problem and I would encourage you to explore a number and discover what works best.
First, you need to confirm that you have data that can be used to predict the outcome, e.g. is it temporally dependent, or whatever it is dependent upon, can the model get access to that.
Then, perhaps explore modeling it as a time series classification problem, e.g. is the even going to occur in this interval. Explore different interval sizes and different input history sizes and see what works.
Thank you, i have a datset en csv format and i can open it in excel, it has data since 1984 until 2019, I want to train an artificial neural network in python i order to make frecastig or predictions about that dataset in csv format, I was thinking in a MLP, coul you help me Jason, a guide pls. Many thanks.
Working on one of the banking use cases i.e. Current account and Saving account attrition
prediction.
We are using the last 6 months data for training, we need to predict customers whose balance will reduce more than 70% with one exception, as long money invested in the same bank it is fine.
Great, if you could suggest, which models or time series models will be the best options to try in this case?
Thank you so much for this post I learned a lot. I am a fan of the ARMAX models in my work as a hydrologist for streamflow forecasting.
I hope you can share something about Gamma autoregressive models or GARMA models which work well even for non-Gaussian time series which the streamflow time series mostly are. Can we do GARMA in python?
This blog is very helpful to a novice like me. I have been running the examples you have provided with some changes to create seasonality for example (period of 10 as in 0 to 9, back to 0, again to 9 with randomness thrown in). Linear regression seems to be better at it than the others which I find surprising. What am I missing?
Hi Jason,
Thank you for all your posts, they are so helpful for people who are starting in this area. I am trying to forecast some data and they recommended me to use NARX, but I haven’t found a good implementation in python. Do you know other method implemented in python similar to NARX?
The line in VARMAX “The method is suitable for multivariate time series without trend and seasonal components and exogenous variables.” is very confusing.
I guess you mean no trend no seaonal but with exogenous?
Hi Jason,
Thanks for the post. It was great and easy to understand for a beginner in time series.
I have data of past 4 years of number of users logged in for a day .
I want to predict the number of users for a day per month.
I used ARIMA but I am getting RMSE 2749 and R2 score 60% .
Can you please suggest methods to increase the accuracy as well as RMSE.
I have a question. What if my data is time series with multiple variables including categorical data, which model should be used for this? For example, i’m predicting The Air pollution level using the previous observation value of Temperature + Outlook (rain or not).
Hi Jason,
Thanks for the great post again, wonderful learning experience.
Do you have R codes for the time series methods you described in your article?
or Can you suggest me a good source where can i get R codes to learn some of these methods?
Thanks for the post. Please do check out AnticiPy which is an open-source tool for forecasting using Python and developed by Sky.
The goal of AnticiPy is to provide reliable forecasts for a variety of time series data, while requiring minimal user effort.
AnticiPy can handle trend as well as multiple seasonality components, such as weekly or yearly seasonality. There is built-in support for holiday calendars, and a framework for users to define their own event calendars. The tool is tolerant to data with gaps and null values, and there is an option to detect outliers and exclude them from the analysis.
Ease of use has been one of our design priorities. A user with no statistical background can generate a working forecast with a single line of code, using the default settings. The tool automatically selects the best fit from a list of candidate models, and detects seasonality components from the data. Advanced users can tune this list of models or even add custom model components, for scenarios that require it. There are also tools to automatically generate interactive plots of the forecasts (again, with a single line of code), which can be run on a Jupyter notebook, or exported as .html or .png files.
I have a question for time series forecasting. Have you heard about Dynamic Time Warping? As far as I know, this is a method for time series classification/clustering, but I think it can also be used for forecasting based on the similar time series. What do you think about this method compared to ARIMA? Do you think it will be better if I combine both two methods? For example, use DTW to group similar time series and then use ARIMA for each group?
Can you please explain why you use len(data) in your predict arguments? I was using the .forecast feature for a while which is for out of sample forecasts but I keep getting an error on my triple expo smoothing. Apparently, the .predict can be used for in-sample prediction as well as out of sample. The arguments are start and end, and you use len(data) for both which is confusing me. Will this really forecast or will it just produce a forecast for months in the past?
Thanks for the reply! I was reading through the explanation in your linked article and it was great. Can the .predict() do multiple periods in the future like .forecast? I was using .forecast(12) for forecasting 12 months into the future.
EDIT: Dumb question- did not read until the end. If you got time, check out my stack post though: https://stackoverflow.com/questions/56709745/statsmodels-operands-could-not-be-broadcast-together-in-pandas-series . I think i found some error in statsmodels as the error is thrown in the .forecast function as it wants me to specify a frequency. I found the potential misstep by reading through the actual code for ExponentialSmoothing and it was the only part that really referenced the frequency. Its either that, or my noob is really showing.
Thanks for the helpful tutorial. I’ve been trying to solve a sales forecast problem but haven’t been any successful. The data is the monthly records of product purchases (counts) with their respective prices for ten years. The records do not show either a significant auto-correlation for a wide range of lags or seasonality. The records are stationary though.
Among the time series models, I have tried (S)ARIMA, exponential methods, the Prophet model, and a simple LSTM. I have also tried regression models using a number of industrial and financial indices and the product price. Unfortunately, no method has led to an acceptable result. With regression models, the test R^2 is always negative.
My questions are:
* What category of problems is this problem more relevant to?
* Do you have any suggestions for possibly suitable approaches to follow for this kind of problems?
I guess that might be the case. I’m also guessing that maybe I don’t have sufficiently relevant explanatory variables to obtain a good regression model. Thanks for your feedback though. And, thanks again for your very helpful tutorials.
Hello Jason,
Do you know if there any way to randomly sample a fitted VARMA time series using the statsmodel library. Just like using the sm.tsa.arma_generate_sample for the ARMA.
I cant seem to see this how this is done anywhere.
Hello Jason,
It is really nice and informative article.
I am having network data and in that data there are different parameters (network incidents) which slows down the network. From the lot of parameters one is network traffic. I have DATE and TIME when network traffic crosses a certain threshold at certain location. I need to draw a predictive model so that i can spot when network traffic is crossing threshold value at a particular location, so that i can take preventive measures prior to occurrence of that parameter at that place.
So please can you suggest me appropriate model for this problem.
We’ve been having trouble with statsmodels’ ARIMA. It just doesn’t work and takes forever. What can you tell us about these issues? Do you know of any alternatives to statsmodels?
Of course there are many regression models available in sklearn. The point is that statsmodels seems to fail miserably, both in time and accuracy. That is, with respect to their arima (family) set of functions.
Question is if you know an alternative python library providing that missing functionality. R works well. Mathematica even better. At the moment, my students are working on interfacing with R, since we have not found a sound Python library for arima.
In time series classification when I am plotting it is showing day wise data . For an instance if i am taking data of past 1 month and applying Autoregression time series classification on it then i am not able to fetch detailed outputs from that chart. From detailed output I mean that 1 incident is occuring number of times in a day so I want visibility in such a way so that everytime an incident occur it should be noticed on that plot.
Is there any precise way by which I can do it.I would be really thankful if you would help me.
Thank you
I can’t seem to make VAR work? Gives me a lot of errors?
—————————————————————————
ValueError Traceback (most recent call last)
in ()
11 # fit model
12 model = VAR(data)
—> 13 model_fit = model.fit()
14 # make prediction
15 yhat = model_fit.forecast(model_fit.y, steps=1)
D:\Users\Berns\Anaconda3\envs\time_series_p27\lib\site-packages\statsmodels\tsa\vector_ar\var_model.pyc in _estimate_var(self, lags, offset, trend)
666 exog = None if self.exog is None else self.exog[offset:]
667 z = util.get_var_endog(endog, lags, trend=trend,
–> 668 has_constant=’raise’)
669 if exog is not None:
670 # TODO: currently only deterministic terms supported (exoglags==0)
D:\Users\Berns\Anaconda3\envs\time_series_p27\lib\site-packages\statsmodels\tsa\vector_ar\util.pyc in get_var_endog(y, lags, trend, has_constant)
36 if trend != ‘nc’:
37 Z = tsa.add_trend(Z, prepend=True, trend=trend,
—> 38 has_constant=has_constant)
39
40 return Z
# VAR example
from statsmodels.tsa.vector_ar.var_model import VAR
from random import random
# contrived dataset with dependency
data = list()
for i in range(100):
v1 = i + random()
v2 = v1 + random()
row = [v1, v2]
data.append(row)
# fit model
model = VAR(data)
model_fit = model.fit()
# make prediction
yhat = model_fit.forecast(model_fit.y, steps=1)
print(yhat)
Jason , need one clarification , for a SARIMAX model or in general in timeseries model, should I use .forecast or .predict ? I have got some differences in result using this two . Please suggest should we use .forecast for train and test ( to get the model fitted values) or should we use .predict for train and test ( to get the model fitted values) . Also for future forecast what need to be used ? Please suggest.
GREAT article. I’ve been looking for an overview like this. I do have a side question about a project I’m working on. I’ve got a small grocery distribution company that wants weekly forecasts of its 4000+ different products/SKU numbers. I’m struggling with how to produce a model that can forecast all these different types of time-series trends and seasonalities. ARIMA with Python doesn’t seem like a good option because of the need to fine tune the parameters. I would love you’re feedback on the following ideas:
1. Use facebooks prophet for a “good enough” model for all different types of time series trends and seasonalities.
2. Separate the grocery items into 11-20 broader category groups and then produce individual models for all those different categories that can then be applied to each individual sku within that category.
LIMITATIONS:
1. I need this process to be computationally efficient, and a for-loop for 4000 items seems like a horrible way to go about this.
QUESTIONS:
1. Do i have to model sales and NOT item quantities and why is that?
2. Is there a way to cluster the items with unsupervised learning just based off their trend and seasonality? Each item does not have good descriptor category so I don’t have much to work with for unsupervised clustering techniques.
Would love to hear any thoughts you have about this type of problem!
I have a simulation model of a system which receives a forecast of a time series as input.
In my scientific work I would like to examine how the performance of the simulation model behaves in relation to the accuracy of the forecast. So I would like to have forecasts between 80% (bad prediction) and 95% (very good prediction) for my time series.
How would you recommend me to generate such pseudo predictions that are meaningful for real predictions? Maybe adjusting a mediocre prediction in both direction?
My problem is that my simulation model receives forecasted time series as input, from which I know from publication papers that they are 95% predictable with a lot of effort and large amounts of training data.
Simple predictions like the ones I make, however, only reach about 80%. Nevertheless, I would like to analyze for my scientific work how much the quality of the simulation model depends on the quality of the forecasts. For example, to say whether the effort for good forecasts is worth it.
Is there a way to make pseudo forecasts of time series for this purpose? As a basis the actual data and the forecast data are available with 80% accuracy.
Part of my question also aimed at how I best proceed when contriving forecasts, as they should be as representative as possible.
The simplest would be to add normal distributed noise to the actual data, but I don’t know if that would be representative if the error was only caused by white-noise.
I have to perform a regression on a data set that has the following data
Age of Person at time of data collection
Date
cholesterol Level
HDL
Event which can be death etc.
Need to find out if there are any patterns in the Cholesterol Level or the HDL Levels prior to the death event.
I’m trying to do this using Jupyter Notebook. Really appreciate your help sincerely.
Do you have any examples that would be close ?
Hey Jason! I am working on supply elasticities for crops in California, I am new using python your post has been incredible useful, I have one question for my model I need to develop a time series dynamic model particularly an Autoregressive Distributed Lag (ARDL) Model, do you have any material related to how perform this type of models on python?
Actually I am very impatient and did not read all the comments but thot that you already knew them 🙂
I need to make a hierarcial and grouped time series in Python.
I need to predict daily & hourly passenger unit by continent or vessel name. I investigated but did not find a solution, could you pls direct me what can I do.
Below are my inputs for the model.
Date & Time Vessel Name Continent Passenger Unit
Hi Jason,
For the simple AR model yhat values is it possible to get the threshold?
data = [x + random() for x in range(1, 100)]
model = AR(data)
model_fit = model.fit()
yhat = model_fit.predict(len(data), len(data))
I see that you are using random function to generate the data sets. Can you suggest any pointers where I can get real data sets to try out these algorithms. If possible IoT data sets would be preferable.
Hey
It is really great to see all this stuff.
Can you please explain me:
which algorithm should i use when i want to predict the voltage of a battery:ie…
battery voltage is recorded from 4.2 to 3.7 is 6 months which algorithm will be the best one to predict the next 6 months
Buenas tardes, se podría utilizar el walk fordward validación para series de tiempo multivariado ? es decir una variable Target y varias variables input, y k te haga la técnica de wall fordward validación?? esa es mi pregunta, saben de algún link o tutoría
Have you ever evaluated the performance of Gaussian processes for time series forecast? Do you have any material on this (or plan to make a post about this)?
Hi Jason,
I’m currently working on a sales data set which has got date, location, item and the unit sales. Time period is 4 yrs data, need to forecast sales for subsequent year -1 month(1-jan to 15 jan). Kindly suggest a model for this multivariate data. Another challenge which i’m currently facing is working with google collab with restricted RAM size of 25 GB and the data set size is 1M rows.
Please suggest a good model which will handle the memory size as well.
Note: I have change the datatype to reduce the memory, still facing the issue
All these methods seem to focus on two data features:
– evenly spaced time series;
– continuous data.
As it is easy to guess, I have a time series which is both irregular and categorical.
Mine is sensor data. These sensors only detect the presence of a person in a room. Each entry gives me the start time, duration and room. This is for monitoring one person with health issues as they move through their flat.
None of the above methods — fantastically depicted, by the way! — is a glove to this hand. I need to detect patterns, trigger alerts on anomalies and predict future anomalies.
Hi Dr.Jason,
I have a question about time-series forecasting.I have to predict the value of Time-series A in the next few hours.And at the same time,the Time-series B have an influence on A.I have two ideas :
1.X_train is time T and Y_train is the value of A_T and B._T
2 X_train is the value of (A_n, A_n-1, A_n-2,…A_n-k, B_n, B_n-1, B_n-2,…,B_n-k) ,Y_train is (A_n+1, A_n+2,…,A_n+m)
Which is better?
One other thing – do you know where I could find an explanation on how the solvers work (relating to the theory) ? Since I’ll need to know what config/parameter changes will do.
I did access the links you posted (e.g. statsmodels.tsa.ar_model.AR API) , but I don’t find these very descriptive, and they are without code examples.
I appreciate that you posted this explanation about the models.
I would like to apply the algorithms to my problem.
I was in doubt as to how to extract the predicted value and the current value of each prediction. I would like to calculate the RMSE, MAPE, among others.
hello Jason, thank you for your post. it is so interesting.
In fact i have a question. I have many time series Trafic per region. I don’t want ot have 40 models to forecast trafic for each region.
Do you have an idea how can i use one model to predict trafic for all regions so one model to a multiple time series?
Dear Jason,
Do you have a blog where you forecast univariate time series using a combination of exponential smoothing and neural network?
I mean where we give lags of time series to exponential smoothing model and the output of exponential smoothing will be the input of a neural network?
Could you also please review the BATS and TBATS models? (De Livera, A.M., Hyndman, R.J., & Snyder, R. D. (2011), Forecasting time series with complex seasonal patterns using exponential smoothing, Journal of the American Statistical Association, 106(496), 1513-1527.) Like SARIMA, they are quite good to deal with seasonal time series. And I am wondering if you have any plan to review Granger causality which is often used to find time series dependencies from multivariate time series? Thank you.
hello
i have a dtaset of many vehicle which ask me to predict output by taking 5 timeseries input and i have to do this for all vehicle almost 100 vehicle to improve accuracy of model,can u help me
Hello, which one of the methods you shown is applicable to very limited data? Like 5 data pts only, univariate. There’s no way to increase it because my subject has just began operating 5 yrs ago and this data should be forecasted annually.
As far as I understand
Single input : Univariate Model
Multiple Input : Multivariate Model
But I guess where each of the models, in which scenario fits exceeds the scope of this topic i.e cheat sheet.
However in short words(since I am a beginner to time-series) can you tell me where each of these model fit in what scenario, so I can have a basic understanding of where to use each model.
However I am having IoT data project in my company. where you get to see monitored data in timestamp. I am being asked to make Pv predictions. Can you suggest a model as well that would be great :
Datetime Pv Min Max Avg
01-01-2019 07:00 20.1 20.1 20.1 20.1
01-01-2019 07:03 20 20 20 20
Thanks for your great tutorial.
Are there any other references for multivariate time-series forecasting with ARIMA and other classical time-series approaches?
I have a question on lag orders, in case of univariate techniques, we can decide p and q values using autocorrelation and partial autocorrelation. But, in the case of multivariate techniques like VAR, VMA, and VARMA how we can decide these lag orders as we have multiple parallel series?
Thank you for the great post.
I have a time series showing departure date, booking date and number of booked passengers. It has both trend and seasonal components and I am working on passengers forecast by considering departure week day, week, month and prior days ( departure date minus booking date). What is your model suggestion to implement? MLP?
I have a multivariate time seies problem formulation that seeks to forecast endogenous variable y, as a function of exogenous variables, x1,x2,…,xn. I thought VARMAX method is applicable in this situation, but it seems VARMAX() can only take multiple exogenous varaibles when I tried to implement the method. However, one edogenous or response variable is needed in my case.
I need to create a baseline model for the stated multivariate problem. I will appreciate your thought in this situation.
I undertand the aspect of the persistence model that would based on univariate data – the respose variable.
For the second option, I guess you meant fiting a linear regression model on a suppervised learning(window) transform of the response variablle, right?
I have a multivariate time seies problem formulation that seeks to forecast endogenous variable y, as a function of exogenous variables, x1,x2,…,xn. I thought VARMAX method is applicable in this situation, but it seems VARMAX() can only take multiple endogenous varaibles when I tried to implement the method. However, one edogenous or response variable is needed in my case.
I need to create a baseline model for the stated multivariate problem. I will appreciate your thought in this situation.
I’m a trainee and are now executing my literature review with the goal to find the right method for solving my internship problem. During my search for relevant articles I come across a lot of your blogs and read some of them to find the method suitable for my problem. However, I still haven’t found it yet. Maybe you can direct me in the right way?
From the department demand planning at my internship organization I get a monthly forecast of the amount of products (medical devices) that will be ordered. I have been instructed to break this monthly forecast down in daily forecast. So, I’m looking for a (machine learning) method that distributes the monthly forecast into daily buckets. Am I looking in the right direction if I’m searching for Time Series forecasting algorithms? I find that a lot of these algorithms can be used to determine the monthly forecast and not the distribution of the monthly forecast into daily buckets.
Can you help me putting me in the right direction for searching algorithms for distribution?
Sorry if the description of my problem and question is a bit hazy 🙂
Thank you for sharing all of these time series examples. I would like to do a weighted time series forecast similar to Holt Winter’s Exponential Smoothing but with multivariate data. What would you recommend for this? The objective is to have a time series forecast with greater weight to more recent data
hi Jason, How can Time Series Analysis be done with Categorical Variables?
Right now I’m doing a project where I have a database of network infrastructure alames in a space of 3 months, in this case a time series problem.
Most of the time series analysis tutorials/textbooks I’ve read about, be they for univariate or multivariate time series data, usually deal with continuous numerical variables.
I currently have a problem at hand that deals with multivariate time series data, but the fields are all categorical variables. Hence, I was wondering if there is any way to use the standard time series analysis techniques (such as ARIMA, ARMA etc.)
Specifically, my data is a stream of alert data, where at each time stamp, information such as the alert monitoring system, the location of the problem etc. are stored in the alert. These fields are all categorical variables.
What advice can you give me regarding the problem, and which ML algorithms are most recommended for this case?
Hi Jaron, there seems to be a bug in your SARIMA example, we get at line
model = SARIMAX(data, order=(1, 1, 1), seasonal_order=(1, 1, 1, 1))
Traceback (most recent call last):
File “”, line 1, in
File “C:\Users\User\AppData\Local\Continuum\anaconda3\envs\Py37_Tf2_Ke2\lib\site-packages\statsmodels\tsa\statespace\sarimax.py”, line 332, in __init__
missing=missing, validate_specification=validate_specification)
File “C:\Users\User\AppData\Local\Continuum\anaconda3\envs\Py37_Tf2_Ke2\lib\site-packages\statsmodels\tsa\arima\specification.py”, line 303, in __init__
raise ValueError(‘Seasonal periodicity must be greater’
ValueError: Seasonal periodicity must be greater than 1.
Sourabh Shrikrishna ChinchanikarJanuary 6, 2021 at 2:45 am#
Hi Jason,
Thank you for sharing all types of time series examples.
In my project work I want to predict vehicle speed considering different inputs such as current speed of vehicle, traffic light states, distance between two vehicles etc. Which approach shall I follow as it has trends and seasonality? its kind of multiple input and singe output scenario.
Thank for the tutorial. I read your book on time series forecasting. Usually, time column is dropped and then, train and test the model. In my work i want to predict the “time”. So i can’t remove “time” column. The time is represented as “timestamp”. Can you please explain me well how to preprocess the data in order to forecast the “time”.
I know how to prepare the data and split the dataset, it’s well explained in your book
Cordially,
Interesting classical tutorials (via statistics analysis) that many times are better than advance Neural ML models such as MLP, LSTM.
But this “statsmodels” library are not uniform on its terminology methods for each type of model, sometimes requiring “.predict()” but other times requiring “.forecast( )”, with different arguments on it (sometimes limits of length prediction, sometimes needing to declare the variable to predict (but other not) …so not uniformity on them on the contrary when you use Keras methods you know which arguments to declare.
in addition to it the exogenous variables as far I understand you need to provide on both model definition and the “.predict()” methods…but at the end the prediction is for one feature (univariate) but on multivariate are so many outputs as multivariate.
Also when you use multivariate models (VAR, VARMA,..) you also get multiple output (as many as multivariate) … I thought you can uses to two features timeseries to produce a complete new one output timeseries …as i thought it is allowed on keras MLP, LSTM, etc…So my intuition fail 🙁
The terminology for orders (p,d,q..P,Q,D, mm) it is also confused. I think finally they refer to lags o timesteps required for each order …
Do you have any updated book about classical methods on python? As mentioned, statsmodels are a mess through the years and in my university project i want to use Simplexponentialsmooth, Holtwinters and ARIMA to predict my data but i’m having a lot of problems about statsmodels.
1) how to check stationarity for multivariate time series analysis and if i use ‘Johansen test’ for multivariate time series how i conclude the results.
2)can you please update here if you have python code for johansen test or any other test is there to check stationarity for multivariate?
3) can i use ACF & PACF values for VARMAX models or it is only supports for univariate models.
I am working in petrochemical sector and want to forecast the production of ethylene in cracker furnace based upon certain no. of chemical parameters like pressure, temperature ,composition of the gas with time, which models should I try to use, like LSTMs or classical time series methods?
Hello Jason,
I have to backcast values of a time series. Is it enough to invert the time series and procede with forecasting?
I was wondering if there is some python code related to backcasting.
Thank you.
Can we use categorical variables(binary) as exogenous variable in Sarimax. I am trying to forecast Premium collection and i want to use covid impact as an exogenous variable.
Just anted to know your thoughts if this can be done. I have weekly collectin data for last 4 years.
It should work as long as you encoded it as a number (e.g. 0 or 1). But you should aware that the SARIMAX model still consider that as a real number. If you are predicting these categorical variable, the model may give you values like 0.7 instead of 0 and 1.
Hi Jason,
How do we forecast time series which is having less than 10 observations ? I have time series for every lab roviding volumes for customer claiming for insurance for a particular disease. Some of these id’s have volumes for all the quarters in 4 years and few id’s have only one insurance claim. So how do I handle this short time series?
Basically your hands are tied under such small amount of data. I know the actuarial have their own tricks to supplement that. For example, they would consider certain thing are in a particular statistical model. Or they would consider disease A and disease B can be mixed together so you can borrow the data from one to another. I do not have the domain knowledge to give more detail but think out of the box is the only way you can do here.
I’m trying to solve a problem using timeseries forecasting. For that, i used the n-beats model and the prediction is ok. However, i need to know how to define an extreme event after i forecast the results.
is there a library already for that? is it a data distribution dependent? as my data doesn’t fall under one of the famous distributions.
That’s really depends on the data distribution. If your data doesn’t fall into anything, try to skew your data a bit (e.g., take logarithm, take square of it, etc.) and find the best match. That’s what we need to do in the real world problems.
Hello Dr. Jason,
Greetings.
I want ro work on a prediction model for number of users in a Telecom Network. Which model do you think would be best for such a scenario?
Best Regards
Thank you for this amazing content. Will you kindly help me with my question?
I’m currently working with data obtained from a test conducted, which contains a lot of values for parameters such as force, time, torque, velocity, etc. I’m still confused about which model to proceed with. This brings me to my question.
How should I select the most suitable model for my time series prediction task?
when we use machine learning models for time series data, we use ARIMA first and then its AR terms as independent variables. Do we have any other criteria that we can use directly by avoiding the use of ARIMA models AR and MA terms, to analyse univariate time series data?
I can show u how wrong u are … u just need to be creative to transform history drawing database to pattern data base in which u don’t predict winning numbers but u analyze pattern occurrence by confidence of appearance. Now u cant do this “manually” but with reinforcement learning by back propagation to correct used NN’s for each new drawing (and its not time series). This way u can achieve up to 80% plus confidence analytically (not always it’s still % of confidence) and u can pick the right pattern (very important) if u consider other database properties behavior (behaves no random but structured). U can’t achieve that the easy way. U have to dive into coding very deep with a lot of testing, u just have to learn to code at phD level. Maybe i said to much as u re phD in this field already.
Hi Cris…Thank you for your feedback! I agree that deep reinforcement learning is an emerging and exciting area of research! Let us know what you find out from your models.
Hi Jason, thanks for such an excellent and comprehensive post on time series. I sincerely appreciate your effort. As you ask for the further topic, just wondering if I can request you for a specific topic I have been struggling to get an output. It’s about Structural Dynamic Factor model ( SDFM) by Barigozzi, M., Conti, A., and Luciani, M. (Do euro area countries respond asymmetrically to the common monetary policy) and Mario Forni Luca Gambetti (The Dynamic Effects of Monetary Policy: A Structural Factor Model Approach). Would it be possible for you to go over and estimate these two models using Python or R? It’s just a request from me and sorry if it doesn’t go with your interest.
Thanks for the suggestion. I’ve not heard of that method before.
I have 1000 time series data and i want to predict next 500 steps. So how can i do this
Fit a model then call the predict() function.
Hi Jason can the above mentioned algorithms can be used for Demand forecast ing as well?
I don’t see why not.
Hi Jason,
Great technical insights. I work in the oil and gas sector and one of our day to day job is to to rate vs time oil rate forecasting. We typically use empirical equations and use regression to fit those empirical input parameters. I was wondering if we could use sophisticated ML techniques to accomplish this?
Thanks!
Thanks.
Yes, perhaps start with the examples here:
https://machinelearningmastery.com/start-here/#deep_learning_time_series
Hello Jason, thanks for your great work.
Can you give an example on how to fit a model and use of predict() function?
Yes, see this:
https://machinelearningmastery.com/arima-for-time-series-forecasting-with-python/
Thanks for sharing
Dear Adrena contact me I may be able to help you of jason agrees, amybe we can work on a research paper together.
#Abhishek I work in oil&gas company too, and I tried to generate forecasts using ML. the problem is that you can’t relay only on prodaction rates, some parameters should be included in the model such as WHFP, BHFP, SEP P, WC… for all wells to accurately estimate the total production. An other problem is that even you came somehow to collect those data you cannot relay on them because generally those data are daily averages and dont reflect the variations in real time. Also, prodaction rates for each well are genarelly calculated using allocation method.
I am working on Time series or Prediction with neural network and SVR, I want to this in matlab by scratch can you give me the references of this materials
Thank you in advance
Sorry, I don’t have any materials for matlab, it is only really used in universities.
not true, but ok
Fair enough.
The demand for “python machine learning” is orders of magnitude higher than “matlab machine learning”, this drives my focus:
https://trends.google.com/trends/explore?date=today%205-y&q=python%20machine%20learning,matlab%20machine%20learning
Hi Jason! From which editor do you import the python code into the webpage of your article? Or what kind of container it that windowed control used to display the python code?
Great question, I explain the software I use for the website here:
https://machinelearningmastery.com/faq/single-faq/what-software-do-you-use-to-run-your-website
Thanks for all the things to try!
I recently stumbled over some tasks where the classic algorithms like linear regression or decision trees outperformed even sophisticated NNs. Especially when boosted or averaged out with each other.
Maybe its time to try the same with time series forecasting as I’m not getting good results for some tasks with an LSTM.
Always start with simple methods before trying more advanced methods.
The complexity of advanced methods just be justified by additional predictive skill.
Hi Jason,
Thanks for this nice post!
You’ve imported the sin function from math many times but have not used it.
I’d like to see more posts about GARCH, ARCH and co-integration models.
Best,
Elie
Thanks, fixed.
I have a post on ARCH (and friends) scheduled.
Will you consider writing a follow-up book on advanced time-series models soon?
Yes, it is written. I am editing it now. The title will be “Deep Learning for Time Series Forecasting”.
CNNs are amazing at time series, and CNNs + LSTMs together are really great.
will the new book cover classical time-series models like VAR, GARCH, ..?
The focus is deep learning (MLP, CNN and LSTM) with tutorials on how to get the most from classical methods (Naive, SARIMA, ETS) before jumping into deep learning methods. I hope to have it done by the end of the month.
This is great news! Don’t you think that R is better suited than Python for classical time-series models?
Perhaps generally, but not if you are building a system for operational use. I think Python is a better fit.
Great to hear this news. May I ask if the book also cover the topic of multivariate and multistep?
Yes, there are many chapters on multi-step and most chapters work with multivariate data.
Well, although it is a really helpful and useful book as it is usually made by Jason, this book does not cover multivariate time series problems, in fact Jason explicitely says “This is not a treatment of complex time series problems. It does
not provide tutorials on advanced topics like multi-step sequence forecasts, multivariate time series problems or spatial-temporal prediction problems.”
Correct. I cover complex topics in “deep learning for time series forecasting”.
Sounds amazing that you finally 😉 are geting the new book out on time-series models – when will it be available to buy?
Thanks. I hope by the end of the month or soon after.
Hi, Can you help me with Arimax ?
I use Prophet.
https://facebook.github.io/prophet/docs/quick_start.html
Also, sometimes FastFourier Transformations gives a good result.
Thanks.
I would second the use of prophet, especially in the context of shock events — this is where this approach has a unique advantage.
Thanks for the suggestion.
Hi,can you pls help to get the method for timeseries forecasting of10000 products at same time .
I have some suggestions here that might help (replace “site” with “product”):
https://machinelearningmastery.com/faq/single-faq/how-to-develop-forecast-models-for-multiple-sites
Hi Arun,
Can you let me know how you worked with fbprophet. I am struggling with the installation of fbprophet module. Since it’s asking for c++ complier. Can you please share how you installed the c++ complier. I tried all ways to resolve it.
Thanks
I have not worked with fbprophet, sorry.
conda install gccNice!
conda install libpython m2w64-toolchain -c msys2
pip install pystan
pip install fbprophet
you need to install above mentioned one by one separately and according to the sequence also..
Thanks!
What are the typical application domain of these algos?
Forecasting a time series across domains.
Hi Jason!
Firstly I congratulate you for your blog. It is helping me a lot in my final work on my bachelor’s degree in Statistics!
What are the assumptions for make forecasting on time series using Machine Learning algorithms? For example, it must to be stationary? Thanks!
Gaussian error, but they work anyway if you violate assumptions.
The methods like SARIMA/ETS try to make the series stationary as part of modeling (e.g. differencing).
You may want to look at power transforms to make data more Gaussian.
so when using machine learning algorithms there is no need to make data stationery?
It depends on the algorithm and the data, yes this is often the case.
Hi Jason
I’m interested in forecasting the temperatures
I’m provided with the previous data of the temperature
Can you suggest me the procedure I should follow in order to solve this problem
Yes, an SARIMA model would be a great place to start.
Hey Jason,
Cool stuff as always. Kudos to you for making me a ML genius!
Real quick:
How would you combine VARMAX with an SVR in python?
Elaboration.
Right now I am trying to predict a y-value, and have x1…xn variables.
The tricky part is, the rows are grouped.
So, for example.
If the goal is to predict the price of a certain car in the 8th year, and I have data for 1200 cars, and for each car I have x11_xnm –> y1_xm data (meaning that let’s say car_X has data until m=10 years and car_X2 has data until m=3 years, for example).
First I divide the data with the 80/20 split, trainset/testset, here the first challenge arises. How to make the split?? I chose to split the data based on the car name, then for each car I gathered the data for year 1 to m. (If this approach is wrong, please tell me) The motivation behind this, is that the 80/20 could otherwise end up with data of all the cars of which some would have all the years and others would have none of the years. aka a very skewed distribution.
Then I create a model using an SVR, with some parameters.
And then I try to predict the y-values of a certain car. (value in year m)
However, I do not feel as if I am using the time in my prediction. Therefore, I turned to VARMAX.
Final question(s).
How do you make a time series prediction if you have multiple groups [in this case 1200 cars, each of which have a variable number of years(rows)] to make the model from?
Am I doing right by using the VARMAX or could you tell me a better approach?
Sorry for the long question and thank you for your patience!
Best,
Den
You can try model per group or across groups. Try both and see what works best.
Compare a suite of ml methods to varmax and use what performs the best on your dataset.
Hi Jason!
Excellent post! I also would like to invite you to know the Fuzzy Time Series, which are data driven, scalable and interpretable methods to analyze and forecast time series data. I have recently published a python library for that on http://petroniocandido.github.io/pyFTS/ .
All feedbacks are welcome! Thanks in advance!
Thanks for sharing.
Hi Jason,
I have multivariate binary time series. If a time series is 0 it shows us the specific transmission line in power system isn’t failed, as soon as it fails it turns to 1 (and never get back to zero). The problem is to understand the influences of line failures on each other.
My teacher told me that to try this by time series analysis tools, but I’m not sure about it.
Is it reasonable to use this methods for such problems? If yes which tools do you suggest?
Thank you so much
Hi, Fatemeh…While I cannot speak specifically to your project, I can recommend the following as a starting point:
https://machinelearningmastery.com/start-here/
Hello Sir, Can you please share an example code using your Fuzzy Logic timeseries library..
I want to implement Fuzzy Logic time series, and i am just a student, so that’s why it will be a great help from you if you will help me in this.
I just need a sample code that is written in python.
Hi Jason,
Thank you so much for the many code examples on your site. I am wondering if you can help an amatur like me on something.
When I pull data from our database, I generally do it for multiple SKU’s at the same time into a large table. Considering that there are thousands of unique SKU’s in the table, is there a methodology you would recommend for generating a forecast for each individual SKU? My initial thought is to run a loop and say something to the effect of: For each in SKU run…Then the VAR Code or the SARIMA code.
Ideally I’d love to use SARIMA, as I think this works the best for the data I am looking to forecast, but if that is only available to one SKU at a time and VAR is not constrained by this, it will work as well. If there is a better methodology that you know of for these, I would gladly take this advice as well!
Thank you so much!
Yes, I’d encourage you to use this methodology:
https://machinelearningmastery.com/how-to-develop-a-skilful-time-series-forecasting-model/
Great post. I’m currently investigating a state space approach to forecasting. Dynamic Linear Modeling using a Kálmán Filter algorithm (West, Hamilton). There is a python package, pyDLM, that looks promising, but it would be great to hear your thoughts on this package and this approach.
Sounds good, I hope to cover state space methods in the future. To be honest, I’ve had limited success but also limited exposure with the methods.
Not familiar with the lib. Let me know how you go with it.
Indeed it’s an excellent lib.
I use it almost everyday and it really improved the effectiveness of my forecasts over any other method.
Hi Jason, I noticed using VARMAX that I had to remove seasonality — enforcing stationarity .. now I have test and predictions data that I cannot plot (I can, but it doesn’t look right _at all_). I’m wondering if there are any built-ins that handle translation to and from seasonality for me? My notebook is online: https://nbviewer.jupyter.org/github/robbiemu/location-metric-data/blob/master/appData%20and%20locationData.ipynb
Typically I would write a function to perform the transform and a sister function to invert it.
I have examples here:
https://machinelearningmastery.com/machine-learning-data-transforms-for-time-series-forecasting/
Does that help?
Thanks for your great tutorial posts. This one was very helpful. I am wondering if there is any method that is suitable for multivariate time series with a trend or/and seasonal components?
Yes, you can try MLPs, CNNs and LSTMs.
You can experiment with each with and without data prep to make the series stationary.
Thanks for your respond. I also have another question I would appreciate if you help me.
I have a dataset which includes multiple time series variables which are not stationary and seems that these variables are not dependent on each other. I tried ARIMA for each variable column, also VAR for the pair of variables, I expected to get better result with ARIMA model (for non-stationarity of time series) but VAR provides much better prediction. Do you have any thought why?
No, go with the method that gives the best performance.
Hi Jason,
In the (S/V)ARIMAX procedure, should I check to see if my exogenous regressors are stationary and difference if them if necessary before fitting?
Y = data2 = [x + random() for x in range(101, 200)]
X = data1 = [x + random() for x in range(1, 100)]
If I don’t, then I can’t tell if a change in X is related to a change in Y, or if they are both just trending with time. The time trend dominates as 0 <= random() <= 1
In R, Hyndman recommends "[differencing] all variables first as estimation of a model with non-stationary errors is not consistent and can lead to “spurious regression”".
https://robjhyndman.com/hyndsight/arimax/
Does SARIMAX handle this automatically or flag me if I have non-stationary regressors?
Thanks
No, the library will not do this for you. Differencing is only performed on the provided series, not the exogenous variables.
Perhaps try with and without and use the approach that results in the lowest forecast error for your specific dataset.
Hi Jason,
Thank you for this wonderful tutorial.
I do have a question regarding data that isn’t continuous, for example, data that can only be measured during daylight hours. How would you approach a time series analysis (forecasting) with data that has this behavior? Fill non-daylight hour data with 0’s or nan’s?
Thanks.
I’d encourage you to test many different framings of the problem to see what works.
If you want to make data contiguous, I have some ideas here:
https://machinelearningmastery.com/handle-missing-timesteps-sequence-prediction-problems-python/
Hey..
Kindly Help us in making hybrid forecasting techniques.
Using two forecasting technique and make a hybrid technique from them.
Like you may use any two techniques mentioned above and make a hybrid technique form them.
Thanks.
Sure, what problem are you having with using multiple methods exactly?
Thank you for your excellent and clear tutorial.
I wondered which is the best way to forecast the next second Packet Error Rate in DSRC network for safety messages exchange between vehicles to decide the best distribution over Access Categories of EDCA.
I hesitated to choose between LSTM or ARMA methodology.
Could you please guide me to the better method of them ?
Kindly, note that I’m beginner in both methods and want to decide the best one to go deep with it because I don’t have enouph time to learn both methods especially they are as I think from different backgrounds.
Thank you in advance.
Best regards,
Mohammad.
I recommend testing a suite of methods in order to discover what works best for your specific problem.
Hi Jason,
Thanks for great post. I have 2 questions. First, is there a way to calculate confidence intervals in HWES, because i could not find any way in the documentation. And second, do we have something like ‘nnetar’ R’s neural network package for time series forecasting available in python.
Regards
I’m not sure if the library has a built in confidence interval, you could calculate it yourself:
https://machinelearningmastery.com/confidence-intervals-for-machine-learning/
What is “nnetar”?
https://www.rdocumentation.org/packages/forecast/versions/8.12/topics/nnetar
Thanks for your reply Jason. “nnetar” is a function in R,
https://www.rdocumentation.org/packages/forecast/versions/8.4/topics/nnetar
it is used for time series forecasting. I could not find anything similar in Python.
but now i am using your tutorial of LSTM for time series forecasting.
And i am facing an issue that my data points are 750. and when i do prediction the way you have mentioned i.e. feed the one step forecast back to the new forecast step. So, the plot of my forecasting is just the repetition of my data. Forecast look just like the cyclic repetition of the training data. I don’t know what am i missing.
Perhaps try this tutorial:
https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/
Hi Jason,
Thank you for this great post!
In VARMAX section, at the end you wrote:
“The method is suitable for univariate time series without trend and seasonal components and exogenous variables.”
I understand from the description of VARMAX that it takes as input, multivariate time series and exogenous variables. No?
Another question, can we use the seasonal_decompose (https://www.statsmodels.org/dev/generated/statsmodels.tsa.seasonal.seasonal_decompose.html) function in python to remove the seasonality and transform our time series to stationary time series? If so, is the result residual (output of seasonal_decompose) is what are we looking for?
Thanks!
Rima
Thanks, fixed.
What about Seasonal_decompose method? Do we use residual result or the trend?
Sorry, I don’t understand, perhaps you can elaborate your question?
The seasonal_decompose function implemented in python gives us 4 resutls: the original data, the seasonal component, the trend component and the residual component. Which component should we use to forecast this curve? the residual or the trend component?
I generally don’t recommend using the decomposed elements in forecasting. I recommend performing the transforms on your data yourself.
Hi Jason,
Could you please help me list down the names of all the models available to forecast a univariate time series?
Thanks!
Does the above post not help?
Hi Jason,
Thank you this was super helpful!
For the AR code, is there any modification I can make so that model predicts multiple periods as opposed to the next one? For example, if am using a monthly time series, and have data up until August 2018, the AR predicts September 2018. Can it predict September 2018, October, 2018, and November 2018 based on the same model and give me these results?
Yes, you can specify the interval for which you need a prediction.
How might I go about doing that? I have read through the statsmodel methods and have not found a variable that allows this
The interval is specified either to the forecast() or the predict() method, I given an example here that applies to most statsmodels forecasting methods:
https://machinelearningmastery.com/make-sample-forecasts-arima-python/
Hi Jason,
I have dynamic demand forecasting problem, i.e. simple time series but with DaysBeforeDeparture complexity added.
Historical data looks like:
daysbeforedeparture – DepartureDate Time – Bookings
175. 21 09 2018. 16 00 00 – 10
Etc.
So taking into account the dynamics (daysbeforedeparture) is vital.
I am trying to use fbProphet model, aggregating time series by hour, but no idea how to deal with DBDeparture feature.
Can you give me some directions?
I think the date is redundant as you already have days before departure.
I think you may have 2 inputs, the days before departure, bookings, and one output, the bookings. This would be a multivariate input and univariate output. Or univariate forecasting with an exog variable.
In fact, if you have sequential data for days before departure, it is also redundant, and you can model bookings directly as a univariate time series.
This may help:
https://machinelearningmastery.com/taxonomy-of-time-series-forecasting-problems/
Hi, thank you so much for your post.
I have a question, have you used or have you any guidelines for the use of neural networks in forescating time series, using CNN and LSTMboth together?
Yes, I have many examples and a book on the topic. You can get started here:
https://machinelearningmastery.com/start-here/#deep_learning_time_series
All methods have common problems. In real life, we do not need to predict the sample data. The sample data already contains the values of the next moment. The so-called prediction is only based on a difference, time lag. That is to say, the best prediction is performance delay. If we want to predict the future, we don’t know the value of the current moment. How do we predict? Or maybe we have collected the present and past values, trained for a long time, and actually the next moment has passed. What need do we have to predict?
You can frame the problem any way you wish, e.g. carefully define what inputs you have and what output you want to predict, then fit a model to achieve that.
Dear Jason : your post and book look interesting , I am interested in forecasting a daily close price for a stock market or any other symbol, data collected is very huge and contain each price ( let’s say one price for each second) , can you briefly tell how we can predict this in general and if your book and example codes if applied will yield to future data.
can we after inputting our data and producing the plot for the past data , can we extend the time series and get the predicted priced for next day/month /year , please explain
This is a common question that I answer here:
https://machinelearningmastery.com/faq/single-faq/can-you-help-me-with-machine-learning-for-finance-or-the-stock-market
Hi Jason,
Thank you for this great post.
I have a requirement of predicting receipt values for open invoices of various customers. I am taking closed invoices – whose receipt amount is used to create training data and open invoices as test data.
Below is the list of columns I will be getting as raw data
For Test Data – RECEIPT_AMOUNT, RECEIPT_DATE will be blank, depicting Open Invoices
For Training Data – Closed Invoices will have receipt amount and receipt date
CUSTOMER_NUMBER
CUSTOMER_TRX_ID
INVOICE_NUMBER
INVOICE_DATE
RECEIPT_AMOUNT
BAL_AMOUNT
CUSTOMER_PROFILE
CITY_STATE
STATE
PAYMENT_TERM
DUE_DATE
PAYMENT_METHOD
RECEIPT_DATE
It would be a great help if you can guide me which algo be suitable for this requirement. I think a multivariate method can satisfy this requirement
Thanks,
AD
I recommend the following process for new predictive modeling problems:
https://machinelearningmastery.com/start-here/#process
Hi Jason,
Are STAR models relevant here as well?
Kindest
Marius
What are star models?
Hi Jason, STAR models are Space-Time Autoregression models. I have the same question. I have a multivariate time-series with additional spatial dimension: latitude and longitude. So we need to account not only for the time lags but also for the spacial interactions. I’m trying to find a clear example in Python with no luck so far…
What are you’re inputs and outputs exactly?
Hi Jason, I have the same question as Kostyantyn.
I have a dataset on weekly detection rates of different bird species in different counties in the US. I want to build forecasting models to predict the detection rate of each species in each county, week by week.
I am thinking of using longitude, latitude as regressors for the model, so that I don’t have to run a model for each county. But I am not sure if I should treat spatial data different from numeric features. And if so, how should I do that?
I’d appreciate any advice you might have!
Thank you,
Nhat
Hi Nhat…It seems as if spatial data would be best viewed as data for classification or clustering as opposed to regression data.
In that case you would benefit from an ensemble model.
https://machinelearningmastery.com/ensemble-machine-learning-with-python-7-day-mini-course/
Hi Jason,
Thanks for this.
I want to forecast whether an event would happen or not. Would that SARMAR actually work work if we have a binary column in it?
How would I accomplish something like this including the time?
Sounds like it might be easier to model the problem as time series classification.
I have some examples of activity recognition, which is time series classification that might provide a good starting point:
https://machinelearningmastery.com/start-here/#deep_learning_time_series
Good morning
A quality cheat sheet for time series, which I took time to re-create and decided to try an augment by adding code snippets for ARCH and GARH
It did not take long to realize that Statsmodels does not have an ARCH function, leading to a google search that took me directly to:
https://machinelearningmastery.com/develop-arch-and-garch-models-for-time-series-forecasting-in-python/
Great work =) Thought to include here as I did not see a direct link, sans your above comment on thinking to do an ARCH and GARCH module.
also for reference:
LSTM time series model
https://machinelearningmastery.com/how-to-develop-lstm-models-for-multi-step-time-series-forecasting-of-household-power-consumption/
MLP and Keras Time Series
https://machinelearningmastery.com/time-series-prediction-with-deep-learning-in-python-with-keras/
Cheers and thank you
-GM
Thanks, and many more here, but neural nets are not really “classic” methods and arch only forecasts volatility:
https://machinelearningmastery.com/start-here/#deep_learning_time_series
Hi Jason,
Thank you very much this paper. I have a time series problem but i can’t find any technique for applying. My dataset include multiple input and one output like multiple linear regression but also it has timestamp. Which algorithm is the best solution for my problem?
Thanks.
I have many exmaples you can get started here:
https://machinelearningmastery.com/start-here/#deep_learning_time_series
Hi Jason,
I have a problem about time series data.
My dataset include multiple input and one output.
Normally it is like multiple linear regression but as additional has timestamp 🙁
So i can’t find any solution or algorithm.
For example: AR, MA, ARIMA, ARIMAX, VAR, SARIMAX or etc.
Which one is the best for my problem?
Thanks.
I recommend testing a suite of methods and discover what works best for your specific dataset.
one thing is there any methods to do grouped forecasting by keys or category so you have lots of forecasts , there is this on R to an extent
I’m not sure I follow, can you elaborate please?
First of all, I have read two of your books(Basics_for_Linear_Algebra_for_Machine_Learning and deep_learning_time_series_forecasting) and the simplicity with which you explain difficult concepts is brilliant. I’m using the second one to face the problem hat I present below.
I’m facing a predicting problem for food alerts. The goal is to predict the variables of the most probable alert in the next x days (also any information I could get about future alerts is really useful for me).Alerts are recorded over time (so it’s a time series problem).
The problem is that observations are not uniform over time (not separated by equal time lapses), i.e: since alerts are only recorded when they happen, there can be one day without alerts and another with 50 alerts. As you indicate in your book, it is a discontiguous time series.
The entry for the possible model could be the alerts (each alert correctly coded as they are categorical variables) of the last x days, but this entry must have a fixed size/format. Since the time windows don’t have the same number of alerts, I don’t know what is the correct way to deal with this problem.
Any data formatting suggestion to make the observations uniform over time?
Or should I just face the problem in a different way (different inputs)?
Thank you for your great work.
Sounds like a great problem!
There are many ways to frame and model the problem and I would encourage you to explore a number and discover what works best.
First, you need to confirm that you have data that can be used to predict the outcome, e.g. is it temporally dependent, or whatever it is dependent upon, can the model get access to that.
Then, perhaps explore modeling it as a time series classification problem, e.g. is the even going to occur in this interval. Explore different interval sizes and different input history sizes and see what works.
Let me know how you go.
Hello Sir,
Thank you for these information
I have a question.
I wanna know if we can use the linear regression model for time series data ?
You can, but it probably won’t perform as well as specalized linear methods like SARIMA or ETS.
I have time series data,i want to plot the seasonality graph from it. I am familiar with holt-winter. Are there any other methods?
You can plot the series directly to see any seasonality that may be present.
Hi Jason Brownlee
like auto_arima function present in the R ,Do we have any functions like that in python
for VAR,VARMAX,SES,HWES etc
Yes, I have written a few examples. Perhaps start here:
https://machinelearningmastery.com/how-to-grid-search-sarima-model-hyperparameters-for-time-series-forecasting-in-python/
Thank you, this list is great primer!
I’m glad it was helpful.
Dear Jason,
Thanks for your valuable afford and explanations in such a simple way…
What about the very beginning models of
– Cumulative
– Naive
– Holt’s
– Dampened Holt’s
– Double ESM
I would be very good to see the structural developments of the code from simple to more complex one.
Thank you very much in advance.
Best regards,
Bilal
Thanks for the suggestion.
I was told to build a bayesian regression forecast
which one of these is the best?
because I did not understand “bayesian regression” meaning
Perhaps ask the person who gave you the assignment what they meant exactly?
Thank you, i have a datset en csv format and i can open it in excel, it has data since 1984 until 2019, I want to train an artificial neural network in python i order to make frecastig or predictions about that dataset in csv format, I was thinking in a MLP, coul you help me Jason, a guide pls. Many thanks.
Sounds great, you can get started here:
https://machinelearningmastery.com/start-here/#deep_learning_time_series
Dear Jason,
Thank you for your well written quick great info.
Working on one of the banking use cases i.e. Current account and Saving account attrition
prediction.
We are using the last 6 months data for training, we need to predict customers whose balance will reduce more than 70% with one exception, as long money invested in the same bank it is fine.
Great, if you could suggest, which models or time series models will be the best options to try in this case?
I recommend this process:
https://machinelearningmastery.com/how-to-develop-a-skilful-time-series-forecasting-model/
Hi Jason,
Thank you so much for this post I learned a lot. I am a fan of the ARMAX models in my work as a hydrologist for streamflow forecasting.
I hope you can share something about Gamma autoregressive models or GARMA models which work well even for non-Gaussian time series which the streamflow time series mostly are. Can we do GARMA in python?
Thanks for the suggestion, I’ll look into the method.
This blog is very helpful to a novice like me. I have been running the examples you have provided with some changes to create seasonality for example (period of 10 as in 0 to 9, back to 0, again to 9 with randomness thrown in). Linear regression seems to be better at it than the others which I find surprising. What am I missing?
You’re probably not missing anything.
If a simpler model works, use it!
Hi Jason,
Thank you for all your posts, they are so helpful for people who are starting in this area. I am trying to forecast some data and they recommended me to use NARX, but I haven’t found a good implementation in python. Do you know other method implemented in python similar to NARX?
You can use a SARIMAX as a NARX, just turn off all the aspects you don’t need.
Hi Jason,
Thank you for all you share us. it’s very helpful.
I’m happy to hear that.
sir,
above 11 models are time series forecasting models, in few section you are discussing about persistence models…what is the difference.
Persistence is a naive model, e.g. “no model”.
Very good work. Thanks for sharing.
The line in VARMAX “The method is suitable for multivariate time series without trend and seasonal components and exogenous variables.” is very confusing.
I guess you mean no trend no seaonal but with exogenous?
Yes, fixed. Thanks!
Hi Jason,
Thanks for the post. It was great and easy to understand for a beginner in time series.
I have data of past 4 years of number of users logged in for a day .
I want to predict the number of users for a day per month.
I used ARIMA but I am getting RMSE 2749 and R2 score 60% .
Can you please suggest methods to increase the accuracy as well as RMSE.
Thanks
Perhaps try some alternate configurations for the ARIMA?
Perhaps try using SARIMA to capture any seasonality?
Perhaps try ETS?
Hello Jason, thanks for your explanation.
I have a question. What if my data is time series with multiple variables including categorical data, which model should be used for this? For example, i’m predicting The Air pollution level using the previous observation value of Temperature + Outlook (rain or not).
Thank you.
It is a good idea to encode categorical data, e.g. with an integer, one hot encoding or embedding.
Hi Menghok, did you get any luck in implementing forecasting problem when you have one more categorical variable in dataset
Hi Jason,
Thanks for the great post again, wonderful learning experience.
Do you have R codes for the time series methods you described in your article?
or Can you suggest me a good source where can i get R codes to learn some of these methods?
Thanks
Sorry, I don’t have R code for time series, perhaps you can start here:
https://machinelearningmastery.com/books-on-time-series-forecasting-with-r/
Thanks for the post. Please do check out AnticiPy which is an open-source tool for forecasting using Python and developed by Sky.
The goal of AnticiPy is to provide reliable forecasts for a variety of time series data, while requiring minimal user effort.
AnticiPy can handle trend as well as multiple seasonality components, such as weekly or yearly seasonality. There is built-in support for holiday calendars, and a framework for users to define their own event calendars. The tool is tolerant to data with gaps and null values, and there is an option to detect outliers and exclude them from the analysis.
Ease of use has been one of our design priorities. A user with no statistical background can generate a working forecast with a single line of code, using the default settings. The tool automatically selects the best fit from a list of candidate models, and detects seasonality components from the data. Advanced users can tune this list of models or even add custom model components, for scenarios that require it. There are also tools to automatically generate interactive plots of the forecasts (again, with a single line of code), which can be run on a Jupyter notebook, or exported as .html or .png files.
Check it out here:
https://pypi.org/project/anticipy/
Thanks for the note.
Hi Jason,
Thanks for this great post!
I have a question for time series forecasting. Have you heard about Dynamic Time Warping? As far as I know, this is a method for time series classification/clustering, but I think it can also be used for forecasting based on the similar time series. What do you think about this method compared to ARIMA? Do you think it will be better if I combine both two methods? For example, use DTW to group similar time series and then use ARIMA for each group?
Thanks
I don’t have any posts on the topic, but I hope to cover it in the future.
Can you please explain why you use len(data) in your predict arguments? I was using the .forecast feature for a while which is for out of sample forecasts but I keep getting an error on my triple expo smoothing. Apparently, the .predict can be used for in-sample prediction as well as out of sample. The arguments are start and end, and you use len(data) for both which is confusing me. Will this really forecast or will it just produce a forecast for months in the past?
Great question.
To predict the next or index beyond the known data.
More details here:
https://machinelearningmastery.com/make-sample-forecasts-arima-python/
Thanks for the reply! I was reading through the explanation in your linked article and it was great. Can the .predict() do multiple periods in the future like .forecast? I was using .forecast(12) for forecasting 12 months into the future.
EDIT: Dumb question- did not read until the end. If you got time, check out my stack post though: https://stackoverflow.com/questions/56709745/statsmodels-operands-could-not-be-broadcast-together-in-pandas-series . I think i found some error in statsmodels as the error is thrown in the .forecast function as it wants me to specify a frequency. I found the potential misstep by reading through the actual code for ExponentialSmoothing and it was the only part that really referenced the frequency. Its either that, or my noob is really showing.
Perhaps this code example will help:
https://machinelearningmastery.com/how-to-grid-search-triple-exponential-smoothing-for-time-series-forecasting-in-python/
Hi Jason,
Thanks for all of your awesome tutorial.
Can you please provide any link of your tutorial which has described forecasting of multivariate time series with a statistical model like VAR?
I do not have a tutorial on VAR, sorry.
OK, thanks for your reply. Hopefully, if you can manage time will come with a tutorial on VAR for us. 🙂
Hi Jason, do you cover all these models using a real dataset in your book?
I focus on AR/ARIMA in the intro book and SARIMA/ETS+deep learning in the deep learning time series book.
Hi Jason,
Thanks for the helpful tutorial. I’ve been trying to solve a sales forecast problem but haven’t been any successful. The data is the monthly records of product purchases (counts) with their respective prices for ten years. The records do not show either a significant auto-correlation for a wide range of lags or seasonality. The records are stationary though.
Among the time series models, I have tried (S)ARIMA, exponential methods, the Prophet model, and a simple LSTM. I have also tried regression models using a number of industrial and financial indices and the product price. Unfortunately, no method has led to an acceptable result. With regression models, the test R^2 is always negative.
My questions are:
* What category of problems is this problem more relevant to?
* Do you have any suggestions for possibly suitable approaches to follow for this kind of problems?
Thank you in advance.
Shabnam
Perhaps the time series is not predictable?
I guess that might be the case. I’m also guessing that maybe I don’t have sufficiently relevant explanatory variables to obtain a good regression model. Thanks for your feedback though. And, thanks again for your very helpful tutorials.
Shabnam
You’re welcome.
Hello Jason,
Do you know if there any way to randomly sample a fitted VARMA time series using the statsmodel library. Just like using the sm.tsa.arma_generate_sample for the ARMA.
I cant seem to see this how this is done anywhere.
Not off hand, sorry.
Hello Jason,
It is really nice and informative article.
I am having network data and in that data there are different parameters (network incidents) which slows down the network. From the lot of parameters one is network traffic. I have DATE and TIME when network traffic crosses a certain threshold at certain location. I need to draw a predictive model so that i can spot when network traffic is crossing threshold value at a particular location, so that i can take preventive measures prior to occurrence of that parameter at that place.
So please can you suggest me appropriate model for this problem.
Perhaps it might be interesting to explore modeling the problem as a time series classification?
Hi, Jason,
We’ve been having trouble with statsmodels’ ARIMA. It just doesn’t work and takes forever. What can you tell us about these issues? Do you know of any alternatives to statsmodels?
ThX,
Juan
Perhaps try a sklearn linear regression model directly?
Dear Jason,
Thanks for the answer.
Of course there are many regression models available in sklearn. The point is that statsmodels seems to fail miserably, both in time and accuracy. That is, with respect to their arima (family) set of functions.
Question is if you know an alternative python library providing that missing functionality. R works well. Mathematica even better. At the moment, my students are working on interfacing with R, since we have not found a sound Python library for arima.
Have a good one.
Juan
I see, good question.
I have found the statsmodels implementation to be reliable, but only if the data is sensible and only if the order is modest.
I cannot recommend another library at this time. R may be slightly more reliable, but is not bulletproof.
Thanks, Jason.
We ended up running everything on Mathematica. It works way better. A more robust python library for ARIMA models is needed.
Best,
Juan
Nice work!
I agree.
In time series classification when I am plotting it is showing day wise data . For an instance if i am taking data of past 1 month and applying Autoregression time series classification on it then i am not able to fetch detailed outputs from that chart. From detailed output I mean that 1 incident is occuring number of times in a day so I want visibility in such a way so that everytime an incident occur it should be noticed on that plot.
Is there any precise way by which I can do it.I would be really thankful if you would help me.
Thank you
Perhaps resample the data to daily before or after modeling, depending on whether you want model data this way or only view forecasts this way.
I can’t seem to make VAR work? Gives me a lot of errors?
—————————————————————————
ValueError Traceback (most recent call last)
in ()
11 # fit model
12 model = VAR(data)
—> 13 model_fit = model.fit()
14 # make prediction
15 yhat = model_fit.forecast(model_fit.y, steps=1)
D:\Users\Berns\Anaconda3\envs\time_series_p27\lib\site-packages\statsmodels\tsa\vector_ar\var_model.pyc in fit(self, maxlags, method, ic, trend, verbose)
644 self.data.xnames[k_trend:])
645
–> 646 return self._estimate_var(lags, trend=trend)
647
648 def _estimate_var(self, lags, offset=0, trend=’c’):
D:\Users\Berns\Anaconda3\envs\time_series_p27\lib\site-packages\statsmodels\tsa\vector_ar\var_model.pyc in _estimate_var(self, lags, offset, trend)
666 exog = None if self.exog is None else self.exog[offset:]
667 z = util.get_var_endog(endog, lags, trend=trend,
–> 668 has_constant=’raise’)
669 if exog is not None:
670 # TODO: currently only deterministic terms supported (exoglags==0)
D:\Users\Berns\Anaconda3\envs\time_series_p27\lib\site-packages\statsmodels\tsa\vector_ar\util.pyc in get_var_endog(y, lags, trend, has_constant)
36 if trend != ‘nc’:
37 Z = tsa.add_trend(Z, prepend=True, trend=trend,
—> 38 has_constant=has_constant)
39
40 return Z
D:\Users\Berns\Anaconda3\envs\time_series_p27\lib\site-packages\statsmodels\tsa\tsatools.pyc in add_trend(x, trend, prepend, has_constant)
97 col_const = x.apply(safe_is_const, 0)
98 else:
—> 99 ptp0 = np.ptp(np.asanyarray(x), axis=0)
100 col_is_const = ptp0 == 0
101 nz_const = col_is_const & (x[0] != 0)
D:\Users\Berns\Anaconda3\envs\time_series_p27\lib\site-packages\numpy\core\fromnumeric.pyc in ptp(a, axis, out, keepdims)
2388 else:
2389 return ptp(axis=axis, out=out, **kwargs)
-> 2390 return _methods._ptp(a, axis=axis, out=out, **kwargs)
2391
2392
D:\Users\Berns\Anaconda3\envs\time_series_p27\lib\site-packages\numpy\core\_methods.pyc in _ptp(a, axis, out, keepdims)
151 def _ptp(a, axis=None, out=None, keepdims=False):
152 return um.subtract(
–> 153 umr_maximum(a, axis, None, out, keepdims),
154 umr_minimum(a, axis, None, None, keepdims),
155 out
ValueError: zero-size array to reduction operation maximum which has no identity
Sorry, I’m not sure about the cause of your error.
Perhaps confirm statsmodels is up to date?
By the way Doc Jason, I just bought your book today this morning. I have a huge use case for a time series at work I can use them for.
Thanks, you have a lot of fun ahead!
I’m here to help if you have questions, email me directly:
https://machinelearningmastery.com/contact/
Darn indentions! Now code works.
# VAR example
from statsmodels.tsa.vector_ar.var_model import VAR
from random import random
# contrived dataset with dependency
data = list()
for i in range(100):
v1 = i + random()
v2 = v1 + random()
row = [v1, v2]
data.append(row)
# fit model
model = VAR(data)
model_fit = model.fit()
# make prediction
yhat = model_fit.forecast(model_fit.y, steps=1)
print(yhat)
I’m very happy to hear that.
More on how copy code from posts here:
https://machinelearningmastery.com/faq/single-faq/how-do-i-copy-code-from-a-tutorial
Jason , need one clarification , for a SARIMAX model or in general in timeseries model, should I use .forecast or .predict ? I have got some differences in result using this two . Please suggest should we use .forecast for train and test ( to get the model fitted values) or should we use .predict for train and test ( to get the model fitted values) . Also for future forecast what need to be used ? Please suggest.
Either, they both do the same thing.
Hi Jason,
GREAT article. I’ve been looking for an overview like this. I do have a side question about a project I’m working on. I’ve got a small grocery distribution company that wants weekly forecasts of its 4000+ different products/SKU numbers. I’m struggling with how to produce a model that can forecast all these different types of time-series trends and seasonalities. ARIMA with Python doesn’t seem like a good option because of the need to fine tune the parameters. I would love you’re feedback on the following ideas:
1. Use facebooks prophet for a “good enough” model for all different types of time series trends and seasonalities.
2. Separate the grocery items into 11-20 broader category groups and then produce individual models for all those different categories that can then be applied to each individual sku within that category.
LIMITATIONS:
1. I need this process to be computationally efficient, and a for-loop for 4000 items seems like a horrible way to go about this.
QUESTIONS:
1. Do i have to model sales and NOT item quantities and why is that?
2. Is there a way to cluster the items with unsupervised learning just based off their trend and seasonality? Each item does not have good descriptor category so I don’t have much to work with for unsupervised clustering techniques.
Would love to hear any thoughts you have about this type of problem!
Thanks for ALL the helpful content.
Kristen
This will give you some ideas of working with multiple SKUs:
https://machinelearningmastery.com/faq/single-faq/how-to-develop-forecast-models-for-multiple-sites
Explore groupings by simple aspects like product category, average sales, average gross in an interval, etc.
Hi Jason,
many thanks for the many informative articles.
I have a simulation model of a system which receives a forecast of a time series as input.
In my scientific work I would like to examine how the performance of the simulation model behaves in relation to the accuracy of the forecast. So I would like to have forecasts between 80% (bad prediction) and 95% (very good prediction) for my time series.
How would you recommend me to generate such pseudo predictions that are meaningful for real predictions? Maybe adjusting a mediocre prediction in both direction?
Thanks and best regards
Anne
Not sure I follow, sorry.
What specific problem are you having exactly?
Excuse me, it was also written a bit misleading.
My problem is that my simulation model receives forecasted time series as input, from which I know from publication papers that they are 95% predictable with a lot of effort and large amounts of training data.
Simple predictions like the ones I make, however, only reach about 80%. Nevertheless, I would like to analyze for my scientific work how much the quality of the simulation model depends on the quality of the forecasts. For example, to say whether the effort for good forecasts is worth it.
Is there a way to make pseudo forecasts of time series for this purpose? As a basis the actual data and the forecast data are available with 80% accuracy.
Thanks and best regards
Anne
Yes, why not contrive forecasts as input and contrive expected outputs and control the entire system – e.g. a controlled experiment.
I believe that is required for what you’re describing.
If you’re new to computational experimental design, perhaps this book will help:
https://amzn.to/2AFTd7t
Thank you very much. I’ll follow up on that tip.
Part of my question also aimed at how I best proceed when contriving forecasts, as they should be as representative as possible.
The simplest would be to add normal distributed noise to the actual data, but I don’t know if that would be representative if the error was only caused by white-noise.
What is your opinion on this issue?
A linear model (e.g. linear dependency) for generating forecasts would be the way I would approach it.
Or a controllable simplification of the data you will be working with for real, e.g. trends/seasonality/level changes/etc.
Hey Jason, I have a requirement
I have to perform a regression on a data set that has the following data
Age of Person at time of data collection
Date
cholesterol Level
HDL
Event which can be death etc.
Need to find out if there are any patterns in the Cholesterol Level or the HDL Levels prior to the death event.
I’m trying to do this using Jupyter Notebook. Really appreciate your help sincerely.
Do you have any examples that would be close ?
Sounds like a great problem, perhaps start with some of the simpler tutorials here:
https://machinelearningmastery.com/start-here/#timeseries
Hey Jason! I am working on supply elasticities for crops in California, I am new using python your post has been incredible useful, I have one question for my model I need to develop a time series dynamic model particularly an Autoregressive Distributed Lag (ARDL) Model, do you have any material related to how perform this type of models on python?
Thanks!
Sorry, I have not seen this model in Python.
If you can’t locate, perhaps you can implement it yourself? Or call a function from an implementation in another language?
Why do you need that specific model?
Hi Jason,
Actually I am very impatient and did not read all the comments but thot that you already knew them 🙂
I need to make a hierarcial and grouped time series in Python.
I need to predict daily & hourly passenger unit by continent or vessel name. I investigated but did not find a solution, could you pls direct me what can I do.
Below are my inputs for the model.
Date & Time Vessel Name Continent Passenger Unit
Thank you,
Perhaps start by fitting one model per “vessel”?
Then later perhaps try fitting models across vessels?
For simple linear models you can start here:
https://machinelearningmastery.com/start-here/#timeseries
For advanced models you can start here:
https://machinelearningmastery.com/start-here/#deep_learning_time_series
Hi Jason,
While forecasting in VAR or VARMAX I’ m not getting the future timestamp associated with it?
Won’t model provide us the future timestamp?
You will know it relative to the end of your training data.
Hi Jason,
For the simple AR model yhat values is it possible to get the threshold?
data = [x + random() for x in range(1, 100)]
model = AR(data)
model_fit = model.fit()
yhat = model_fit.predict(len(data), len(data))
Thanks
What threshold?
The forecast, lowerbound and upperbound?
I think you might be referring to the prediction interval.
This gives an example:
https://machinelearningmastery.com/time-series-forecast-uncertainty-using-confidence-intervals-python/
Hi
Looking at the AR.predict documentation here:
http://www.statsmodels.org/dev/generated/generated/statsmodels.tsa.ar_model.AR.predict.html#statsmodels.tsa.ar_model.AR.predict
With the given parameters as params, start, end and dynamic, I can’t correspond them to the method call you have as:
yhat = model_fit.predict(len(data), len(data))
For which exact documented parameters do you pass over len(data) twice?
Thanks.
That makes a single step prediction for the next time step at the end of the data.
Alternately, you can use model.forecast() in some cases/models.
Thanks Jason. This post is really awesome.
I see that you are using random function to generate the data sets. Can you suggest any pointers where I can get real data sets to try out these algorithms. If possible IoT data sets would be preferable.
Yes, right here:
https://machinelearningmastery.com/faq/single-faq/where-can-i-get-a-dataset-on-___
Hey
It is really great to see all this stuff.
Can you please explain me:
which algorithm should i use when i want to predict the voltage of a battery:ie…
battery voltage is recorded from 4.2 to 3.7 is 6 months which algorithm will be the best one to predict the next 6 months
Good question, this will help:
https://machinelearningmastery.com/how-to-develop-a-skilful-time-series-forecasting-model/
Buenas tardes, se podría utilizar el walk fordward validación para series de tiempo multivariado ? es decir una variable Target y varias variables input, y k te haga la técnica de wall fordward validación?? esa es mi pregunta, saben de algún link o tutoría
Yes, the same walk-forward validation approach applies and you can evaluate each variate separately or all together if they have the same units.
Hi,
Have you ever evaluated the performance of Gaussian processes for time series forecast? Do you have any material on this (or plan to make a post about this)?
It’s probably not a “classical” method though.
Thanks!
I do not. I hope to cover it in the future.
Hi Jason,
I’m currently working on a sales data set which has got date, location, item and the unit sales. Time period is 4 yrs data, need to forecast sales for subsequent year -1 month(1-jan to 15 jan). Kindly suggest a model for this multivariate data. Another challenge which i’m currently facing is working with google collab with restricted RAM size of 25 GB and the data set size is 1M rows.
Please suggest a good model which will handle the memory size as well.
Note: I have change the datatype to reduce the memory, still facing the issue
I recommend this process:
https://machinelearningmastery.com/how-to-develop-a-skilful-time-series-forecasting-model/
Can you briefly explain about Fbprophet, it’s pros and cons etc.
Thanks for the suggestion.
I have a tutorial on the topic written, it will appear soon.
Hi Jason,
All these methods seem to focus on two data features:
– evenly spaced time series;
– continuous data.
As it is easy to guess, I have a time series which is both irregular and categorical.
Mine is sensor data. These sensors only detect the presence of a person in a room. Each entry gives me the start time, duration and room. This is for monitoring one person with health issues as they move through their flat.
None of the above methods — fantastically depicted, by the way! — is a glove to this hand. I need to detect patterns, trigger alerts on anomalies and predict future anomalies.
Any suggestions?
Yes, see this:
https://machinelearningmastery.com/faq/single-faq/how-do-i-handle-discontiguous-time-series-data
I have error executing VARMAX
“ValueError: endog and exog matrices are different sizes”
Perhaps confirm that the inputs to your model have the same length.
Hello,
Are you sure the code is well-formatted ?
”
for i in range(100):
v1 = random()
“
Fixed, thanks!
Hi Dr.Jason,
I have a question about time-series forecasting.I have to predict the value of Time-series A in the next few hours.And at the same time,the Time-series B have an influence on A.I have two ideas :
1.X_train is time T and Y_train is the value of A_T and B._T
2 X_train is the value of (A_n, A_n-1, A_n-2,…A_n-k, B_n, B_n-1, B_n-2,…,B_n-k) ,Y_train is (A_n+1, A_n+2,…,A_n+m)
Which is better?
Test and compare results.
It might be easier to model it as a multivariate input time series forecasting problem.
The tutorials here will help:
https://machinelearningmastery.com/start-here/#deep_learning_time_series
Thanks! It definitely helps me a lot !
I’m happy to hear that.
Hope for your suggestions.
Thanks and best regards.
Lanet
You’re welcome.
Hi Jason
I get numerous warnings for some of the models (e.g. robustness warnings, and model convergence warnings, etc.) Do you get the same?
Thanks,
Dean
Sometimes. You can ignore them for now, or explore model configs/data preps that are more kind to the underlying solver.
Thanks Jason.
One other thing – do you know where I could find an explanation on how the solvers work (relating to the theory) ? Since I’ll need to know what config/parameter changes will do.
I did access the links you posted (e.g. statsmodels.tsa.ar_model.AR API) , but I don’t find these very descriptive, and they are without code examples.
Yes, a textbook on convex optimization, wikipedia, the original papers, and maybe the scipy API.
Hi Jason Brownlee,
I appreciate that you posted this explanation about the models.
I would like to apply the algorithms to my problem.
I was in doubt as to how to extract the predicted value and the current value of each prediction. I would like to calculate the RMSE, MAPE, among others.
You’re welcome.
You can make a prediction by fitting the model on all available data and then calling model.predict()
hello Jason, thank you for your post. it is so interesting.
In fact i have a question. I have many time series Trafic per region. I don’t want ot have 40 models to forecast trafic for each region.
Do you have an idea how can i use one model to predict trafic for all regions so one model to a multiple time series?
Thank you in advance
Yes, I list some ideas here:
https://machinelearningmastery.com/faq/single-faq/how-to-develop-forecast-models-for-multiple-sites
thank you for your answer. for my project i want to have one model to predict all the values for all cities.
I don’t know wich types of algorithms can i use that accept multiseries can you help me with this thank you
Follow this framework to discover the algorithm that works best for your specific data:
https://machinelearningmastery.com/how-to-develop-a-skilful-time-series-forecasting-model/
thank you for your response. In fact do you have an examples of link or article with python that forecast with multiple independant time series
my data is in this format:
date ville X(predicted value)
01/01/2016 A 2
01/02/2016 A 4
01/02/2016 A 5
01/01/2016 B 10
01/02/2016 B 11
01/02/2016 B 13
Yes many – search the blog, perhaps start with the simple examples in this tutorial:
https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/
I want to fit a curve on time series data. Will this methods work for it or it can be useful for prediction only?
Perhaps use the methods here:
https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.curve_fit.html
Dear Jason,
Do you have a blog where you forecast univariate time series using a combination of exponential smoothing and neural network?
I mean where we give lags of time series to exponential smoothing model and the output of exponential smoothing will be the input of a neural network?
Thank you
I don not, sorry.
Hi Jason,
Could you also please review the BATS and TBATS models? (De Livera, A.M., Hyndman, R.J., & Snyder, R. D. (2011), Forecasting time series with complex seasonal patterns using exponential smoothing, Journal of the American Statistical Association, 106(496), 1513-1527.) Like SARIMA, they are quite good to deal with seasonal time series. And I am wondering if you have any plan to review Granger causality which is often used to find time series dependencies from multivariate time series? Thank you.
Thanks for the suggestion!
hello
i have a dtaset of many vehicle which ask me to predict output by taking 5 timeseries input and i have to do this for all vehicle almost 100 vehicle to improve accuracy of model,can u help me
This may give you some ideas:
https://machinelearningmastery.com/faq/single-faq/how-to-develop-forecast-models-for-multiple-sites
Hello Sir,
The SARIMAX model you explained here is for single step. Can you let me know, how to do it for multi step ( with steps = 96 in future) forecasting.
Specify the number of steps required in the call to the forecast() function.
Hi Jeson,
Great info. Can you suggest the algorithms which cover multivariate with trend and seasonality?
Neural nets:
https://machinelearningmastery.com/start-here/#deep_learning_time_series
Hello, which one of the methods you shown is applicable to very limited data? Like 5 data pts only, univariate. There’s no way to increase it because my subject has just began operating 5 yrs ago and this data should be forecasted annually.
Thanks
With limited data, usually we do the reverse: Assume the model/method, and fit with the limited data as much as we can, and use it for forecasting.
hey, Jason. there are some changes on the statsmodels library. e.g. on the arima is now
>> statsmodels.tsa.arima.model
Thanks.
Hi Jason. Thank you for the great post.
It is not very clear to me what is the difference between a multivariate and an exogenous time series
Can you please clarify that?
Thank you
Thanks.
My understanding:
Multivariate refers to multiple parallel input time series that are modeled as sych (lags obs)
Exogenous variables are simply static regressors that are added to the model (e.g. linear regression).
statsmodels.tsa.arima_model.ARMA:
Deprecated since version 0.12: Use statsmodels.tsa.arima.model.ARIMA instead…
Thanks. I will schedule time to update it.
Jason, thanks for all your posts!
I am a bit confused on When to use VAR and when to use ARIMAX? Any explanation would be much appreciated….
VAR for when you have multivariate input.
ARX (ARIMAX), the “X”, for when you have a univariate time series with exogenous variables:
https://en.wikipedia.org/wiki/Exogenous_and_endogenous_variables
If you’re unsure, try both and one will be a natural fit for your data and another won’t. It might be that clear.
Hi Sir, Thanks for the amazing tutorial
1. For the SARIMAX seasonal component “m” should it be different than 1 ,get an error when 1 is used .
2. In the predict function why we pass the length as 99 if we want next prediction
It might depend on your data.
You specify the index or date you want to forecast in the predict() function.
I was curious to know when to use which model? And conditions to be met before using a model. Can I get some points on this?
Determine if you have multiple or single inputs and/or multiple or single outputs then select a model that supports your requirements.
As far as I understand
Single input : Univariate Model
Multiple Input : Multivariate Model
But I guess where each of the models, in which scenario fits exceeds the scope of this topic i.e cheat sheet.
However in short words(since I am a beginner to time-series) can you tell me where each of these model fit in what scenario, so I can have a basic understanding of where to use each model.
However I am having IoT data project in my company. where you get to see monitored data in timestamp. I am being asked to make Pv predictions. Can you suggest a model as well that would be great :
Datetime Pv Min Max Avg
01-01-2019 07:00 20.1 20.1 20.1 20.1
01-01-2019 07:03 20 20 20 20
Yes, the example for each model show’s the type of example where it is appropriate.
Hi, I am trying to work on sarimax but I am still confused on how to set the order and seasonal order?
Use a grid search:
https://machinelearningmastery.com/how-to-grid-search-sarima-model-hyperparameters-for-time-series-forecasting-in-python/
Hi Sir,
Thanks for your great tutorial.
Are there any other references for multivariate time-series forecasting with ARIMA and other classical time-series approaches?
Thanks in advance.
The above VAR methods are a good place to start.
Hi Jason,
I have a question on lag orders, in case of univariate techniques, we can decide p and q values using autocorrelation and partial autocorrelation. But, in the case of multivariate techniques like VAR, VMA, and VARMA how we can decide these lag orders as we have multiple parallel series?
Thanks,
Ajay
It can get tricky. Perhaps fall back to a grid search.
Hello Prof Brownlee
I really need help in respect to my work.
I am trying something on modelling structural breaks in futures and wants to try something unique with Python.
How can you help me, I will like to get through with your email please
This process may help you work through your project:
https://machinelearningmastery.com/start-here/#process
Hi Bob,
How did you go with your structural breaks modeling? I am interested to do the same but is getting stuck looking for a way through.
Hi Jason,
Can you demonstrate using Dynamic Regression model? I need help with Dynamic Regression model for my project.
where nt is ARMA process
yt = a + ν(B)xt + nt
Thank you so much for all your great tutorials.
Thanks for the suggestion.
Thanks Jason. Here is Rob’s link. Appreciate if you can do it in python.
https://robjhyndman.com/uwa2017/3-2-Dynamic-Regression.pdf
Thanks for the suggestion.
Hi Jason
Thank you for the great post.
I have a time series showing departure date, booking date and number of booked passengers. It has both trend and seasonal components and I am working on passengers forecast by considering departure week day, week, month and prior days ( departure date minus booking date). What is your model suggestion to implement? MLP?
thanks
I recommend testing a suite of different data preparation methods, different modeling algorithms and different configurations in order to discover what works best for your dataset:
https://machinelearningmastery.com/how-to-develop-a-skilful-time-series-forecasting-model/
Hi Dr. Jason,
I have a multivariate time seies problem formulation that seeks to forecast endogenous variable y, as a function of exogenous variables, x1,x2,…,xn. I thought VARMAX method is applicable in this situation, but it seems VARMAX() can only take multiple exogenous varaibles when I tried to implement the method. However, one edogenous or response variable is needed in my case.
I need to create a baseline model for the stated multivariate problem. I will appreciate your thought in this situation.
Perhaps you can use a persistence model as the baseline.
Then maybe use a linear regression model on window transform of the data as a baseline linear model.
Thanks, Dr. Jason.
I undertand the aspect of the persistence model that would based on univariate data – the respose variable.
For the second option, I guess you meant fiting a linear regression model on a suppervised learning(window) transform of the response variablle, right?
Yes, this can help you prepare the data:
https://machinelearningmastery.com/time-series-forecasting-supervised-learning/
And this:
https://machinelearningmastery.com/convert-time-series-supervised-learning-problem-python/
Thanks!
Revised: “multiple endogenous variables”
Hi Dr. Jason,
I have a multivariate time seies problem formulation that seeks to forecast endogenous variable y, as a function of exogenous variables, x1,x2,…,xn. I thought VARMAX method is applicable in this situation, but it seems VARMAX() can only take multiple endogenous varaibles when I tried to implement the method. However, one edogenous or response variable is needed in my case.
I need to create a baseline model for the stated multivariate problem. I will appreciate your thought in this situation.
Hello Dr. Jason,
I’m a trainee and are now executing my literature review with the goal to find the right method for solving my internship problem. During my search for relevant articles I come across a lot of your blogs and read some of them to find the method suitable for my problem. However, I still haven’t found it yet. Maybe you can direct me in the right way?
From the department demand planning at my internship organization I get a monthly forecast of the amount of products (medical devices) that will be ordered. I have been instructed to break this monthly forecast down in daily forecast. So, I’m looking for a (machine learning) method that distributes the monthly forecast into daily buckets. Am I looking in the right direction if I’m searching for Time Series forecasting algorithms? I find that a lot of these algorithms can be used to determine the monthly forecast and not the distribution of the monthly forecast into daily buckets.
Can you help me putting me in the right direction for searching algorithms for distribution?
Sorry if the description of my problem and question is a bit hazy 🙂
Already thanks for your time.,
Jimmy,
The netherlands
I would recommend testing a suite of different algorithms on your problem in order to discover what works.
Perhaps this framework will help:
https://machinelearningmastery.com/how-to-develop-a-skilful-time-series-forecasting-model/
Thank you for sharing all of these time series examples. I would like to do a weighted time series forecast similar to Holt Winter’s Exponential Smoothing but with multivariate data. What would you recommend for this? The objective is to have a time series forecast with greater weight to more recent data
The model will learn we weighting required to give the best forecast. No need to do this manually.
Thanks what about in the extreme case now with Covid where there are only several months of post-Covid data vs years of less useful historic data?
No difference.
hi Jason, How can Time Series Analysis be done with Categorical Variables?
Right now I’m doing a project where I have a database of network infrastructure alames in a space of 3 months, in this case a time series problem.
Most of the time series analysis tutorials/textbooks I’ve read about, be they for univariate or multivariate time series data, usually deal with continuous numerical variables.
I currently have a problem at hand that deals with multivariate time series data, but the fields are all categorical variables. Hence, I was wondering if there is any way to use the standard time series analysis techniques (such as ARIMA, ARMA etc.)
Specifically, my data is a stream of alert data, where at each time stamp, information such as the alert monitoring system, the location of the problem etc. are stored in the alert. These fields are all categorical variables.
What advice can you give me regarding the problem, and which ML algorithms are most recommended for this case?
Encode them as numbers, then use any model you like, e.g. linear, machine learning, deep learning, etc.
Popular methods include ordinal encoding, one hot encoding, embedding.
Hi Dr Jason,
I want to do forecasting of telecom data.
The timeseries data has records in millions ranging from 1 million to 10 million.
Should i first normalize the timeseries data using log, min-max or any other method ?
Also, is it better to remove/replace outliers in time series data ?
Thanks & Regards,
Rajiv Kamrani
Perhaps work with a small sample of the data.
It can be a good idea to scale, transform data, remove outliers, etc. try it and see if it improves model performance.
Hi Jaron, there seems to be a bug in your SARIMA example, we get at line
model = SARIMAX(data, order=(1, 1, 1), seasonal_order=(1, 1, 1, 1))
Traceback (most recent call last):
File “”, line 1, in
File “C:\Users\User\AppData\Local\Continuum\anaconda3\envs\Py37_Tf2_Ke2\lib\site-packages\statsmodels\tsa\statespace\sarimax.py”, line 332, in __init__
missing=missing, validate_specification=validate_specification)
File “C:\Users\User\AppData\Local\Continuum\anaconda3\envs\Py37_Tf2_Ke2\lib\site-packages\statsmodels\tsa\arima\specification.py”, line 303, in __init__
raise ValueError(‘Seasonal periodicity must be greater’
ValueError: Seasonal periodicity must be greater than 1.
I can confirm the code operates correctly, I just re-ran it.
I have some suggestions for you:
https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me
Hi Jason,
Thank you for sharing all types of time series examples.
In my project work I want to predict vehicle speed considering different inputs such as current speed of vehicle, traffic light states, distance between two vehicles etc. Which approach shall I follow as it has trends and seasonality? its kind of multiple input and singe output scenario.
Thank you!!
You’re welcome.
I recommend testing a suite of different algoritms in order to discover what works well or best for your specific dataset.
Thanks for your reply. Can I implement SARIMA for multiple imput and single output or do I have to use VAR or VARMA?
Yes, it would be SARIMAX and additional input would be considered exogenous input variables.
Hi Jason
Thank for the tutorial. I read your book on time series forecasting. Usually, time column is dropped and then, train and test the model. In my work i want to predict the “time”. So i can’t remove “time” column. The time is represented as “timestamp”. Can you please explain me well how to preprocess the data in order to forecast the “time”.
I know how to prepare the data and split the dataset, it’s well explained in your book
Cordially,
Thanks.
Perhaps you can frame it as a classification ask, e.g. will an event occur in the next time step: 0 or 1.
Hi Jason:
Interesting classical tutorials (via statistics analysis) that many times are better than advance Neural ML models such as MLP, LSTM.
But this “statsmodels” library are not uniform on its terminology methods for each type of model, sometimes requiring “.predict()” but other times requiring “.forecast( )”, with different arguments on it (sometimes limits of length prediction, sometimes needing to declare the variable to predict (but other not) …so not uniformity on them on the contrary when you use Keras methods you know which arguments to declare.
in addition to it the exogenous variables as far I understand you need to provide on both model definition and the “.predict()” methods…but at the end the prediction is for one feature (univariate) but on multivariate are so many outputs as multivariate.
Also when you use multivariate models (VAR, VARMA,..) you also get multiple output (as many as multivariate) … I thought you can uses to two features timeseries to produce a complete new one output timeseries …as i thought it is allowed on keras MLP, LSTM, etc…So my intuition fail 🙁
The terminology for orders (p,d,q..P,Q,D, mm) it is also confused. I think finally they refer to lags o timesteps required for each order …
Astute observations!
Agreed, the library has been a bit of a mess for years.
Hello Jason,
Do you have any updated book about classical methods on python? As mentioned, statsmodels are a mess through the years and in my university project i want to use Simplexponentialsmooth, Holtwinters and ARIMA to predict my data but i’m having a lot of problems about statsmodels.
A new library can help a lot as well
I do cover SARIMAX and ETS in this book:
https://machinelearningmastery.com/deep-learning-for-time-series-forecasting/
1) how to check stationarity for multivariate time series analysis and if i use ‘Johansen test’ for multivariate time series how i conclude the results.
2)can you please update here if you have python code for johansen test or any other test is there to check stationarity for multivariate?
3) can i use ACF & PACF values for VARMAX models or it is only supports for univariate models.
Sorry, I cannot help you with these questions, I don’t have tutorials on these topics.
I am working in petrochemical sector and want to forecast the production of ethylene in cracker furnace based upon certain no. of chemical parameters like pressure, temperature ,composition of the gas with time, which models should I try to use, like LSTMs or classical time series methods?
Perhaps try a suite of techniques and discover what works well or best for your specific dataset.
How can we use each of the 11 methods? I mean under what circumstances or conditions should one apply AR, MA, ARMA, ARIMA, or SARIMA method.
Each section gives some small idea of when to use it, e.g. when you have a trend or seasonality or exog variable, etc.
Hey Jason,
One question wavelets are getting a hot topic.
You have anything on this topic
Not at this stage.
Doctor Brownlee, good morning.
Thank you very much for the books you have written and the publications you make.
I was wondering if you know any library or function that calculates the best values for the parameters (order and seasonal_orden), when using SARIMAX.
Thank you.
You can use a grid search, see this example:
https://machinelearningmastery.com/how-to-grid-search-sarima-model-hyperparameters-for-time-series-forecasting-in-python/
Doctor thank you very much!
You’re welcome.
Thank You for the post
You’re welcome.
Hello Jason,
I have to backcast values of a time series. Is it enough to invert the time series and procede with forecasting?
I was wondering if there is some python code related to backcasting.
Thank you.
Perhaps try a few methods and discover what works well or bes. Compare everything to simple persistence model.
Hi Jason,
Can we use categorical variables(binary) as exogenous variable in Sarimax. I am trying to forecast Premium collection and i want to use covid impact as an exogenous variable.
Just anted to know your thoughts if this can be done. I have weekly collectin data for last 4 years.
It should work as long as you encoded it as a number (e.g. 0 or 1). But you should aware that the SARIMAX model still consider that as a real number. If you are predicting these categorical variable, the model may give you values like 0.7 instead of 0 and 1.
Hi Jason,
How do we forecast time series which is having less than 10 observations ? I have time series for every lab roviding volumes for customer claiming for insurance for a particular disease. Some of these id’s have volumes for all the quarters in 4 years and few id’s have only one insurance claim. So how do I handle this short time series?
Basically your hands are tied under such small amount of data. I know the actuarial have their own tricks to supplement that. For example, they would consider certain thing are in a particular statistical model. Or they would consider disease A and disease B can be mixed together so you can borrow the data from one to another. I do not have the domain knowledge to give more detail but think out of the box is the only way you can do here.
Hi jason,
I’m trying to solve a problem using timeseries forecasting. For that, i used the n-beats model and the prediction is ok. However, i need to know how to define an extreme event after i forecast the results.
is there a library already for that? is it a data distribution dependent? as my data doesn’t fall under one of the famous distributions.
Thanks and would be grateful if you replied.
That’s really depends on the data distribution. If your data doesn’t fall into anything, try to skew your data a bit (e.g., take logarithm, take square of it, etc.) and find the best match. That’s what we need to do in the real world problems.
Hello Dr. Jason,
Greetings.
I want ro work on a prediction model for number of users in a Telecom Network. Which model do you think would be best for such a scenario?
Best Regards
Hi Prashant…While I cannot speak directly to your particular application, the following may be of interest to you:
https://machinelearningmastery.com/when-to-use-mlp-cnn-and-rnn-neural-networks/
Hi Jason
Thank you for this amazing content. Will you kindly help me with my question?
I’m currently working with data obtained from a test conducted, which contains a lot of values for parameters such as force, time, torque, velocity, etc. I’m still confused about which model to proceed with. This brings me to my question.
How should I select the most suitable model for my time series prediction task?
Hi Aravind…Thank you for the kind words and feedback!
I would start with the following two:
https://machinelearningmastery.com/sarima-for-time-series-forecasting-in-python/
https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/
Hi Jason
This is great and helped me a lot
i have a request can you do this for autoML models
Thank you for the feedback and suggestion!
when we use machine learning models for time series data, we use ARIMA first and then its AR terms as independent variables. Do we have any other criteria that we can use directly by avoiding the use of ARIMA models AR and MA terms, to analyse univariate time series data?
Hi Shahfahad…If you are dealing with multivariate data, you should investigate CNN and LSTM models.
Great Job! I have a question: Do you think which Moder should be the best for Lottery Prediction? Thanks.
Hi Andy…Sorry, I cannot help you develop a predictive model for the lottery.
The draw of numbers in a lottery are random events, by design.
A draw of random events cannot be predicted.
Even if you have a large dataset with hundreds or thousands of previous lottery draws.
I can show u how wrong u are … u just need to be creative to transform history drawing database to pattern data base in which u don’t predict winning numbers but u analyze pattern occurrence by confidence of appearance. Now u cant do this “manually” but with reinforcement learning by back propagation to correct used NN’s for each new drawing (and its not time series). This way u can achieve up to 80% plus confidence analytically (not always it’s still % of confidence) and u can pick the right pattern (very important) if u consider other database properties behavior (behaves no random but structured). U can’t achieve that the easy way. U have to dive into coding very deep with a lot of testing, u just have to learn to code at phD level. Maybe i said to much as u re phD in this field already.
Hi Cris…Thank you for your feedback! I agree that deep reinforcement learning is an emerging and exciting area of research! Let us know what you find out from your models.