Bitcoin merupakan salah satu jenis cryptocurrency yang banyak digunakan karena transaksinya yang aman, cepat, dan berpotensi memberikan keuntungan signifikan. Namun, volatilitas harga yang tinggi membuat aktivitas transaksi berisiko, karena pergerakan harga tidak hanya dipengaruhi oleh faktor internal, tetapi juga faktor eksternal seperti sentimen publik dan Google Trends Index (GTI). Penelitian ini bertujuan membandingkan algoritma XGBoost regression dan LSTM for regression dalam memprediksi harga penutupan bitcoin dengan mengintegrasikan variabel harga harian, sentiment, dan GTI ke dalam model regresi yang disesuaikan dengan karakteristik data penelitian, dimana data yang yang digunakan bersifat non-linear, tidak berdistribusi normal, dan mengandung unsur time series. Berdasarkan hasil pengujian, model XGBoost regression terbaik diperoleh pada skenario dengan variabel eksternal. Namun menghasilkan nilai RMSE sebesar 5169,898 USD dan R2-Score sebesar -13%, yang menunjukkan adanya overfitting dan model kurang tepat untuk data time series. Sebaliknya, model LSTM for regression dengan variabel eksternal dan kombinasi hyperparameter terbaik menunjukkan performa yang lebih unggul dengan RMSE sebesar 1378,55 USD dan R2-Score sebesar 92%. Model ini tidak menunjukkan indikasi overfitting dan mampu mereplikasi pola pergerakan harga secara akurat. Hal ini menunjukkan bahwa LSTM for regression lebih mampu mengenali pola temporal dalam data historis. Selain itu, fitur harga historis, khususnya Open teridentifikasi sebagai variabel paling dominan berdasarkan hasil analisis menggunakan metode SHAP. Abstract Bitcoin is one type of cryptocurrency that is widely used because its transactions are safe, fast, and have the potential to provide significant profits. However, high price volatility makes transaction activities risky, because price movements are not only influenced by internal factors, but also external factors such as public sentiment and the Google Trends Index (GTI). This study aims to compare the XGBoost regression and LSTM for regression algorithms in predicting bitcoin closing prices by integrating daily price, sentiment, and GTI variables into a regression model that is adjusted to the characteristics of the research data, where the data used is non-linear, not normally distributed, and contains time series elements. Based on the test results, the best XGBoost regression model was obtained in the scenario with external variables. However, it produces an RMSE value of 5169.898 USD and an R2-Score of -13%, which indicates overfitting and the model is less appropriate for time series data. In contrast, the LSTM for regression model with external variables and the best combination of hyperparameters shows superior performance with an RMSE of 1378.55 USD and an R2-Score of 92%. This model does not show any indication of overfitting and is able to replicate price movement patterns accurately. This shows that LSTM for regression is better able to recognize temporal patterns in historical data. In addition, historical price features, especially Open, are identified as the most dominant variables based on the results of the analysis using the SHAP method.
Bahasa Indonesia
Bitcoin merupakan salah satu jenis cryptocurrency yang banyak digunakan karena transaksinya yang aman, cepat, dan berpotensi memberikan keuntungan signifikan. Namun, volatilitas harga yang tinggi membuat aktivitas transaksi berisiko, karena pergerakan harga tidak hanya dipengaruhi oleh faktor internal, tetapi juga faktor eksternal seperti sentimen publik dan Google Trends Index (GTI). Penelitian ini bertujuan membandingkan algoritma XGBoost regression dan LSTM for regression dalam memprediksi harga penutupan bitcoin dengan mengintegrasikan variabel harga harian, sentiment, dan GTI ke dalam model regresi yang disesuaikan dengan karakteristik data penelitian, dimana data yang yang digunakan bersifat non-linear, tidak berdistribusi normal, dan mengandung unsur time series. Berdasarkan hasil pengujian, model XGBoost regression terbaik diperoleh pada skenario dengan variabel eksternal. Namun menghasilkan nilai RMSE sebesar 5169,898 USD dan R2-Score sebesar -13%, yang menunjukkan adanya overfitting dan model kurang tepat untuk data time series. Sebaliknya, model LSTM for regression dengan variabel eksternal dan kombinasi hyperparameter terbaik menunjukkan performa yang lebih unggul dengan RMSE sebesar 1378,55 USD dan R2-Score sebesar 92%. Model ini tidak menunjukkan indikasi overfitting dan mampu mereplikasi pola pergerakan harga secara akurat. Hal ini menunjukkan bahwa LSTM for regression lebih mampu mengenali pola temporal dalam data historis. Selain itu, fitur harga historis, khususnya Open teridentifikasi sebagai variabel paling dominan berdasarkan hasil analisis menggunakan metode SHAP. Abstract Bitcoin is one type of cryptocurrency that is widely used because its transactions are safe, fast, and have the potential to provide significant profits. However, high price volatility makes transaction activities risky, because price movements are not only influenced by internal factors, but also external factors such as public sentiment and the Google Trends Index (GTI). This study aims to compare the XGBoost regression and LSTM for regression algorithms in predicting bitcoin closing prices by integrating daily price, sentiment, and GTI variables into a regression model that is adjusted to the characteristics of the research data, where the data used is non-linear, not normally distributed, and contains time series elements. Based on the test results, the best XGBoost regression model was obtained in the scenario with external variables. However, it produces an RMSE value of 5169.898 USD and an R2-Score of -13%, which indicates overfitting and the model is less appropriate for time series data. In contrast, the LSTM for regression model with external variables and the best combination of hyperparameters shows superior performance with an RMSE of 1378.55 USD and an R2-Score of 92%. This model does not show any indication of overfitting and is able to replicate price movement patterns accurately. This shows that LSTM for regression is better able to recognize temporal patterns in historical data. In addition, historical price features, especially Open, are identified as the most dominant variables based on the results of the analysis using the SHAP method.
Affiliations come from OpenAlex; journal OAI-PMH metadata carries none. Authors without an OpenAlex ID are matched by name, which can merge or split people.