Prediksi Harga Bitcoin Menggunakan Long Short Term Memory Dengan Optimasi Hyperparameter

Adi, Garin Legantoro and , Azizah Fatmawati, S.T, M.Cs. (2026) Prediksi Harga Bitcoin Menggunakan Long Short Term Memory Dengan Optimasi Hyperparameter. Skripsi thesis, Universitas Muhammadiyah Surakarta.

	PDF (Naskah Publikasi) NASKAH PUBLIKASI.pdf Download (688kB)
	PDF (Surat Pernyataan Publikasi) SURAT PERNYATAAN.pdf Restricted to Repository staff only Download (831kB)

Abstract

The high price volatility of Bitcoin presents both a significant forecasting challenge and an opportunity for developing robust predictive models. This study proposes a deep learning based approach by utilizing the Long Short-Term Memory (LSTM) architecture optimized through Grid Search hyperparameter tuning. Daily Bitcoin closing price data (BTC-USD) spanning 2021 to 2025 were collected from Yahoo Finance and processed through normalization and sequence construction stages prior to model training. Two experimental scenarios were compared: an LSTM model with default parameter settings and one subjected to systematic hyperparameter optimization. Model performance was assessed using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE). The optimal configuration identified consisted of 100 LSTM units, a batch size of 32, and a learning rate of 0.01. Grid Search optimization yielded a substantial improvement in predictive accuracy, reducing RMSE by 23.18% from USD 2,968.22 to USD 2,280.21 and lowering MAE from USD 2,251.83 to USD 1,750.84. Subsequently, external testing was conducted using 59 external data points from the period of January 1 to February 28, 2026, employing a rolling one-step-ahead forecasting method without retraining. In the external testing, the default model achieved an RMSE of USD 9,986.93 and an MAE of USD 7,560.51, whereas the Grid Search model yielded an RMSE of USD 11,945.32 and an MAE of USD 10,080.43. These results indicate that the default model possesses superior generalization capabilities on external data. Thus, while hyperparameter optimization can enhance performance during internal testing, the optimal configuration must still be tested on new data to verify the model's generalization ability.

Item Type:	Thesis (Skripsi)
Uncontrolled Keywords:	Bitcoin, LSTM, Grid Search, time series, price prediction
Subjects:	H Social Sciences > HG Finance > Investasi T Technology > Information Technology T Technology > Information Technology > Software. Aplication T Technology > Information Technology > Software. Aplication > Software Engineering
Divisions:	Fakultas Komunikasi dan Informatika > S1 Teknik Informatika
Depositing User:	GARIN LEGANTORO ADI
Date Deposited:	04 Aug 2026 04:27
Last Modified:	04 Aug 2026 04:27
URI:	http://eprints.ums.ac.id/id/eprint/146739

Actions (login required)

View Item