A Clustering-Based Machine Learning Approach to Forecast Cryptocurrency Prices

Forecasting cryptocurrency prices is challenging due to their highly vola–tile and unpredictable nature. This study presents a high-frequency forecasting model that combines changepoint analysis with clustering-based machine lear–ning techniques. The pruned exact linear time (PELT) method was applied to intraday historical data from Bitcoin (BTC), Ethereum (ETH), and Cardano (ADA) for a one-year period, from August 2022 to August 2023, to detect volatility shifts, dividing the data into distinct regions. These regions were then grouped using the K-means algorithm based on characteristics like price and volume variance. Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and Random Forest (RF) models were trained on both clustered and non-clustered data to predict cryptocurrency closing prices. The results showed that using clustered original data improved forecasting accuracy in most cases. For ADA, SVR’s Mean Absolute Error (MAE) dropped from 0.0018 to 0.0003, and for RF, MAE improved from 0.0010 to 0.0004. Similar improvements were observed for BTC and ETH. These results show that clustering volatility regions enables machine learning models to more accurately capture price dynamics, resulting in better and more reliable forecasts. This study demonstrates the effectiveness of integrating changepointbased clustering with machine learning models to improve short-term cryptocurrency price forecasts.