Data Aggregation and Demand Prediction

We study how retailers can use data aggregation and clustering to improve demand prediction. High accuracy in demand prediction allows retailers to effectively manage their inventory as well as mitigate stock-outs and excess supply. A typical retail setting involves predicting demand for hundreds of items simultaneously. Although some items have a large amount of historical data, others were recently introduced and, thus, transaction data can be scarce. A common approach is to cluster several items and estimate a joint model for each cluster. In this vein, one can estimate some model parameters by aggregating the data from several items and other parameters at the individual-item level. We propose a practical method referred to as data aggregation with clustering ($\mathsf{DAC}$), which balances the tradeoff between data aggregation and model flexibility. $\mathsf{DAC}$ allows us to predict demand while optimally identifying the features that should be estimated at the (i) item, (ii) cluster, and (iii) aggregate levels. We show that the $\mathsf{DAC}$ algorithm yields a consistent and normal estimate, along with improved prediction errors relative to the decentralized benchmark, which estimates a different model for each item. Using both simulated and real data, we illustrate $\mathsf{DAC}$’s improvement in prediction accuracy relative to a wide range of common benchmarks. Interestingly, the $\mathsf{DAC}$ algorithm has theoretical and practical advantages and helps retailers uncover meaningful managerial insights.