Demand Forecasting based on Pairwise Item Associations

Forecasting is an essential task conducted regularly by competitive retailers around the world. Most retail decisions are made based on the demand forecasts which may or may not be accurate in the first place. In this study, a framework for forecasting weekly demands of retail items is proposed via linear regression models within item groups that incorporate both positive and negative item associations. In addition to pairwise item associations found by utilizing transactional data, our framework incorporates item similarities based on weekly sales figures to group the similar items. Grouping items can be regarded as a form of variable selection to prevent the overfitting in the prediction models. The regression results of the framework and benchmark linear regression models are reported for a real world dataset provided by an apparel retailer. The results show that the regression models provide better estimates within multi-item groups compared to the single item models.