Assessing the Performance of Hierarchical Forecasting Methods on the Retail Sector

Area of Science:

• Operations Research
• Business Analytics
• Time Series Forecasting

Background:

• Retailers require demand forecasts across various aggregation levels for supply chain decisions.
• Ensuring forecast consistency across hierarchical levels is crucial for aligned decision-making.
• The choice between independent and hierarchical forecasting methods impacts accuracy and coherence.

Purpose of the Study:

• To empirically compare the performance of independent versus reconciled forecasting approaches.
• To provide guidelines for retailers on selecting appropriate forecasting methods.
• To investigate forecast accuracy improvements using real retail data.

Main Methods:

• Utilized state space models and ARIMA for base forecast generation.
• Selected optimal models using the bias-corrected Akaike information criterion.
• Applied the Minimum Trace Shrinkage (MinT-Shrink) estimator for forecast reconciliation.

Main Results:

• Reconciled forecasts using MinT-Shrink consistently outperformed ARIMA base forecasts in accuracy.
• Accuracy gains were observed across all hierarchy levels and forecast horizons.
• Forecast accuracy improvements were more pronounced at higher aggregation levels.

Conclusions:

• Hierarchical forecasting with reconciliation, specifically MinT-Shrink, enhances forecast accuracy for retailers.
• Reconciliation effectively reintegrates lost information from lower aggregation levels into higher ones.
• These findings support improved management decisions through more accurate, coherent demand forecasts.