Food prices worldwide have risen substantially following recent geopolitical and supply-chain disruptions. New Zealand, with its remote location and absence of land-based freight connections to other national economies, provides an informative case study for food-price forecasting. This study analyses monthly item-level food prices from June 2006 to September 2024 and compares Exponential Smoothing, SARIMAX, principal-component vector autoregression (PCA-VAR), a global long short-term memory network (LSTM), a transfer-learning LSTM, and four transparent baseline models within a leakage-safe evaluation framework. A fixed 50/50 SARIMAX–Exponential Smoothing combination is constructed by jointly selecting its component specifications using blocked expanding-window validation data; equal-weight averaging itself is not presented as a new forecasting algorithm. All preprocessing, scaling, parameter selection, and hybrid construction are completed before the untouched October 2022 to September 2024 test period is evaluated. Of 155 candidate series, 135 satisfied the training-only temporal-coverage rule. On the common test panel, the hybrid achieved the lowest mean absolute error (MAE = 0.4463), a 1.33% reduction relative to the Simple Average benchmark, together with the lowest symmetric mean absolute percentage error (sMAPE = 5.7926) and mean absolute scaled error (MASE = 1.0722). Exponential Smoothing achieved the lowest mean squared error (MSE = 1.0149) and root mean squared error (RMSE = 1.0074). The modest MAE gain and predominantly non-significant item-level Diebold–Mariano comparisons indicate that the main contribution is the reproducible, leakage-safe framework for model selection, multi-metric evaluation, uncertainty reporting, and Business Intelligence integration rather than a claim of universal performance superiority.