| This study aims to predict cryptocurrency prices using machine learning algorithms and to determine the most successful method through multi-criteria decision-making (MCDM) techniques. A multidimensional dataset was constructed for Bitcoin, Ethereum, BNB, Ripple, and Dogecoin using daily data from January 1, 2018, to December 31, 2023, incorporating price movements, technical indicators, investor sentiment, and macroeconomic factors. Prediction models such as SVR, RF, XGBoost, and LSTM were applied, and their performances were evaluated using error metrics like R², MAE, MSE, and RMSE. The importance levels of variables were analyzed through permutation importance for SVR and LSTM, and embedded importance calculation methods for RF and XGBoost. Based on model performances, a decision matrix was created, criterion weights were calculated using the CRITIC method, and rankings were conducted using TOPSIS, ARAS, and CODAS methods. The most successful algorithm was determined using the Copeland method. According to the results, the XGBoost algorithm demonstrated the highest overall performance. The LSTM algorithm ranked second, followed by RF in third and SVR in fourth place. Additionally, the findings indicate that technical analysis variables play a decisive role in model performance, whereas macroeconomic and sentiment indicators provide limited contribution. |
| Keywords: |
Cryptocurrency, Price Prediction, Machine Learning, Extreme Gradient Boosting, Support Vector Machine, Random Forest, Long Short-Term Memory, MCDM |
|
