Multivariate Time Series Forecasting


This post was originally published by Mohammad Masum at Towards Data Science

Applying Vector Autoregressive (VAR) model to a real-world multivariate dataset

A univariate time series data contains only one single time-dependent variable while a multivariate time series data consists of multiple time-dependent variables. We generally use multivariate time series analysis to model and explain the interesting interdependencies and co-movements among the variables. In the multivariate analysis — the assumption is that the time-dependent variables not only depend on their past values but also show dependency between them. Multivariate time series models leverage the dependencies to provide more reliable and accurate forecasts for a specific given data, though the univariate analysis outperforms multivariate in general[1]. In this article, we apply a multivariate time series method, called Vector Auto Regression (VAR) on a real-world dataset.

VAR model is a stochastic process that represents a group of time-dependent variables as a linear function of their own past values and the past values of all the other variables in the group.

For instance, we can consider a bivariate time series analysis that describes a relationship between hourly temperature and wind speed as a function of past values [2]:

temp(t) = a1 + w11* temp(t-1) + w12* wind(t-1) + e1(t-1)

wind(t) = a2 + w21* temp(t-1) + w22*wind(t-1) +e2(t-1)

where a1 and a2 are constants; w11, w12, w21, and w22 are the coefficients; e1 and e2 are the error terms.

Statmodels is a python API that allows users to explore data, estimate statistical models, and perform statistical tests [3]. It contains time series data as well. We download a dataset from the API.

To download the data, we have to install some libraries and then load the data:

import pandas as pd
import statsmodels.api as sm
from statsmodels.tsa.api import VAR
data = sm.datasets.macrodata.load_pandas().data

The output shows the first two observations of the total dataset:

A snippet of the dataset

The data contains a number of time-series data, we take only two time-dependent variables “realgdp” and “realdpi” for experiment purposes and use “year” columns as the index of the data.

data1 = data[["realgdp", 'realdpi']]
data1.index = data["year"]


A snippet of the data

Let’s visualize the data:

data1.plot(figsize = (8,5))

Both of the series show an increasing trend over time with slight ups and downs.

Before applying VAR, both the time series variable should be stationary. Both the series are not stationary since both the series do not show constant mean and variance over time. We can also perform a statistical test like the Augmented Dickey-Fuller test (ADF) to find stationarity of the series using the AIC criteria.

adfuller_test = adfuller(data1['realgdp'], autolag= "AIC")
print("ADF test statistic: {}".format(adfuller_test[0]))
print("p-value: {}".format(adfuller_test[1]))


Result of the statistical test for stationarity

In both cases, the p-value is not significant enough, meaning that we can not reject the null hypothesis and conclude that the series are non-stationary.

As both the series are not stationary, we perform differencing and later check the stationarity.

data_d = data1.diff().dropna()

The series becomes stationary after first differencing of the original series as the p-value of the test is statistically significant.

ADF test for one differenced realgdp data

The series becomes stationary after first differencing of the original series as the p-value of the test is statistically significant.

ADF test for one differenced realdpi data

In this section, we apply the VAR model on the one differenced series. We carry-out the train-test split of the data and keep the last 10-days as test data.

train = data_d.iloc[:-10,:]
test = data_d.iloc[-10:,:]

In the process of VAR modeling, we opt to employ Information Criterion Akaike (AIC) as a model selection criterion to conduct optimal model identification. In simple terms, we select the order (p) of VAR based on the best AIC score. The AIC, in general, penalizes models for being too complex, though the complex models may perform slightly better on some other model selection criterion. Hence, we expect an inflection point in searching the order (p), meaning that, the AIC score should decrease with order (p) gets larger until a certain order and then the score starts increasing. For this, we perform grid-search to investigate the optimal order (p).

forecasting_model = VAR(train)results_aic = []
for p in range(1,10):
results =

In the first line of the code: we train VAR model with the training data. Rest of code: perform a for loop to find the AIC scores for fitting order ranging from 1 to 10. We can visualize the results (AIC scores against orders) to better understand the inflection point:

import seaborn as sns
plt.plot(list(np.arange(1,10,1)), results_aic)

Investigating optimal order of VAR models

From the plot, the lowest AIC score is achieved at the order of 2 and then the AIC scores show an increasing trend with the order p gets larger. Hence, we select the 2 as the optimal order of the VAR model. Consequently, we fit order 2 to the forecasting model.

let’s check the summary of the model:

results =

The summary output contains much information:

We use 2 as the optimal order in fitting the VAR model. Thus, we take the final 2 steps in the training data for forecasting the immediate next step (i.e., the first day of the test data).

Forecasting test data

Now, after fitting the model, we forecast for the test data where the last 2 days of training data set as lagged values and steps set as 10 days as we want to forecast for the next 10 days.

laaged_values = train.values[-2:]forecast = pd.DataFrame(results.forecast(y= laaged_values, steps=10), index = test.index, columns= ['realgdp_1d', 'realdpi_1d'])forecast

The output:

First differenced forecasts

We have to note that the aforementioned forecasts are for the one differenced model. Hence, we must reverse the first differenced forecasts into the original forecast values.

forecast["realgdp_forecasted"] = data1["realgdp"].iloc[-10-1] +   forecast_1D['realgdp_1d'].cumsum()forecast["realdpi_forecasted"] = data1["realdpi"].iloc[-10-1] +      forecast_1D['realdpi_1d'].cumsum() 


Forecasted values for 1 differenced series and for the original series

The first two columns are the forecasted values for 1 differenced series and the last two columns show the forecasted values for the original series.

Now, we visualize the original test values and the forecasted values by VAR.

Original and Forecasted values for realgdp and realdpi

The original realdpi and the forecasted realdpi show a similar pattern throwout the forecasted days. For realgdp: the first half of the forecasted values show a similar pattern as the original values, on the other hand, the last half of the forecasted values do not follow similar pattern.

To sum up, in this article, we discuss multivariate time series analysis and applied the VAR model on a real-world multivariate time series dataset.

You can also read the article — A real-world time series data analysis and forecasting, where I applied ARIMA (univariate time series analysis model) to forecast univariate time series data.




Spread the word

This post was originally published by Mohammad Masum at Towards Data Science

Related posts