<![CDATA[Predicting non-stationary, non-normal hourly foreign exchange rates using probability-based classifiers is challenging, and the standard Gaussian assumption for Naive Bayes can be fragile. This study evaluated and compared three Naive Bayes representations for hourly foreign exchange rate prediction to examine the impact of feature discretization. The representations included Gaussian Naive Bayes on continuous indicators, equal-width binned Bernoulli Naive Bayes, and equal-frequency binned Bernoulli Naive Bayes. Relative Strength Index, Average True Range, and Moving Average Convergence Divergence indicators were computed from hourly historical bars for three major currency pairs. The models were trained in Python, exported via the Open Neural Network Exchange, and integrated into MetaTrader 5 for backtesting under a standardized execution gate. Backtests revealed that the quantile-binned equal-frequency model achieved consistent profitability across all three currency pairs, whereas the Gaussian and uniform-binned models demonstrated unstable performance and significant drawdowns. The findings suggested that quantile-based discretization mitigated the impact of outliers and improved classifier robustness in non-stationary market environments.]]>
Copyrights © 2026
