Electronic Theses and Dissertation

Volatilitas harga bitcoin yang tinggi dan karakteristik datanya yang non-stasioner menjadikan peramalan harga sebagai tantangan kompleks yang memerlukan pendekatan pemodelan berbasis deep learning. penelitian ini bertujuan untuk membandingkan kinerja tiga arsitektur deep learning, yaitu long short-term memory (lstm), gated recurrent unit (gru), dan transformer, dalam meramalkan harga penutupan harian bitcoin terhadap rupiah (btc/idr). data yang digunakan merupakan data historis harga penutupan harian yang diperoleh dari coinmarketcap, mencakup periode agustus 2010 hingga maret 2026, dengan total 5.714 observasi. seluruh model dioptimalkan menggunakan metode grid search terhadap kombinasi hyperparameter yang telah ditentukan, serta dievaluasi pada tiga skenario pembagian data, yaitu 90:10, 80:20, dan 70:30. kinerja model diukur menggunakan metrik mae, rmse, dan mape. hasil evaluasi menunjukkan bahwa model gru dengan rasio pembagian data 90:10 menghasilkan performa terbaik, dengan nilai mae dan rmse paling kecil serta mape sebesar 2,80%. model gru selanjutnya digunakan untuk meramalkan harga bitcoin selama 120 hari ke depan (april–juli 2026), dengan proyeksi harga berkisar antara rp1,116 miliar hingga rp1,154 miliar dan tren penurunan bertahap yang semakin stabil mendekati akhir periode.

The high price volatility and non-stationary nature of Bitcoin data make price forecasting a complex challenge that requires deep learning-based modeling approaches. This study aims to compare the performance of three deep learning architectures Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer in forecasting the daily closing price of Bitcoin against the Indonesian Rupiah (BTC/IDR). The dataset consists of historical daily closing prices obtained from CoinMarketCap, covering the period from August 2010 to March 2026, with a total of 5,714 observations. All models were optimized using grid search across predefined hyperparameter combinations and evaluated under three data split scenarios: 90:10, 80:20, and 70:30. Model performance was assessed using MAE, RMSE, and MAPE metrics. The evaluation results indicate that the GRU model with a 90:10 data split achieved the best performance, with the lowest MAE and RMSE values and a MAPE of 2.80%. The best-performing GRU model was subsequently applied to forecast Bitcoin prices for 120 days ahead (April–July 2026), projecting prices in the range of IDR 1.116 billion to IDR 1.154 billion, with a gradual downward trend that becomes increasingly stable toward the end of the forecast period.