Last Updated on

Selecting a time series forecasting model is just the beginning.

Using the chosen model in practice can pose challenges, including data transformations and storing the model parameters on disk.

In this tutorial, you will discover how to finalize a time series forecasting model and use it to make predictions in Python.

After completing this tutorial, you will know:

- How to finalize a model and save it and required data to file.
- How to load a finalized model from file and use it to make a prediction.
- How to update data associated with a finalized model in order to make subsequent predictions.

Discover how to prepare and visualize time series data and develop autoregressive forecasting models in my new book, with 28 step-by-step tutorials, and full python code.

Let’s get started.

**Update Feb/2017**: Updated layout and filenames to separate the AR case from the manual case.**Updated Apr/2019**: Updated the link to dataset.**Updated Aug/2019**: Updated CSV file loading.

## Process for Making a Prediction

A lot is written about how to tune specific time series forecasting models, but little help is given to how to use a model to make predictions.

Once you can build and tune forecast models for your data, the process of making a prediction involves the following steps:

**Model Selection**. This is where you choose a model and gather evidence and support to defend the decision.**Model Finalization**. The chosen model is trained on all available data and saved to file for later use.**Forecasting**. The saved model is loaded and used to make a forecast.**Model Update**. Elements of the model are updated in the presence of new observations.

We will take a look at each of these elements in this tutorial, with a focus on saving and loading the model to and from file and using a loaded model to make predictions.

Before we dive in, let’s first look at a standard univariate dataset that we can use as the context for this tutorial.

### Stop learning Time Series Forecasting the *slow way*!

Take my free 7-day email course and discover how to get started (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

## Daily Female Births Dataset

This dataset describes the number of daily female births in California in 1959.

The units are a count and there are 365 observations. The source of the dataset is credited to Newton (1988).

Download the dataset and place it in your current working directory with the filename “*daily-total-female-births.csv*“.

We can load the dataset as a Pandas series. The snippet below loads and plots the dataset.

1 2 3 4 5 6 |
from pandas import read_csv from matplotlib import pyplot series = read_csv('daily-total-female-births.csv', header=0, index_col=0) print(series.head()) series.plot() pyplot.show() |

Running the example prints the first 5 rows of the dataset.

1 2 3 4 5 6 |
Date 1959-01-01 35 1959-01-02 32 1959-01-03 30 1959-01-04 31 1959-01-05 44 |

The series is then graphed as a line plot.

## 1. Select Time Series Forecast Model

You must select a model.

This is where the bulk of the effort will be in preparing the data, performing analysis, and ultimately selecting a model and model hyperparameters that best capture the relationships in the data.

In this case, we can arbitrarily select an autoregression model (AR) with a lag of 6 on the differenced dataset.

We can demonstrate this model below.

First, the data is transformed by differencing, with each observation transformed as:

1 |
value(t) = obs(t) - obs(t - 1) |

Next, the AR(6) model is trained on 66% of the historical data. The regression coefficients learned by the model are extracted and used to make predictions in a rolling manner across the test dataset.

As each time step in the test dataset is executed, the prediction is made using the coefficients and stored. The actual observation for the time step is then made available and stored to be used as a lag variable for future predictions.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 |
from pandas import read_csv from matplotlib import pyplot from statsmodels.tsa.ar_model import AR from sklearn.metrics import mean_squared_error import numpy # create a difference transform of the dataset def difference(dataset): diff = list() for i in range(1, len(dataset)): value = dataset[i] - dataset[i - 1] diff.append(value) return numpy.array(diff) # Make a prediction give regression coefficients and lag obs def predict(coef, history): yhat = coef[0] for i in range(1, len(coef)): yhat += coef[i] * history[-i] return yhat series = read_csv('daily-total-female-births.csv', header=0, index_col=0) # split dataset X = difference(series.values) size = int(len(X) * 0.66) train, test = X[0:size], X[size:] # train autoregression model = AR(train) model_fit = model.fit(maxlag=6, disp=False) window = model_fit.k_ar coef = model_fit.params # walk forward over time steps in test history = [train[i] for i in range(len(train))] predictions = list() for t in range(len(test)): yhat = predict(coef, history) obs = test[t] predictions.append(yhat) history.append(obs) error = mean_squared_error(test, predictions) print('Test MSE: %.3f' % error) # plot pyplot.plot(test) pyplot.plot(predictions, color='red') pyplot.show() |

Running the example first prints the Mean Squared Error (MSE) of the predictions, which is 52 (or about 7 births on average, if we take the square root to return the error score to the original units).

This is how well we expect the model to perform on average when making forecasts on new data.

1 |
Test MSE: 52.696 |

Finally, a graph is created showing the actual observations in the test dataset (blue) compared to the predictions (red).

This may not be the very best possible model we could develop on this problem, but it is reasonable and skillful.

## 2. Finalize and Save Time Series Forecast Model

Once the model is selected, we must finalize it.

This means save the salient information learned by the model so that we do not have to re-create it every time a prediction is needed.

This involves first training the model on all available data and then saving the model to file.

The *statsmodels* implementations of time series models do provide built-in capability to save and load models by calling *save()* and *load()* on the fit ARResults object.

For example, the code below will train an AR(6) model on the entire Female Births dataset and save it using the built-in *save()* function, which will essentially pickle the *ARResults* object.

The differenced training data must also be saved, both the for the lag variables needed to make a prediction, and for knowledge of the number of observations seen, required by the *predict()* function of the *ARResults* object.

Finally, we need to be able to transform the differenced dataset back into the original form. To do this, we must keep track of the last actual observation. This is so that the predicted differenced value can be added to it.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 |
# fit an AR model and save the whole model to file from pandas import read_csv from statsmodels.tsa.ar_model import AR from statsmodels.tsa.ar_model import ARResults import numpy # create a difference transform of the dataset def difference(dataset): diff = list() for i in range(1, len(dataset)): value = dataset[i] - dataset[i - 1] diff.append(value) return numpy.array(diff) # load dataset series = read_csv('daily-total-female-births.csv', header=0, index_col=0) X = difference(series.values) # fit model model = AR(X) model_fit = model.fit(maxlag=6, disp=False) # save model to file model_fit.save('ar_model.pkl') # save the differenced dataset numpy.save('ar_data.npy', X) # save the last ob numpy.save('ar_obs.npy', [series.values[-1]]) |

This code will create a file *ar_model.pkl* that you can load later and use to make predictions.

The entire training dataset is saved as *ar_data.npy* and the last observation is saved in the file *ar_obs.npy* as an array with one item.

The NumPy save() function is used to save the differenced training data and the observation. The load() function can then be used to load these arrays later.

The snippet below will load the model, differenced data, and last observation.

1 2 3 4 5 6 7 8 |
# load the AR model from file from statsmodels.tsa.ar_model import ARResults import numpy loaded = ARResults.load('ar_model.pkl') print(loaded.params) data = numpy.load('ar_data.npy') last_ob = numpy.load('ar_obs.npy') print(last_ob) |

Running the example prints the coefficients and the last observation.

1 2 3 |
[ 0.12129822 -0.75275857 -0.612367 -0.51097172 -0.4176669 -0.32116469 -0.23412997] [50] |

I think this is good for most cases, but is also pretty heavy. You are subject to changes to the statsmodels API.

My preference is to work with the coefficients of the model directly, as in the case above, evaluating the model using a rolling forecast.

In this case, you could simply store the model coefficients and later load them and make predictions.

The example below saves just the coefficients from the model, as well as the minimum differenced lag values required to make the next prediction and the last observation needed to transform the next prediction made.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 |
# fit an AR model and manually save coefficients to file from pandas import read_csv from statsmodels.tsa.ar_model import AR import numpy # create a difference transform of the dataset def difference(dataset): diff = list() for i in range(1, len(dataset)): value = dataset[i] - dataset[i - 1] diff.append(value) return numpy.array(diff) # load dataset series = read_csv('daily-total-female-births.csv', header=0) X = difference(series.values) # fit model window_size = 6 model = AR(X) model_fit = model.fit(maxlag=window_size, disp=False) # save coefficients coef = model_fit.params numpy.save('man_model.npy', coef) # save lag lag = X[-window_size:] numpy.save('man_data.npy', lag) # save the last ob numpy.save('man_obs.npy', [series.values[-1]]) |

The coefficients are saved in the local file *man_model.npy*, the lag history is saved in the file *man_data.npy*, and the last observation is saved in the file *man_obs.npy*.

These values can then be loaded again as follows:

1 2 3 4 5 6 7 8 |
# load the manually saved model from file import numpy coef = numpy.load('man_model.npy') print(coef) lag = numpy.load('man_data.npy') print(lag) last_ob = numpy.load('man_obs.npy') print(last_ob) |

Running this example prints the loaded data for review. We can see the coefficients and last observation match the output from the previous example.

1 2 3 4 |
[ 0.12129822 -0.75275857 -0.612367 -0.51097172 -0.4176669 -0.32116469 -0.23412997] [-10 3 15 -4 7 -5] [50] |

Now that we know how to save a finalized model, we can use it to make forecasts.

## 3. Make a Time Series Forecast

Making a forecast involves loading the saved model and estimating the observation at the next time step.

If the ARResults object was serialized, we can use the *predict()* function to predict the next time period.

The example below shows how the next time period can be predicted.

The model, training data, and last observation are loaded from file.

The period is specified to the *predict()* function as the next time index after the end of the training data set. This index may be stored directly in a file instead of storing the entire training data, which may be an efficiency.

The prediction is made, which is in the context of the differenced dataset. To turn the prediction back into the original units, it must be added to the last known observation.

1 2 3 4 5 6 7 8 9 10 11 12 |
# load AR model from file and make a one-step prediction from statsmodels.tsa.ar_model import ARResults import numpy # load model model = ARResults.load('ar_model.pkl') data = numpy.load('ar_data.npy') last_ob = numpy.load('ar_obs.npy') # make prediction predictions = model.predict(start=len(data), end=len(data)) # transform prediction yhat = predictions[0] + last_ob[0] print('Prediction: %f' % yhat) |

Running the example prints the prediction.

1 |
Prediction: 46.755211 |

We can also use a similar trick to load the raw coefficients and make a manual prediction.

The complete example is listed below.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 |
# load a coefficients and from file and make a manual prediction import numpy def predict(coef, history): yhat = coef[0] for i in range(1, len(coef)): yhat += coef[i] * history[-i] return yhat # load model coef = numpy.load('man_model.npy') lag = numpy.load('man_data.npy') last_ob = numpy.load('man_obs.npy') # make prediction prediction = predict(coef, lag) # transform prediction yhat = prediction + last_ob[0] print('Prediction: %f' % yhat) |

Running the example, we achieve the same prediction as we would expect, given the underlying model and method for making the prediction are the same.

1 |
Prediction: 46.755211 |

## 4. Update Forecast Model

Our work is not done.

Once the next real observation is made available, we must update the data associated with the model.

Specifically, we must update:

- The differenced training dataset used as inputs to make the subsequent prediction.
- The last observation, providing a context for the predicted differenced value.

Let’s assume the next actual observation in the series was 48.

The new observation must first be differenced with the last observation. It can then be stored in the list of differenced observations. Finally, the value can be stored as the last observation.

In the case of the stored AR model, we can update the *ar_data.npy* and *ar_obs.npy* files. The complete example is listed below:

1 2 3 4 5 6 7 8 9 10 11 12 13 14 |
# update the data for the AR model with a new obs import numpy # get real observation observation = 48 # load the saved data data = numpy.load('ar_data.npy') last_ob = numpy.load('ar_obs.npy') # update and save differenced observation diffed = observation - last_ob[0] data = numpy.append(data, [diffed], axis=0) numpy.save('ar_data.npy', data) # update and save real observation last_ob[0] = observation numpy.save('ar_obs.npy', last_ob) |

We can make the same changes for the data files for the manual case. Specifically, we can update the *man_data.npy* and *man_obs.npy* respectively.

The complete example is listed below.

1 2 3 4 5 6 7 8 9 10 11 12 13 |
# update the data for the manual model with a new obs import numpy # get real observation observation = 48 # update and save differenced observation lag = numpy.load('man_data.npy') last_ob = numpy.load('man_obs.npy') diffed = observation - last_ob[0] lag = numpy.append(lag[1:], [diffed], axis=0) numpy.save('man_data.npy', lag) # update and save real observation last_ob[0] = observation numpy.save('man_obs.npy', last_ob) |

We have focused on one-step forecasts.

These methods would work just as easily for multi-step forecasts, by using the model repetitively and using forecasts of previous time steps as input lag values to predict observations for subsequent time steps.

## Consider Storing All Observations

Generally, it is a good idea to keep track of all the observations.

This will allow you to:

- Provide a context for further time series analysis to understand new changes in the data.
- Train a new model in the future on the most recent data.
- Back-test new and different models to see if performance can be improved.

For small applications, perhaps you could store the raw observations in a file alongside your model.

It may also be desirable to store the model coefficients and required lag data and last observation in plain text for easy review.

For larger applications, perhaps a database system could be used to store the observations.

## Summary

In this tutorial, you discovered how to finalize a time series model and use it to make predictions with Python.

Specifically, you learned:

- How to save a time series forecast model to file.
- How to load a saved time series forecast from file and make a prediction.
- How to update a time series forecast model with new observations.

Do you have any questions?

Ask your questions in the comments below and I will do my best to answer.

Great Work sir.

is it ok to apply time series analysis to predict patients status in ICU using ICU Database such as MIMIC II database? if it is then would be kind enough to guide me on this.

Thanking you

It is not my area of expertise. It might be a fun learning exercise.

This process will guide you through your predictive modeling problem:

http://machinelearningmastery.com/start-here/#process

Thanks for the tutorial.

I noticed that when I ran the following code:

series = pd.read_csv(‘daily-total-female-births.csv’, header=0)

print(series.head())

series.plot()

plt.show()

I am not getting the same graph. The line graph appears more flat and the months are not showing. Was there additional codes not shown in the tutorial? Also, I tried using pyplot.show() so I update it to plt.show() and the code ran fine.

Thanks for your help.

Hi Hoang,

There’s no additional code.

Confirm you have the latest version of Pandas and matplotlib installed. Also confirm that your data file contains date-times and observations with no additional data or footer information.

Thanks Jason, for your quick response.

I am using Pandas v0.19.2 and matplotlib v2.0.0.

The additional data was deleted in the data file (thanks to your suggestion) and now the graph looks very similar, however the dates are missing. I’m assuming it’s because the indices are displayed on the x-axis. It was corrected once I set ‘Date’ as the index.

series=series.set_index(‘Date’)

The graph doesn’t look exactly like your graph but it’s close enough.

Glad to hear you made some progress, thanks for providing the details. It may help others.

Hi Jason, Thank you so much for your nice examples.

I’ve been wasting my time playing with game results datasets and these first step into prediction was really handy. Usually, in these games there is no tendency at all, all balls have the same fair chance of getting out. How would you approach such predictions, for example for Euromillions?

Where can I find some more advanced time-series predictions for multivariate cases? I would like to predict the outcome of 3 variables that linked to each other. Similar code to this example would be perfect…

Thank you

Do you mean lotto? I believe that is random and not predictable.

Once you frame your time series as a supervised learning problem, you can use a suite of machine learning methods.

For predicting 3 variables, I would recommend a neural network model.

Hi Jason,

Great content as always.

I had a question.

1) Why did you not use ARIMA model here and why AR? Is there a difference?

There is a difference, AR does not include the differencing and moving average. It’s a simpler method.

It is just an example for making predictions, you could swap in ARIMA if you like.

Hi Jason,

Really good stuff!

# make prediction

predictions = model.predict(start=len(data), end=len(data))

here the minimum value for start should be window (6 in this case) and less than the end time.

Thanks Martha.

“These methods would work just as easily for multi-step forecasts, by using the model repetitively and using forecasts of previous time steps as input lag values to predict observations for subsequent time steps.”

Hello, Jason! I’m a little new to this forecasting modeling approach, and was able to successfully able to run your tutorial and examples no problem. However, as I’m still poking around to see how this works, is there a good example you could point to that explains your above quote? I get that the predict function looks at coef[0] to ultimately perform its calculation, and our forecasting approach provides the next time step’s prediction, but how do I tell the function “look at the NEXT time step after that one…” etc? Is it coef[1]…coef[n]?

Apologies if these are really dumb questions, I find this fascinating at how accurate it can be on even real world data, and want to do statistical analyses to find out when the “long term forecast” falls off a cliff when it comes to accuracy, I just haven’t been able to figure out how to ask for time steps farther away than 1.

It is a good question.

You can use the model recursively but taking the forecast and using it as an observation in order to make the next forecast.

This post might have more details on the topic:

https://machinelearningmastery.com/multi-step-time-series-forecasting/

Jason – awesome example.

Would you use the same method if you say, wanted to predict occupancy on a given floor, of a given building – provided you have time-series data of all of the login’s (and logoffs) of all occupants that come onto the floor – dating back say 1 or 2 years?

Also, would you use the same method if you wanted to “predict” this occupancy from NOW to say, 2 – 3 days in advance – for a specific day? how would the time-step look like for the next day or in 2 days?

Perhaps. Try it and see how it performs on your data.

This is a great tutorial. I have a question why did you transform the data?

And why using difference ?

Is it necessary to transform the data ?

The data had a trend, I removed it via differencing.

Learn more about differencing here:

https://machinelearningmastery.com/remove-trends-seasonality-difference-transform-python/

How can I check the prediction performance

You can calculate the error between predictions and actual values.

Hello,

I will need your advice,

In fact, I have a large mass of data (almost 200 000), and I want to predict about 1 000 points.

Which algorithm would be most appropriate?

My database is univariate (just a data vector)

If you can direct me to a demo of “online prediction”, with real-time adaptation.

I recommend that you follow this process:

https://machinelearningmastery.com/how-to-develop-a-skilful-time-series-forecasting-model/

Hey,

I tried your code with your dataset, but i do not seem to be getting the same RMSE as your code. Is there any additional code you have missed out in your tutorial?

Thanks

~M

Perhaps some of these suggestions will help:

https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me

Dear Sir,

How can we plot graph with time as X-axis not the integers from csv file. If I directly read_csv file and plot then it works but after any type prediction model it always plot with integers values but I want Time as X-axis.

Thanks Sir in Advance

This post will help you get started with plotting:

https://machinelearningmastery.com/data-visualization-methods-in-python/

HI Jason ,

Not sure how do i fix the following error

maxlag should be 27 model_fit = model.fit(maxlag=6,disp=False)

maxlag should be < nobs

Yes.

Perhaps it is related to the amount of data used to fit the model?

Hi Jason,

Thanks for responding. I am struggling here.

I am using Weekly data for last 3 years (156 points) to check the forecast against the current years actuals which is 22 weeks. 178 points in total. If the model fits, i intend to use it for the following year. Data is seasonal as well. I guess i can run this code on Power BI to get the forecast for multiple sites(669 in total)?

Good question about multiple sites, you have many options.

This might give you some ideas:

https://machinelearningmastery.com/faq/single-faq/how-to-develop-forecast-models-for-multiple-sites

Hi Sir ,

Just a basic question here why are substracting the value from previous ones ?

Doesn’t the algorithm find the patter and help us in predictions?

Also sir my r2 value is .35 will that ok or should i work further ?

Really sorry if that was a very bad question but I’m new here so asked

Thanks

It is called differencing and it removes the trend, more here:

https://machinelearningmastery.com/difference-time-series-dataset-python/

If there is a dependent variable that has to be predicted and a datetime variable with no time information along with few categorical and numerical variables, can we split the date variable to day week, year month and drop the date column and predict the output , if so what type of variables will be the converted date time variable

Typically we drop the date/time variable once the data has a consistent interval/spacing.

Hi Jason,

Thanks for the good information on ARIMA.could you please let me know how to create a dataframe with actual and forecast and 95% confidence interval .waiting for your reply

Regards

Prakash Hullathi

Here is an example of calculating a predicting interval with ARIMA:

https://machinelearningmastery.com/time-series-forecast-uncertainty-using-confidence-intervals-python/

Hai Jason,

Thanks for your good tutorial about time series forecasting.

Could you please tell me can time series data forecast in Support Vector Regression (SVR)?

Yes, I don’t see why not.

do you have tutorial about it?

Yes, this might help:

https://machinelearningmastery.com/multi-step-time-series-forecasting-with-machine-learning-models-for-household-electricity-consumption/

thank you

I have followed your tutorial – once again. This time taking great care to copy the information.

My OS is Ubuntu 18.04

Vanilla install and everything updated.

I have created a Python3 environment and I have installed all the required module as required by your tutorial.

I have downloaded the dataset

This is the version of the modules I have installed.

scipy: 1.3.1

numpy: 1.17.0

matplotlib: 3.1.1

pandas: 0.25.1

sklearn: 0.21.3

statsmodels: 0.10.1

Dataset 93, Validation 12

When I execute your code I am confronted with an error.

File “complete.py”, line 47, in

X = X.astype(‘float32’)

ValueError: could not convert string to float: ‘1964-01’

Is it that the modules for this tutorial are older than the ones that I have installed or vise versa.

I could do with a bit of help. This is the only thing that is holding me back for purchasing your book.

Yes the API has changed.

I have updated all examples in the tutorial.

Thanks.

ValueError: end is before start

Im getting this error while trying to find the prediction for observation =48.

Do u have any solution?

I have some suggestions here that may help:

Consider aggressively cutting the code back to the minimum required. This will help you isolate the problem and focus on it.

Consider cutting the problem back to just one or a few simple examples.

Consider finding other similar code examples that do work and slowly modify them to meet your needs. This might expose your misstep.

Consider posting your question and code to StackOverflow.

Hello,

let talk about the problem of prediction of user-preferred channel each hour(time series prediction problem). Suppose that I have many users, each user has his own list of channels. So each hour the user will watch a specific channel (each user has his own behavior). in my DB I have +40K users and almost 50 channels. Those 40K users are in fact the clusters for more than 60 Million users. So we can’t cluster anymore. I couldn’t find a way to build one model that predicts for each user, at a specific time, which channel that the user can watch. Per in mind that I solved the problem for one user (I mean based on the one user data, I build an LSTM model that predict the user preferred channel at a specific time). But the problem is generalized to 40K users. Should I build 40K models which mean model for each user !!, or there is alternative solution ?!

Great question.

Firstly, it sounds like a classification problem, e.g. 50 channels are 50 classes. If you are modeling over time, it sounds like a time series classification problem.

You have lots of data, so I’d recommend testing a range of framing of the problem to see what works. E.g. model per customer, per customer group, across all customers, per day, per week, across all time, etc.

Then, we don’t know how hard the problem is so start with a naive model, compare to a linear model, some ml models, and even some deep learning models to see how learnable each framing happens to be.

Does that help?

Hi Jason,

thanks for you post.

I want to ask if there is example about long-term time series prediction based on LSTM model?

by direct prediction method or recurrent method?

Looking forward to your reply!

Best regards

Hao

Yes, this will help:

https://machinelearningmastery.com/faq/single-faq/how-do-you-use-lstms-for-multi-step-time-series-forecasting

Hello, I have managed to replicate your plots so far. However rather than plotting the actual and predicted birthrate difference (end of Section 1), I would like to plot the birthrate itself. How could I do that?

You can call pyplot.plot() and pass in any data you want.