ARTIFICIAL INTELLIGENCE AND ECONOMETRIC VALIDATION: A THREE-REGIME APPROACH TO FORECASTING RETAIL CHAIN FINANCIAL PERFORMANCE

Abstract

Artificial intelligence and machine learning models are increasingly used to forecast the financial performance of retail chains, yet their genuine practical effectiveness depends heavily on the choice of econometric validation method. Using a panel dataset of 158,715 store-day observations from 140 outlets of a confectionery-café chain operating in Uzbekistan, this thesis comparatively evaluates XGBoost, CatBoost, and a regularized recurrent (GRU) neural network in forecasting revenue and profitability across three validation regimes of increasing strictness. The results show that model accuracy (R²) declines sharply as the validation regime becomes stricter, that tree-based ensembles are markedly more stable than the recurrent model, and that Decision Curve Analysis confirms substantial managerial value even where predictive accuracy is modest. The findings demonstrate the critical importance of validation-methodology choice when deploying artificial intelligence in financial practice.