I have to do forecasting of sales that is how much sales of a product is going to happen in a particular store. I have time series data for last two years and doing forecasting for 2014. The variables are promotion flag ( Yes/ No ), promotion period, location in a store, price discount. These all are categorical variables.

For this I am using regression method where, dependent variable is sales, and independent variables are categorical variables mentioned above. This analysis is done in SPSS where I have used step-wise and backward regression.

Below is the link for the data:


I want to know, the regression model is under-forecasting? Is there a way to improve the forecast?

  • $\begingroup$ Make and post plots with residuals on the y axis and various different things on the x axis (e.g. 1-predicted values, 2-continuous predictors) $\endgroup$ Sep 16, 2014 at 15:11

2 Answers 2


You should investigate adding the impact of the day-of-the-week, weekly indicators,monthly indicators and the effect of holidays. Holiday/events routinely have lead and lag effects. Furthermore there may be level shifts in your data or local time trends . Additionally there may be an ARIMA component that needs to be included. Please post your data for one of the stores and I will be more specific.

You might want to look at

Wrong predictions for weekend, but good predictions for weekdays

as it discusses your problem/opportunity

  • $\begingroup$ I have included the day of weeks, still it is not giving better result. How we can include ARIMA components in the model? If I am not wrong, are you trying to say, use ARIMA model with all these regressor variables instead of regression model. $\endgroup$
    – Arushi
    Sep 18, 2014 at 9:06
  • $\begingroup$ As delivered your data is not in correct form. Please contact me offline and I will guide you in how to prepare it for analysis. $\endgroup$
    – IrishStat
    Sep 18, 2014 at 16:28

Its important to define the problem correctly and then you can model it. I am bit confused reading the problem.

If you are trying to predict value of Response variable SALES using Prediction variables like Price, Promo and / or Ad Placement then use Multiple Regression Model. Then you can try to fit it.

On other-hand if you are developing a Forecasting Model for SALES for future then use Exp (Double) Smoothing, Winter Method and Arima model etc.

If you are doing later and the results are not good then use esemble methods.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.