Instamart’s Generative AI Forecasting: Scaling from Flat Forecasts to Deep Hierarchies

Swiggy's Instamart team shares how they improved demand forecasting for quick commerce using Google's TimesFM foundation model combined with hierarchical forecasting. The system addresses two key challenges: warehouse SKU demand forecasting and Orders per Day (OPD) forecasting for rider fleet planning. Key techniques include availability de-biasing to correct censored sales data, Group SKU (GSKU) aggregation to reduce variance from product substitution, and a dual-hierarchy framework that separates the business objective level from the optimal modeling level. TimesFM is used in zero-shot mode with covariate support via ridge regression. Results include 21–38% relative accuracy improvements across product categories for warehouse forecasting and a 7.5% WMAPE improvement with 50% runtime reduction for OPD forecasting compared to the previous TFT-based system.