Weather-Informed Machine Learning for Predictive Modeling in Residential Air-Conditioning Supply Chain Management

Abstract

Accurately forecasting residential air-conditioner demand is challenging because heat, humidity, and promotion timing generate sharp, region-specific sales swings; poor forecasts translate into over-/under-stocking and higher working-capital costs. This study’s objective is to build and compare weather-informed, region-aware machine-learning models that improve short-horizon planning for retailers. Using five years (2018–2022) of weekly data from five regions, we engineer climatic, promotional, socioeconomic, calendar, and lag features and evaluate linear regression, random forest, and XGBoost under a time-based split (train: 2018–2021; test: 2022) with MAE, RMSE, and R², complemented by permutation importance and SHAP for interpretation. XGBoost performs best (MAE 13.57; RMSE 16.73; R² = 0.913), and importance analyses consistently rank average temperature, promotion status, and a composite socioeconomic index as the dominant drivers; a Dhaka case plot shows close tracking of seasonal peaks and troughs. Coupling accurate heat forecasts with planned promotions can meaningfully reduce error and inventory risk; the socioeconomic index and regional dummies guide geographic allocation, while simple lags offer incremental week-ahead adjustments. The approach is directly actionable for weather-sensitive consumer durables and transferable to adjacent markets.