Averaged bars for cryptocurrency price forecasting across different horizons

Abstract

Technical analysis uses past price movements and patterns to predict future trends and help traders make informed decisions about their cryptocurrency portfolios. This study investigates the effectiveness of different forecasting algorithms and features in predicting the future log return of cryptocurrency close price across various horizons. Specifically, we compare the performance of AdaBoost, light gradient boosting machine (LightGBM), random forest (RF), and k-nearest neighbor (KNN) regressors using Kline open, high, low, close (OHLC) prices data and averaged bars (Heikin-Ashi) features. Our analysis covers ten of the most capitalized cryptocurrencies: Cardano, Avalanche, Binance Coin, Bitcoin, Dogecoin, Polkadot, Ethereum, Solana, Tron, and Ripple. We have observed nuanced patterns in predictive performance across different cryptocurrencies, forecasting horizons and features. Then we have found that AdaBoost and RF models consistently exhibit a competitive performance, with LightGBM showing promising results for specific cryptocurrencies. The impact of forecast horizons on forecasting performance underscores the need for tailored forecasting models. In summary, the use of Kline OHLC data as features outperforms averaged bars in forecasting the first and second horizons, while averaged bars outperform Kline OHLC data for mid- to relatively long-term horizons (starting from the third horizon). Our findings suggest that averaged bars merit more attention from researchers instead of relying solely on Kline OHLC data.