Pengembangan Sistem Peringatan Dini Fluktuasi Harga Beras Sumatra: Integrasi Model Ekonometrika dan Machine Learning

Abstract

Fluktuasi harga beras di Sumatra dipengaruhi oleh ketimpangan pasokan yang memerlukan sistem deteksi dini akurat. Penelitian ini mengevaluasi kinerja peramalan model integrasi ekonometrika dan machine learning pada 822 deret waktu nasional melalui prosedur horse race komprehensif yang melibatkan 471.684 iterasi komputasi. Temuan menunjukkan bahwa klaster model multivariat (ARDL) secara konsisten mengungguli model univariat dalam menangkap sinyal arbitrase nasional. Teridentifikasi adanya transisi dominasi dari faktor internal pada horizon harian menuju faktor laten nasional pada horizon bulanan. Strategi regularisasi Ridge terbukti paling tangguh, sekaligus mengonfirmasi teori ilusi sparsitas di mana informasi harga tersebar secara akumulatif dalam jaringan pasar. Hasil studi ini merekomendasikan implementasi sistem peringatan dini adaptif berbasis subregional guna menjaga stabilitas harga pangan di Pulau Sumatra.

Rice price volatility in Sumatra is driven by supply imbalances requiring accurate early detection. This study evaluates the performance of integrated econometric and machine learning models using 822 national time series through a comprehensive horse race procedure involving 471,684 computational iterations. Findings reveal that multivariate model clusters (ARDL) consistently outperform univariate models by capturing national arbitrage signals. A dominance shift was observed from internal factors at daily horizons to national latent factors at monthly horizons. The Ridge regularization technique emerged as the most robust strategy, confirming the illusion of sparsity theory where price information is accumulatively distributed across the market network. This research recommends implementing an adaptive sub-regional early warning system to maintain food price stability across Sumatra.