Last Updated on December 10, 2020

The autoregression integrated moving average model or ARIMA model can seem intimidating to beginners.

A good way to pull back the curtain in the method is to to use a trained model to make predictions manually. This demonstrates that ARIMA is a linear regression model at its core.

Making manual predictions with a fit ARIMA models may also be a requirement in your project, meaning that you can save the coefficients from the fit model and use them as configuration in your own code to make predictions without the need for heavy Python libraries in a production environment.

In this tutorial, you will discover how to make manual predictions with a trained ARIMA model in Python.

Specifically, you will learn:

- How to make manual predictions with an autoregressive model.
- How to make manual predictions with a moving average model.
- How to make predictions with an autoregression integrated moving average model.

**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 dive in.

**Updated Apr/2019**: Updated the link to dataset.**Updated Aug/2019**: Updated data loading to use new API.**Updated Dec/2020**: Updated ARIMA API to the latest version of statsmodels.

## Minimum Daily Temperatures Dataset

This dataset describes the minimum daily temperatures over 10 years (1981-1990) in the city Melbourne, Australia.

The units are in degrees Celsius and there are 3,650 observations. The source of the data is credited as the Australian Bureau of Meteorology.

Download the dataset and place it into your current working directory with the filename “*daily-minimum-temperatures.csv*“.

The example below demonstrates how to load the dataset as a Pandas Series and graph the loaded dataset.

1 2 3 4 5 |
from pandas import read_csv from matplotlib import pyplot series = read_csv('daily-minimum-temperatures.csv', header=0, index_col=0) series.plot() pyplot.show() |

Running the example creates a line plot of the time series.

### 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.

## ARIMA Test Setup

We will use a consistent test harness to fit ARIMA models and evaluate their predictions.

First, the loaded dataset is split into a train and test dataset. The majority of the dataset is used to fit the model and the last 7 observations (one week) are held back as the test dataset to evaluate the fit model.

A walk-forward validation, or rolling forecast, method is used as follows:

- Each time step in the test dataset is iterated.
- Within each iteration, a new ARIMA model is trained on all available historical data.
- The model is used to make a prediction for the next day.
- The prediction is stored and the “real” observation is retrieved from the test set and added to the history for use in the next iteration.
- The performance of the model is summarized at the end by calculating the root mean squared error (RMSE) of all predictions made compared to expected values in the test dataset.

Simple AR, MA, ARMA and ARMA models are developed. They are unoptimized and are used for demonstration purposes. You will surely be able to achieve better performance with a little tuning.

The ARIMA implementation from the statsmodels Python library is used and AR and MA coefficients are extracted from the ARIMAResults object returned from fitting the model.

The ARIMA model supports forecasts via the predict() and the forecast() functions.

Nevertheless, we will make manual predictions in this tutorial using the learned coefficients.

This is useful as it demonstrates that all that is required from a trained ARIMA model is the coefficients.

The coefficients in the statsmodels implementation of the ARIMA model do not use intercept terms. This means we can calculate the output values by taking the dot product of the learned coefficients and lag values (in the case of an AR model) and lag residuals (in the case of an MA model). For example:

1 |
y = dot_product(ar_coefficients, lags) + dot_product(ma_coefficients, residuals) |

The coefficients of a learned ARIMA model can be accessed from aARIMAResults object as follows:

**AR Coefficients**: model_fit.arparams**MA Coefficients**: model_fit.maparams

We can use these retrieved coefficients to make predictions using the following manual *predict()* function.

1 2 3 4 5 |
def predict(coef, history): yhat = 0.0 for i in range(1, len(coef)+1): yhat += coef[i-1] * history[-i] return yhat |

For reference, you may find the following resources useful:

Hi Jason,

Thanks for this tutorial.

Now the error occure as follow, Please advise me about it.

I think also that the code do not have an error.

—> 23 model_fit = model.fit(trend=’nc’, disp=False)

ValueError: could not convert string to float: ?0.2

thanks

Open the downloaded data file and delete all instances of the “?” character.

It worked fine.

I did not think that it was caused by converting the file to CSV.

Thank you, Jason.

Glad to hear it!

Hi Jason

Really appreciate for this article, I’ve got one question: in this example, why do we use an iterative way to determine the ARIMA parameters, can we fix the model parameters before the loop in test dataset and then doing the validation process?

Thanks a lot for some more insights on it.

I’m not sure what you mean by iterative? Can you please elaborate?

Thanks for the feedback, Jason, what I meant before is:

…

for t in range(len(test)):

model = ARIMA(history, order=(1,1,1))

…

we see that in each loop, we train the model again and get a new set of parameters. Why not train the model just based on the training set and all parameters in the model is fixed, then loop for all test set and validate the error for each of them.

Thanks again for your reply.

You can, but if we have new data (e.g. it is the next month and a new observation is available) then we should use it.

That is what we are simulating here. It is called walk forward validation:

http://machinelearningmastery.com/backtest-machine-learning-models-time-series-forecasting/

Thanks a lot Jason.

Nice example from that link. I find that there’re plenty of quite useful and interesting information from your posts, I will go through others and post questions (if I have).

Again, thanks for the work, great job. ðŸ™‚

Thanks Luca!

Let’s say I have two data points in a week recorded over several years, for example data from Monday and Friday. Would this affect the model? Would this be relevant for the difference function?

It may, it depends on the problem.

I have several ARIMA tutorials read on this site. Some use the difference function some not- like the parameter tunings. When do I need the difference function.

When you have a trend, please read this:

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

Hi Jason

Since the dataset seems to have strong seasonality, in this case, do we need to first decompose the data by removing the seasonality factor, and then, applying models such as ARIMA?

Thanks

It would be a good idea to seasonally adjust the dataset first:

http://machinelearningmastery.com/time-series-seasonality-with-python/

Hi Jason,

Thanks for all the ARIMA tutorials! I’m preparing to analyze some gait data from a biomechanics lab I worked in last summer and they’re helping me to get to hang of the model.

Consider a neural net model?

Hi Jason,

Imagine a scenario where we are using the fitted ARIMA model i.e. the coefficients on a new dataset (out of sample). Calculating the AR part is easy. Use the previous data of length ‘p’ (history). Whereas for the MA part, one needs the residuals for past ‘q’ values. Without fitting for the new data seen, how are the the residuals calculated for the new data. Do we use the residuals found in the training data? This is useful for for ARIMA models with q>0 (i.e. having MA coefficients).

Good question. The ARIMA model will make these available in “model_fit.resid” I believe.

thanks for post, it is helpful.

But I have question, why the history data need to append test data for each loop:

model = ARIMA(history, order=(1,0,0))

obs = test[t]

history.append(obs)

If I do’t know the test data value, for example just predict 6 month or 1 years days(365 days range) value, how to use the model to predict. Just like machine learning, used the train set to train model, and predict new data.

Because we are evaluating the model using walk-forward validation:

http://machinelearningmastery.com/backtest-machine-learning-models-time-series-forecasting/

Hey Jason,

Thanks a lot for these ARIMA posts, I know close to nothing about time series and I’m learning this stuff to work on a project to find a cross-correlation matrix. I need to pre-whiten the data and I used your tutorial to fit a model to one of my series. Next I have to apply this model (filter) to another series and I was wondering if you could elaborate on how to do that using predict. I’ve been scouring the internet for some material, and I don’t know what exactly the contents and format of the ‘params’ field is supposed to be.

Thanks a lot!

This might be a good place to start with time series:

https://machinelearningmastery.com/start-here/#timeseries

I don’t have material on how to whiten a time series, sorry.

Bonjour, est-il possible dâ€™intÃ©grer d’autres variables dans le modÃ¨le ARIMA pour faire de la prÃ©diction ?

Je n etrouve pas de rÃ©ponses concernant cette question.

Yes, it is called exogenous variables.

Sorry, I don’t have an example.

Hi! How can I predct for example 20 reading forward?

mod =ARIMA(X, order=(self.p, self.d, self.q)

res = mod.fit()

res.forecast(20)

Perhaps this post will help:

https://machinelearningmastery.com/make-sample-forecasts-arima-python/

Hello Jason,

Here in example you said 3650 observations and have taken last 7 days for test set.. I am lil confused about my data set. I have weekly data of 7 weather parameters for 30 years. How can I select the training and test set?

Perhaps use a little trial and error and see how much history is required to fit a skilful model with your specific data.

This post might give you ideas:

https://machinelearningmastery.com/sensitivity-analysis-history-size-forecast-skill-arima-python/

I tried both your method to manually predict the series (which I think correct) and statsmodels’ .predict(). Do you have an explanation on why there are not exactly same? statsmodels is not that clear about how they perform predictions. Thanks!

The predict() and the forecast() functions should return the same results.

Learn how to use them here:

https://machinelearningmastery.com/make-sample-forecasts-arima-python/

I have a question, how do they acf and pacf were obtained here? I have a similar dataset with almost 1100 observations, and based on my acf and pacf, I have a significant lag from 1 to 27 for the acf. I want to know also if there is a limitation for the ARIMA model when we have those high values. Thanks

I used the statsmodels plot functions.

hi,I think there is a big error in the ARIMA model about the difference.

When we use AIRMA(p,d,q), if d not equal to 0, the model has done the difference step.

So the predict result data also has been diffenenced, so we need to restore the data.

But the restored result (no difference) is wrong. The predict error is very big.

In your many ARIMA examples, such as ARIMA(5,1,0), the d = 1, but I didn’t find the restored difference.

If the ARIMA(5,0,0), it doesn’t need to restore the difference.

The difference I mean the parameter d in the ARIMA(p,d,q).

I think maybe the python model ARIMA has little error.

The model will invert the difference (if needed) when a prediction is made.

Hi, Jason, I have a question about the train and validation.

In the four examples, you split the data into train and test part. But in your prediction, you also add the data in test to the history. Is it OK to do in this way? According to my understanding, it is not right to use the data in test set.

Another question is about the prediction. According to the document (https://www.statsmodels.org/dev/generated/statsmodels.tsa.arima_model.ARMAResults.forecast.html#statsmodels.tsa.arima_model.ARMAResults.forecast), we can predict the future in many steps (not only 1). Why you do not use the steps directly?

Yes, this is called walk forward validation, more here:

https://machinelearningmastery.com/backtest-machine-learning-models-time-series-forecasting/

I saw the link. It is detailed. I also checked other web pages. It seems that “walk forward validation ” is uncommon in machine learning. If we want to compare the result with other methods. What we should do?

If we use other complex methods and use “walk forward validation”, it will wast a lot of time to train the model.

Thanks.

It is common in time series forecasting problems.

Other methods should use the approach, and if they do not, such as using cross validation, their results are almost certainly invalid.

How can I calculate training error for time series?

Typically we use mean squared error (MSE) or root mean squared error (RMSE).

This is great, thank you! I’ve been looking for a way to manually calculate prediction from parameters because I need to save parameters for forecasting in another context. And I just couldn’t find any documentation that convinced me how to do it… so thanks for this!

I’m happy it helped!

Yes we calculate RMSE but how? What will be the actual and predicted values?

You can use sklearn to calculate RMSE.

Calculate MSE, then take the square root:

https://scikit-learn.org/stable/modules/generated/sklearn.metrics.mean_squared_error.html

Hi Jason,

Firstly, thanks for the tutorial. I have a couple of questions in my mind. Firstly, if I do multi-step ahead forecasting and defining AR and MA terms with some number means that the model looks back that amount out of back steps and do the forecast for the future? If I introduce the whole historical time series to the model, how the model uses that whole historical data to do multi-step ahead forecasting? how the coefficients in the AR and MA terms are updated? I would very thankful If I understand them.

Kind Regards,

Gunay

Perhaps this will help:

https://machinelearningmastery.com/make-sample-forecasts-arima-python/

Hi Jason,

Thank you for the tutorial! I have a question about settting yhat. What would yhat equal in the (0,0,0), (0, 1, 1), (1,1,0), and (0,1,0) cases and, for those cases, what parameters should we use for the predict function?

Thank you,

Emilio

I’m not sure I follow.

yhat is the prediction from the model.

The order (e.g. (0,1,0)) is the configuration of the model.

Does that help?

So say I have a 0,1,0 Arima model. I need to have a constant term to be able to run it. How should I incorporate the constant term into the predict function? For the 0,1,0 case and other cases as well?

Good question, perhaps check the source code for exactly how the statsmodels implementation incorporates the constant.

Or just use the statsmodels model directly:

https://machinelearningmastery.com/make-sample-forecasts-arima-python/

Thanks for the article. Can I ask then how to use this model to predict future days? This practice shows prediction for current data.

TQ

Yes, I show how here:

https://machinelearningmastery.com/make-sample-forecasts-arima-python/

Thanks for writing this. I always enjoy your articles.

How would this work for higher order models like ARIMA(4,1,1)?

Same general idea, although it might be easier to resort to the forecast() or predict() functions:

https://machinelearningmastery.com/make-sample-forecasts-arima-python/

This is my code:

from statsmodels.tsa.arima_model import ARIMA

from sklearn.metrics import mean_squared_error

train= training_set[‘Close’].values.reshape(-1, 1) #reshaping training values

test = test_set.values.reshape(-1, 1)

history = [x for x in train]

predictions = list()

for t in range(len(test)):

model = ARIMA(history, order=(1,2,1))

model_fit = model.fit(trend=’nc’, disp=False)

ar_coef, ma_coef = model_fit.arparams, model_fit.maparams

resid = model_fit.resid

diff = difference(history)

yhat = history[-1] + predict(ar_coef, diff) + predict(ma_coef, resid)

predictions.append(yhat)

obs = test[t]

history.append(obs)

print(‘>predicted=%.3f, expected=%.3f’ % (yhat, obs))

rmse = sqrt(mean_squared_error(test, predictions))

print(‘Test RMSE: %.3f’ % rmse)

Output:

>predicted=3366.015, expected=3370.000

>predicted=3371.046, expected=3390.000

>predicted=3369.670, expected=3370.000

>predicted=3392.017, expected=3377.800

>predicted=3373.568, expected=3460.000

>predicted=3369.171, expected=3460.000

>predicted=3452.457, expected=3400.000

>predicted=3469.630, expected=3458.100

>predicted=3402.823, expected=3490.000

>predicted=3449.645, expected=3490.000

>predicted=3488.853, expected=3442.200

>predicted=3498.426, expected=3468.300

>predicted=3447.526, expected=3489.000

>predicted=3465.161, expected=3510.000

>predicted=3486.675, expected=3485.000

>predicted=3513.211, expected=3498.900

>predicted=3489.170, expected=3500.000

>predicted=3499.777, expected=3580.000

>predicted=3493.101, expected=3495.200

>predicted=3583.419, expected=3545.500

>predicted=3503.000, expected=3720.000

—————————————————————————

ValueError Traceback (most recent call last)

in

8 for t in range(len(test)):

9 model = ARIMA(history, order=(1,2,1))

—> 10 model_fit = model.fit(trend=’nc’, disp=False)

11 ar_coef, ma_coef = model_fit.arparams, model_fit.maparams

12 resid = model_fit.resid

~/anaconda3/lib/python3.7/site-packages/statsmodels/tsa/arima_model.py in fit(self, start_params, trend, method, transparams, solver, maxiter, full_output, disp, callback, start_ar_lags, **kwargs)

1155 arima_fit.mle_retvals = mlefit.mle_retvals

1156 arima_fit.mle_settings = mlefit.mle_settings

-> 1157

1158 return ARIMAResultsWrapper(arima_fit)

1159

~/anaconda3/lib/python3.7/site-packages/statsmodels/tsa/arima_model.py in fit(self, start_params, trend, method, transparams, solver, maxiter, full_output, disp, callback, start_ar_lags, **kwargs)

944 kwargs.setdefault(‘pgtol’, 1e-8)

945 kwargs.setdefault(‘factr’, 1e2)

–> 946 kwargs.setdefault(‘m’, 12)

947 kwargs.setdefault(‘approx_grad’, True)

948 mlefit = super(ARMA, self).fit(start_params, method=solver,

~/anaconda3/lib/python3.7/site-packages/statsmodels/tsa/arima_model.py in _fit_start_params(self, order, method, start_ar_lags)

560 pgtol=1e-7, factr=1e3,

561 bounds=bounds, iprint=-1)

–> 562 start_params = mlefit[0]

563 if self.transparams:

564 start_params = self._transparams(start_params)

~/anaconda3/lib/python3.7/site-packages/statsmodels/tsa/arima_model.py in _fit_start_params_hr(self, order, start_ar_lags)

546 return start_params

547

–> 548 def _fit_start_params(self, order, method, start_ar_lags=None):

549 if method != ‘css-mle’: # use Hannan-Rissanen to get start params

550 start_params = self._fit_start_params_hr(order, start_ar_lags)

ValueError: The computed initial MA coefficients are not invertible

You should induce invertibility, choose a different model order, or you can

pass your own start_params.

Yes, some configurations will result in an unstable model.

Hi sir, i have problem with forecasting between using this manual predicting and using the library..

I compared the 2 result but it showed different result.

I use this model fitting :

model_fit = model.fit(trend=’nc’, disp=False). the library model i used show constant result . For example [10.234,10.235,10.235,….10.235,] . Can you help me please ðŸ™‚

You might need to dig into the statsmodels source code to see what additional steps are being used.

I wonder that why we don’t use constanta in this tutorial and why disp= False?. I tried to make the parameter same but the result show different to the forecast.

We set display to False to remove the verbose output.

Hi Sir, thank you for the impact.

My data has been preprocessed to a range of -1 and 1. I’m doing a walk-forward validation

for ARIMA and the model would predict extremely high values outside this range for some observations.

Is there any reason for that?

Thank you.

Perhaps the model needs to be further tuned for your dataset?

Perhaps you need to scale data to a different range?

Perhaps try an alternate model?