FOOD PRICE SPIKE DETECTION IN INDONESIA: A COMPARATIVE STUDY OF DEEP LEARNING ARCHITECTURES AND SPATIOTEMPORAL PRICE PROPAGATION ANALYSIS

Abstract

Food price spikes pose a severe threat to household economic security and national price stability in developing nations, yet existing early warning mechanisms are predominantly reactive and fail to explicitly capture spatial shock propagation. This paper introduces an end-to-end multi-task deep learning pipeline designed to simultaneously perform multi-step log-return price forecasting and forward-looking spike classification over a shared feature encoder. Deployed on a 33 variable daily panel dataset (2017–2025) across three strategic Indonesian commodities: rice, shallots, and curly red chilli and six representative provinces, the framework integrates a three-layer class imbalance mitigation strategy consisting of DiceFocal loss, class-balanced sampling, and post-training threshold calibration. A controlled benchmarking of four state-of-the-art architectures demonstrates distinct specialization: PatchTST yields superior threshold-independent discriminability (AUPRC = 0.797) via local patch tokenization, the recurrent LSTM baseline minimizes quantitative regression error (MAE = 0.360) due to its sequential inductive bias, and Crossformer achieves the highest sensitivity (Recall = 0.860). Crucially, by leveraging the inverted attention mechanism of iTransformer (Parameters ≈ 83K), we extract a latent cross-variate dependency matrix that unmasks asymmetric upstream-to-downstream price transmission corridors, identifying West Java and North Sumatra rice producers as critical systemic shock anchors. These insights establish an empirically grounded decision-support tool for institutional bodies like BULOG and BAPANAS, facilitating a policy paradigm shift from blanket retail market interventions toward proactive, localized upstream supply stabilization.