A Minimax Optimization Model for Inflation Prediction Under Monetary Growth Uncertainty

Abstract

Forecasting inflation is unusually difficult in economies undergoing rapid financial deepening, because the money growth–price relationship is unstable and the best-performing model changes across macroeconomic regimes. Conventional practice selects models by average out-of-sample accuracy, which systematically favors strategies that excel in calm periods and fail in turbulent ones. This study develops and validates a minimax-regret forecast combination framework, in which inflation forecasts are combined by minimizing the maximum regret across turbulence scenarios rather than average error. Using monthly Chinese data from the Federal Reserve Economic Data archive the M2 money stock (MYAGM2CNM189N, January 1999–August 2019) and the headline consumer price index (CHNCPIALLMINMEI) six competing one-step-ahead models are estimated under a rolling-origin design and evaluated over 52 origins between 2011 and 2019. The premise holds: the money growth–inflation correlation is weak on average (−0.086) but highly unstable, with the rolling 24-month correlation ranging from −0.925 to +0.843. Over the pure out-of-sample window (May 2015–July 2019), the minimax combination reduces worst-scenario RMSE from 0.444 to 0.393 — 11.5% below the best conventional model while matching equal weighting within 0.002 RMSE on average and beating the benchmark overall (0.367 versus 0.391). In stress tests with adverse shocks up to 20%, it attains the lowest maximum forecast displacement at every level (0.894, versus 0.908 conventional and 1.074 for ridge regression). Minimax-regret combination thus delivers materially more robust inflation forecasts under monetary transmission uncertainty, providing a replicable, model-agnostic prescriptive layer for policy institutions.