Perbandingan Peramalan Harga Penutup Saham Harian BMRI Menggunakan ARIMA dan LSTM Berbasis Recursive dan WalkForward Validation

Abstract

Pergerakan harga saham yang dinamis dan non-linear memerlukan metode peramalan yang akurat untuk mitigasi risiko investasi. Penelitian ini membandingkan kinerja model ARIMA dan LSTM dalam meramalkan harga penutupan harian saham PT Bank Mandiri (Persero) Tbk (BMRI) periode Juli 2003 hingga September 2025 menggunakan strategi recursive forecasting dan walk-forward validation. Hasil menunjukkan bahwa pada recursive forecasting, ARIMA(0,1,0) menghasilkan MAPE 1.91%, sedangkan LSTM mendapatkan MAPE 2.12%, tetapi kedua model cenderung menghasilkan prediksi statis sehingga kurang adaptif terhadap fluktuasi harian pasar. Pada walk-forward validation, performa model meningkat signifikan terhadap volatilitas pasar. Model ARIMA(2,1,4) menjadi yang terbaik pada basis ini dengan MAPE 1.34%, sedangkan LSTM memperoleh MAPE 4.04%. Penelitian menyimpulkan bahwa pendekatan recursive forecasting kurang memadai sebagai dasar pengambilan keputusan investasi, sedangkan kombinasi ARIMA dan walk-forward validation memberikan hasil peramalan yang lebih baik dibandingkan LSTM berbasis walk-forward validation.

The dynamic and non-linear movement of stock prices requires accurate forecasting methods for investment risk mitigation. This study compares the performance of ARIMA and LSTM models in forecasting the daily closing stock prices of PT Bank Mandiri (Persero) Tbk (BMRI) from July 2003 to September 2025 using recursive forecasting and walk-forward validation strategies. The results show that under recursive forecasting, ARIMA achieved a MAPE of 1.91% and an RMSE of 114.74, while LSTM obtained a MAPE of 2.12% and an RMSE of 125.43. However, both models tended to produce static predictions and were less adaptive to daily market fluctuations. Under walk-forward validation, model performance improved significantly in handling market volatility. The ARIMA(2,1,4) model achieved the best performance with a MAPE of 1.34% and an RMSE of 81.95, while LSTM obtained a MAPE of 4.04% and an RMSE of 214.00. The study concludes that recursive forecasting is less suitable as a basis for investment decision-making, whereas the combination of ARIMA and walk-forward validation provides best precise forecasting results than walk-forward-based LSTM.