Author(s) :

Brennan Park.

Volume/Issue :

Abstract :

This study examined whether transforming relationships among OHLC prices into relative and normalized numerical features can support short-term ETF return prediction. Candlestick charts display relationships among the open, high, low, and close prices through components such as the body, shadows, closing position, and opening gap. Based on these price relationships, this study defined eight numerical candlestick features and compared their prediction performance with raw OHLC information and eight technical indicators. OHLC refers to open, high, low, and close prices, and technical indicators are commonly used features in stock prediction research. The experiment used eight ETFs: SPY, QQQ, DIA, IWM, XLF, XLK, XLE, and XLV and predicted the cumulative log return over the next five trading days from 2014 to 2025. Ridge regression and random forest regressor were used as prediction models. Principal component analysis was also applied to examine whether reducing redundancy within the features affected prediction performance. The results showed that the numerical candlestick-based representations generally recorded lower observed RMSE and MAE values than the raw OHLC and technical indicators representations. Numerical candlestick with PCA also recorded the highest observed directional accuracy among the five feature representations in both prediction models. These findings suggest that explicitly encoding relative and normalized relationships within existing OHLC information may provide a more suitable input representation for short-term ETF return prediction.