Deteksi Anomali Harga Bitcoin Menggunakan Pendekatan Hybrid Long Short-Term Memory dan Metode Deteksi Outlier

Dilla Meylia, Prof. Dr. Abdurakhman, S.Si., M.Si.

2025 | Skripsi | STATISTIKA

Pasar cryptocurrency, khususnya Bitcoin, dikenal memiliki volatilitas harga yang sangat tinggi sehingga sering kali memunculkan pergerakan harga ekstrem yang dapat dianggap sebagai anomali. Deteksi terhadap anomali penting untuk mendukung pengambilan keputusan investasi dan manajemen risiko. Penelitian ini bertujuan mengembangkan pendekatan teknikal untuk mendeteksi dan memprediksi anomali harga Bitcoin secara unsupervised dengan menggabungkan model prediksi Long Short-Term Memory (LSTM) dan berbagai metode deteksi anomali, yaitu Z-score, One-Class Support Vector Machine (SVM), Isolation Forest, serta LSTM Autoencoder. Hasil penelitian menunjukkan bahwa model prediktif LSTM terbaik diperoleh pada konfigurasi rolling window 60 hari, yang berhasil menangkap dinamika harga Bitcoin dengan baik, dibuktikan dengan nilai MAPE sebesar 2,15%. Dari segi deteksi anomali, Isolation Forest terbukti paling stabil dan konsisten tanpa asumsi distribusi khusus. Sementara itu, LSTM Autoencoder dengan Rolling Z-score menunjukkan sensitivitas yang tinggi terhadap perilaku ekstrem pasar, dengan distribusi error yang mengikuti pola power-law berekor tebal. Sebaliknya, metode Z-score kurang sesuai karena bergantung pada asumsi distribusi normal, dan One-Class SVM menunjukkan ketidakstabilan hasil. Temuan ini menegaskan bahwa pemilihan metode deteksi anomali harus disesuaikan dengan tujuan sistem yang dibangun, baik untuk kestabilan deteksi maupun kepekaan terhadap anomali ekstrem pasar cryptocurrency.

The cryptocurrency market, particularly Bitcoin, is known for its extreme price volatility, often exhibiting sharp movements that may be classified as anomalies. Detecting such anomalies is crucial to support investment decision-making and risk management. This study aims to develop a technical approach for unsupervised anomaly detection and prediction of Bitcoin price movements by combining a Long Short-Term Memory (LSTM) prediction model with several anomaly detection methods, including Z-score, One-Class Support Vector Machine (SVM), Isolation Forest, and LSTM Autoencoder. The best-performing LSTM model was configured with a 60-day rolling window, effectively capturing Bitcoin's price dynamics with a MAPE of 2.15%. Among the detection methods, Isolation Forest proved to be the most stable and consistent, requiring no assumptions about data distribution. Meanwhile, the LSTM Autoencoder combined with Rolling Z-score showed high sensitivity to extreme market behavior, with reconstruction error distributions following a heavy-tailed power-law pattern. In contrast, the Z-score method was less suitable due to its reliance on the normality assumption, and One-Class SVM results were unstable. These findings emphasize that the choice of anomaly detection method should align with the system's purpose, whether prioritizing detection stability or responsiveness to extreme anomalies in the highly dynamic cryptocurrency market.

Kata Kunci : deteksi anomali, Bitcoin, unsupervised learning, Long Short-Term Memory, unsupervised learning, Isolation Forest, Autoencoder, Z-score, One-Class SVM