Machine Learning Analysis of Japanese Candlestick Patterns in Financial Data: An Empirical Application Based on 10 Years of Daily Data
Year: 2026
Type: conference-paper
Abstract: This study examines whether Japanese candlestick patterns provide predictive power in financial time series, either alone or combined with technical indicators, using daily data from eight liquid assets over a ten-year period. Rule-based pattern detection was integrated with returns, volatility, momentum, moving averages, ATR, RSI, MACD, and volume. Logistic regression and random forest models were evaluated using a chronological train-validation-test split. On 20,147 observations, random forest achieved ROC-AUC=0.544 in validation but dropped to 0.486 in testing, while logistic regression reached 0.518. A simple strategy yielded −16.2% return, whereas buy-and-hold achieved 40.0%. Results indicate that candlestick patterns do not provide stable alpha at daily frequency but remain useful as auxiliary features capturing behavioral patterns.
Author: Bora Canbula
Institution: Manisa Celal Bayar University
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7202191925