Regime-Aware Consumer Price Index Dynamics in India:A Multi-Methodology Econometric and Machine Learning Framework, 1952-2024
Arnav Aggarwal
21/07/2026
India's Consumer Price Index (CPI) is influenced by both national economic issues and global economic pressures. This study develops a forecasting framework for predicting CPI using nine economic variables—food price inflation (CFPI), wholesale inflation (WPI), Brent crude oil prices, INR–USD exchange rates, money supply (M3), government expenditure, housing prices, and unemployment—across 73 annual observations (1952–April 2024). I examine 15 forecasting models, including classical time-series models, GARCH-family volatility models, ensemble machine learning methods (XGBoost, LightGBM, Random Forest, Lasso, and Ridge), and LSTM deep learning. Stationarity of the CPI series is confirmed through Augmented Dickey–Fuller, KPSS, Phillips–Perron, and Zivot–Andrews tests. Bai–Perron structural break analysis identifies three major breaks in the data: 1964 (pre-Green Revolution and global food crisis), 1974 (first oil price shock), and 2014 (Reserve Bank of India's inflation-targeting framework). GARCH(1,1) conditional variance estimation yields a persistence coefficient of α̂ + β̂ = 0.86, consistent with periods of the Great Inflation and the Great Moderation. In out-of-sample forecasting, XGBoost achieves the strongest predictive performance (RMSE = 0.0073, MAPE = 12.46%), representing a 73% reduction in RMSE relative to the best classical benchmark, MA(1). Feature importance analysis further indicates that CFPI accounts for approximately 60% of total explanatory importance across the testing period. These findings have direct implications for monetary policy design and real-time inflation surveillance in emerging market economies.