Peramalan Harga Beras Real Time melalui Horse Race Model di Sulawesi

Abstract

Harga beras di Pulau Sulawesi rentan terhadap gejolak akibat ketimpangan pasokan antardaerah, integrasi pasar spasial, dan keterkaitan lintas kualitas beras. Model ekonometrika konvensional kurang mampu menangani data berdimensi tinggi sehingga diperlukan pendekatan peramalan yang lebih andal sebagai landasan intervensi pasar yang bersifat pre-emptive untuk mendukung ketahanan pangan. Penelitian ini bertujuan menganalisis performa akurasi peramalan pada berbagai horizon waktu, mengidentifikasi model dan teknik regularisasi terbaik melalui horse race, serta mengevaluasi kontribusi sinyal harga nasional terhadap stabilitas akurasi jangka panjang. Data deret waktu harian periode April 2019 sampai Januari 2026 dianalisis menggunakan regresi teregularisasi (lasso, ridge, dan elastic net) dengan spesifikasi nested ARDL pada empat horizon peramalan. Hasil penelitian menunjukkan akurasi peramalan menurun seiring bertambahnya horizon (predictability decay). Model AR lasso terbukti paling unggul untuk peramalan jangka pendek. Model yang mengintegrasikan faktor nasional dengan elastic net mendominasi pada horizon jangka panjang. Integrasi sinyal harga nasional terbukti meningkatkan stabilitas akurasi peramalan jangka panjang.

Rice prices in Sulawesi are prone to volatility due to regional supply disparities, spatial market integration, and cross-grade price effects. Conventional econometric approaches are less capable of handling high dimensional data, thus necessitating a more reliable forecasting framework as the basis for pre-emptive market intervention to support food security. This study aims to analyze forecasting accuracy across multiple time horizons, identify the optimal model and regularization technique through a horse race mechanism, and evaluate the contribution of national price signals to long-term forecast stability. Daily time series data spanning April 2019 to January 2026 were analyzed using penalized regressions (lasso, ridge, and elastic net) within nested ARDL specifications across four forecasting horizons. The results show that forecast accuracy declines as horizons extend (predictability decay). The AR lasso model proves most effective for short-term forecasting. The model integrating national factors with elastic net dominates at longer horizons. The integration of national price signals improves forecast stability at longer horizons.