The problem we are tackling here is “Design Compensation Incentives Based on performance of the Store Managers”. We can take the store sales as a proxy to indicate the performance of the manager. The model can be used to set sales target to their respective store managers. The variable payout is dependent on the manager’s ability to exceed the target.
The cost of errors in this context is very high. Under-prediction will result in high unachievable targets resulting in demotivation or attrition. Over-prediction will set easy and low targets leaving a lot of money on the table. Hence, we need to create a forecasting model which can accurately predict sales of the next 2 weeks.
The data in hand is the weekly sales data from 45 stores, across 99 departments along with a few additional external information such as unemployment rate, CPI, etc. From the time series of the sales of a certain store, we can see that the data is seasonal i.e. there is a similar pattern followed every year. There is no increase in the y-o-y level.
Final Forecasting Method
After evaluating multiple forecasting methods, a regression model taking the week index (week #) and accounting for seasonality (repeated patterns across the year) and holidays gives the best results. It’s performance as compared to the benchmark model can be observed below.
Though the regression model gives a better forecast, the costs incurred to run the same is huge. Hence, Walmart should use the benchmark model to design the store managers’ incentives. However, it should be noted that this cannot be used in presence of external shocks such as recession and heavy competition.