Perbandingan Model LSTM dan GRU dalam Peramalan Harga Cryptocurrency serta Pengukuran Risikonya dengan Value at Risk

Abstract

Cryptocurrency memiliki volatilitas tinggi sehingga peramalan harga dan pengukuran risiko penting dalam pengambilan keputusan investasi. Penelitian ini bertujuan menganalisis karakteristik harga penutupan harian dan log return Bitcoin (BTC) dan Ethereum (ETH), membandingkan kinerja model long short-term memory (LSTM) dan gated recurrent unit (GRU) dalam peramalan harga, serta mengukur risiko pasar menggunakan Value at Risk (VaR). Data berupa harga penutupan harian BTC dan ETH periode 1 Januari 2021 hingga 31 Desember 2025 sebanyak 1.826 observasi per aset. Kinerja model dievaluasi menggunakan root mean square error (RMSE), mean absolute error (MAE), dan mean absolute percentage error (MAPE), sedangkan VaR dihitung dengan historical simulation. Hasil menunjukkan bahwa log return kedua aset stasioner, tidak normal, dan mengindikasikan fat tails, dengan volatilitas tahunan ETH lebih tinggi daripada BTC. Berdasarkan MAPE, GRU sedikit lebih baik daripada LSTM, dengan nilai terbaik 1.62% untuk BTC dan 3.16% untuk ETH. Nilai VaR dari return prediksi LSTM dan GRU lebih kecil secara absolut dibandingkan VaR empiris, yang menunjukkan bahwa return hasil prediksi cenderung menghasilkan estimasi risiko yang lebih rendah terhadap potensi kerugian ekstrem.

Cryptocurrencies are highly volatile assets, making price forecasting and risk measurement essential for investment decision-making. This study analyzes the daily closing prices and log returns of Bitcoin (BTC) and Ethereum (ETH), compares the performance of long short-term memory (LSTM) and gated recurrent unit (GRU) models for price forecasting, and measures market risk using Value at Risk (VaR). The data set consists of daily closing prices from 1 January 2021 to 31 December 2025, comprising 1,826 observations for each asset. Model performance is evaluated using root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), while VaR is estimated using the historical simulation method. The results indicate that the log return series of both assets are stationary, non-normally distributed, and exhibit fat-tailed characteristics, with ETH showing higher annual volatility than BTC. Based on the MAPE criterion, the GRU model performs slightly better than the LSTM, model achieving the best MAPE values of 1.62% for BTC and 3.16% for ETH. Furthermore, the VaR values derived from LSTM- and GRU-predicted returns are lower in absolute terms than empirical VaR values, indicating that predicted returns tend to generate lower estimates of potential extreme losses. These findings suggest that GRU provides superior forcasting performance, while both models may underestimate extreme market risk in cryptocurrency returns