A Machine Learning-Based Early Warning System for Inflation Crisis Prediction: A Multi-Country Panel Study

Abstract

Inflation crises on economy can be mentioned, the current situation with modern models designed for forecasting such phenomena is not flawless concerning the definition of global crises, look-ahead bias in feature definition, and the absence of benchmarking in statistics. We would like to offer our readers the ML-based solution that enables us to forecast inflation crises for a panel of 152 countries for years from 1980 till 2024 according to the information obtained from the IMF WEO database. Using the variables such as GDP growth, unemployment rate, balance of current accounts, and government debt to GDP ratio as the inputs, we are able to predict inflation crisis occurrences using z-score methodology without the leakage problem existing when dealing with percentiles of global crises definition. LR, RF, and GBM classifiers have been trained and analyzed as opposed to economic benchmarks AR(1) and VAR(1) through walk-forward temporal cross-validation. The results of our research show that RF and GBM classifiers exhibit remarkable out-of-sample predictive ability in terms of AUC values being equal to 0.841 and 0.848 respectively (it was proven to be statistically significant against benchmark AR(1) with AUC equal to 0.770, p-value equal to 0.002, 95% confidence interval of [0.022; 0.110] for block-country approach with 2,000 draws).