Temperature shocks and food inflation: Multicountry evidence from visual time-series transformers and attention-based...

This study examines whether monthly temperature change contains predictive information for food price inflation (FPI) across nine economies-Brazil, France, Germany, India, Japan, Mexico, the Netherlands, Russia, and the United States-by combining visual time-series representations with deep learning (DL) and explainable machine learning (ML). We assemble country-specific series covering consumer food prices and the following predictors: Temperature change, exchange rate (USD), Brent oil price, industrial production, EPU, and VIX. Building on these data, we benchmark conventional ML/DL regressors (CNN, GRU, LSTM, SVR, RF, KNN, LGBM) against a two-stage hybrid pipeline, GAFWave-A2FSNet: (i) Gramian Angular Field (GAF) and Continuous Wavelet Transform (CWT) convert sequences into images; (ii) BEiT extracts embeddings; (iii) an AutoEncoder-Attention feature selection (A2FS) condenses representations; (iv) country-level regressors are trained and evaluated via a strictly time-aware 5-fold time-series cross-validation (TSCV) framework to ensure robust out-of-sample generalization and prevent temporal leakage; (v) interpretability is established using grouped feature importance, partial dependence (PDP), point-derivative sensitivity, and residual-based CUSUM/CUSUMSQ diagnostics. The results indicate that models trained on raw tabular inputs often exhibit limited and unstable out-of-sample performance, whereas the proposed hybrid framework yields consistent improvements in predictive accuracy across all countries. These findings suggest that visual time-frequency representations enhance generalization under realistic chronological validation settings. Temperature change emerges as a consistently relevant, yet heterogeneous, predictor of FPI, with its predictive contribution varying across countries and operating through nonlinear and threshold-dependent patterns. Stability diagnostics further support the temporal consistency of model performance despite localized fluctuations. Overall, the findings highlight the importance of incorporating climate-related information into inflation forecasting models and support the further evaluation of integrated representation-based hybrid frameworks in heterogeneous macroeconomic environments.