Prediksi harga Bitcoin dan Ethereum berdasarkan persentase kenaikan atau penurunan dalam jangka panjang dan pendek menggunakan metode Stacked LSTM

The high volatility of cryptocurrency prices, such as Bitcoin (BTC) and Ethereum (ETH), presents a significant challenge for investors, requiring accurate price prediction models for both short-term and long-term strategies. Deep learning methods, particularly Stacked Long Short-Term Memory (LSTM), show great potential for handling complex and non-linear time series data. This research aims to evaluate the performance of the Stacked LSTM model in predicting the daily closing prices of BTC and ETH, as well as to analyze the potential of both assets for short-term (7-day) and long-term (90-day) investment.
This study utilizes a Stacked LSTM architecture trained on daily historical data (Open, High, Low, Close, Volume). The data underwent a pre-processing stage, including Min-Max Scaling normalization. Model performance was quantitatively evaluated using 5-Fold Cross-Validation with Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R² Score metrics. A qualitative analysis was conducted by visually comparing the prediction results on 7-day and 90-day horizons.
The results indicate that the Stacked LSTM model is highly accurate and stable. Quantitative evaluation yielded an average R² Score of 0.9821 for Bitcoin and 0.9677 for Ethereum, with very low MAE and RMSE values. Visually, the short-term (7-day) predictions proved to be highly precise and nearly identical to the actual data for both assets. For the long term (90 days), the model successfully captured the general price movement trend, although its precision decreased in capturing sharp daily fluctuations. The Stacked LSTM model proved effective for predicting the prices of both assets. For short-term investment, both assets show equally good predictability potential. For long-term investment, Bitcoin (BTC) demonstrates slightly more stable and better trend predictability than Ethereum (ETH) based on the model's results.