Perbandingan Performa Model Deep Learning dalam Prediksi Harga Protein Hewani dengan Pendekatan Multiple-Input Multiple-Output

Abstract

Komoditas protein hewani, yaitu daging ayam, daging sapi, dan telur ayam, berperan penting dalam upaya penanggulangan stunting di Indonesia. Harga komoditas tersebut berfluktuasi akibat ketergantungan pada biaya pakan, gangguan rantai pasok, dan perubahan permintaan, sehingga prediksi harga yang akurat diperlukan untuk mendukung kebijakan stabilisasi harga. Penelitian ini bertujuan membandingkan performa model recurrent neural network (RNN), long short-term memory (LSTM), dan gated recurrent unit (GRU) menggunakan pendekatan multiple-input multiple-output (MIMO) dalam memprediksi harga harian ketiga komoditas secara simultan. Data yang digunakan berupa 2.922 amatan harian periode 1 Januari 2018 hingga 31 Desember 2025 dari laman Pusat Informasi Harga Pangan Strategis (PIHPS) Bank Indonesia, dioptimalkan menggunakan time series 5-fold cross-validation dengan skema expanding window. Hasil penelitian menunjukkan bahwa ketiga model memberikan performa yang baik dengan MAPE yang lebih kecil dari 5%, namun model MIMO-GRU memberikan performa terbaik didasarkan nilai metrik evaluasi terkecil yaitu RMSE sebesar 1.112,686 dan MAPE sebesar 1,423%, lebih baik dibandingkan MIMO LSTM dan MIMO RNN pada data uji. Hasil peramalan 30 hari ke depan mengindikasikan keterbatasan model dalam mengantisipasi pembalikan arah tren, mengingat perubahan harga komoditas pangan secara aktual dapat terjadi sewaktu-waktu akibat faktor-faktor yang tidak sepenuhnya tercermin dalam data historis.

Animal protein commodities, namely broiler chicken, beef, and chicken eggs, play an important role in stunting prevention efforts in Indonesia. The prices of these commodities fluctuate due to dependence on feed costs, supply chain disruptions, and changes in demand, making accurate price prediction necessary to support price stabilization policies. This study aims to compare the performance of recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU) models using a multiple-input multiple-output (MIMO) approach in simultaneously predicting the daily prices of the three commodities. The data used consisted of 2,922 daily observations covering the period from January 1, 2018 to December 31, 2025, obtained from the Strategic Food Price Information Center (PIHPS) of Bank Indonesia, optimized using time series 5-fold cross-validation with an expanding window scheme. The results showed that all three models performed well with MAPE values below 5%; however, the MIMO-GRU model achieved the best performance based on the smallest evaluation metric values, with an overall RMSE of 1,112.686 and MAPE of 1.423%, outperforming MIMO-LSTM and MIMO-RNN on the test data. The 30-day ahead forecasting results indicate limitations of the model in anticipating trend reversals, given that actual changes in food commodity price trends can occur at any time due to factors not fully reflected in historical data.