Implementasi Arsitektur Temporal Fusion Transformer (Tft) Berbasis Interpretable Multi-Horizon Forecasting Prediksi Volatilitas Harga Aset Kripto Solana

Abstract

The cryptocurrency market, specifically Solana (SOL), exhibits extreme volatility that is difficult for conventional algorithms to predict accurately. Additionally, advanced Deep Learning models often act as black-boxes. This study implements the Temporal Fusion Transformer (TFT) architecture for robust and interpretable multi-horizon volatility prediction of SOL/USDT. High-frequency historical data (1-hour intervals) was acquired via Binance API from January 2023 to December 2024. Data preprocessing included technical indicator feature engineering, Z-Score normalization, and chronological data splitting. Hyperparameters were optimized using Optuna. The TFT architecture performed exceptionally, achieving a Root Mean Squared Error (RMSE) of 2.5191 USDT and a Coefficient of Determination (R2) of 0.9987. Probabilistically, the model successfully mapped risk ranges with a Weighted Mean Quantile Loss of 0.4629. Explainable AI analysis revealed that predictions were heavily influenced by historical closing price, RSI, Volume, and MACD. Innovatively, TFT outputs were integrated into a Decision Support System utilizing Google Gemini 2.5-Flash to automatically generate market analysis narratives. In conclusion, this study produced a robust forecasting model, successfully resolving the black-box problem into a transparent virtual analyst assistant.