Chasing Alpha in Bitcoin: Forecasting Price Direction and Generating Alpha with Random Forest and LASSO-Logistic
(2025) NEKN02 20251Department of Economics
- Abstract
- This study examines whether machine learning (ML) models can forecast daily Bitcoin price
directions and generate statistically significant alpha in trading strategies. Employing a dual-model approach, we utilize a Random Forest (RF) classifier, leveraging its non-linear strengths and a LASSO-penalized logistic regression as a linear benchmark. Both models are trained on a comprehensive dataset of Bitcoin related, macroeconomic, on-chain, and technical indicators from January 2018 to March 2025. Automated hyperparameter tuning was employed to optimize model performance, assessed through classification metrics (accuracy, log-loss, AUC-ROC) and simulated trading outcomes (directional, long-only, short-only strategies). RF outperforms... (More) - This study examines whether machine learning (ML) models can forecast daily Bitcoin price
directions and generate statistically significant alpha in trading strategies. Employing a dual-model approach, we utilize a Random Forest (RF) classifier, leveraging its non-linear strengths and a LASSO-penalized logistic regression as a linear benchmark. Both models are trained on a comprehensive dataset of Bitcoin related, macroeconomic, on-chain, and technical indicators from January 2018 to March 2025. Automated hyperparameter tuning was employed to optimize model performance, assessed through classification metrics (accuracy, log-loss, AUC-ROC) and simulated trading outcomes (directional, long-only, short-only strategies). RF outperforms LASSO, achieving 56.14% out-of-sample accuracy (compared to 51.14% achieved by LASSO), and superior returns (e.g., 107.75% for the directional strategy), surpassing a buy-and-hold benchmark. However, neither model delivers statistically significant alpha against Bitcoin CAPM, Equity CAPM, or Fama-French models at a 0.05 significance level, though the strategies exhibit notable exposures to systematic risk factors. These findings suggest that while ML enhances forecasting and portfolio performance in cryptocurrency markets, it struggles to produce consistent alpha, supporting the efficient market hypothesis. (Less) - Popular Abstract
- This study examines whether machine learning (ML) models can forecast daily Bitcoin price
directions and generate statistically significant alpha in trading strategies. Employing a dual-model approach, we utilize a Random Forest (RF) classifier, leveraging its non-linear strengths and a LASSO-penalized logistic regression as a linear benchmark. Both models are trained on a comprehensive dataset of Bitcoin related, macroeconomic, on-chain, and technical indicators from January 2018 to March 2025. Automated hyperparameter tuning was employed to optimize model performance, assessed through classification metrics (accuracy, log-loss, AUC-ROC) and simulated trading outcomes (directional, long-only, short-only strategies). RF outperforms... (More) - This study examines whether machine learning (ML) models can forecast daily Bitcoin price
directions and generate statistically significant alpha in trading strategies. Employing a dual-model approach, we utilize a Random Forest (RF) classifier, leveraging its non-linear strengths and a LASSO-penalized logistic regression as a linear benchmark. Both models are trained on a comprehensive dataset of Bitcoin related, macroeconomic, on-chain, and technical indicators from January 2018 to March 2025. Automated hyperparameter tuning was employed to optimize model performance, assessed through classification metrics (accuracy, log-loss, AUC-ROC) and simulated trading outcomes (directional, long-only, short-only strategies). RF outperforms LASSO, achieving 56.14% out-of-sample accuracy (compared to 51.14% achieved by LASSO), and superior returns (e.g., 107.75% for the directional strategy), surpassing a buy-and-hold benchmark. However, neither model delivers statistically significant alpha against Bitcoin CAPM, Equity CAPM, or Fama-French models at a 0.05 significance level, though the strategies exhibit notable exposures to systematic risk factors. These findings suggest that while ML enhances forecasting and portfolio performance in cryptocurrency markets, it struggles to produce consistent alpha, supporting the efficient market hypothesis. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9195781
- author
- Dabija Ciurea, Stefan LU and Schmitt, Jonas LU
- supervisor
- organization
- course
- NEKN02 20251
- year
- 2025
- type
- H1 - Master's Degree (One Year)
- subject
- keywords
- Machine Learning, Forecasting, Bitcoin, Cryptocurrency, Jensen’s Alpha
- language
- English
- id
- 9195781
- date added to LUP
- 2025-09-12 10:41:44
- date last changed
- 2025-09-12 10:41:44
@misc{9195781, abstract = {{This study examines whether machine learning (ML) models can forecast daily Bitcoin price directions and generate statistically significant alpha in trading strategies. Employing a dual-model approach, we utilize a Random Forest (RF) classifier, leveraging its non-linear strengths and a LASSO-penalized logistic regression as a linear benchmark. Both models are trained on a comprehensive dataset of Bitcoin related, macroeconomic, on-chain, and technical indicators from January 2018 to March 2025. Automated hyperparameter tuning was employed to optimize model performance, assessed through classification metrics (accuracy, log-loss, AUC-ROC) and simulated trading outcomes (directional, long-only, short-only strategies). RF outperforms LASSO, achieving 56.14% out-of-sample accuracy (compared to 51.14% achieved by LASSO), and superior returns (e.g., 107.75% for the directional strategy), surpassing a buy-and-hold benchmark. However, neither model delivers statistically significant alpha against Bitcoin CAPM, Equity CAPM, or Fama-French models at a 0.05 significance level, though the strategies exhibit notable exposures to systematic risk factors. These findings suggest that while ML enhances forecasting and portfolio performance in cryptocurrency markets, it struggles to produce consistent alpha, supporting the efficient market hypothesis.}}, author = {{Dabija Ciurea, Stefan and Schmitt, Jonas}}, language = {{eng}}, note = {{Student Paper}}, title = {{Chasing Alpha in Bitcoin: Forecasting Price Direction and Generating Alpha with Random Forest and LASSO-Logistic}}, year = {{2025}}, }