# Forecasting on multiple timeseries data with limited data points

I'm predicting operational expense of a stores of a company. I have only six years of data per store at a daily granularity. I want to train a model to predict the next years operational expense. In addition to the operational expense historical data, I have different attributes of the store like location, area of store, no of floors, lighting type etc.

I'm not sure what model to choose for this problem. Should I formulate this as a timeseries problem or a regression problem? How do I decide on how to approach this problem? What are the different considerations while choosing the algorithm

• Could you be more formal on your issue? For example, do you have a single store or stores? Please observe that you have used stores in the first sentence; I am assuming the former. Moreover, I also assume that you would like to use the additional information for prediction, right? Feb 11, 2023 at 8:33
• Yes, I have around 200 stores data for the last 6 years. I want to use additional information for prediction. Feb 11, 2023 at 10:50

The first thing that should be done in this situation is to decompose the time series and analyse the trend and seasonality in isolation.

After this, one could consider the following:

1. Time-series approach

If one is looking to use a time series model such as ARIMA to forecast operational expense without regard to any of the independent variables - then this would be suited to such a model if the data has predictive seasonality patterns, e.g. operational expenses might be higher every Easter and Christmas, for instance. If there is a clear longer-term trend in the data, then this will also be easier for ARIMA to interpret.

2. Regression-based approach

However, if one finds that the data is volatile and the seasonality and trend are not necessarily predictable - then a linear regression could be used to investigate how the independent variables (e.g. area of store, no of floors, etc) influence the dependent variable. However, one would need to test for autocorrelation in this instance and remedy it if necessary, in order to ensure that the dependent variable is not being influenced by lagged values of itself.

One would also need to screen the regression for multicollinearity to ensure that there does not exist multiple variables in the regression that have similar theoretical meaning and are both influencing the dependent variable. As one example, both the area of store and number of floors variables are in essence measuring the size of the store - so there might be multicollinearity present.

Heteroscedasticity should also be screened for - which is an issue if the variance of the error term is not constant. This might be the case if certain stores are larger than others - thus the results will be skewed towards the operational expenses of the larger stores.

Ultimately, my first step would be to decompose the time series data for analysis. Then, proceed with the most suitable approach.