Swarm-Calibrated Ordinal Forests for Risk-Aware Forecasting in Non-Stationary Intraday Financial Time Series

Ahmed S. Salama, Faculty of Computers and Information Technology, Innovation University, Cairo, EgyptFollow
Ahmed Nagm, Faculty of Computers and Information Technology, Innovation University, Cairo, EgyptFollow

ORCID

Ahmed S. Salama: https://orcid.org/0000-0002-1066-8261

Ahmed Nagm: https://orcid.org/0000-0003-1801-0381

Keywords

Swarm intelligence, Particle swarm optimization, Ordinal patterns, Random forest, Intraday financial forecasting, Non-stationary time series, Transaction cost-sensitive learning

Article Type

Original Article

Abstract

Intraday non-stationary financial time series are difficult to model because of regime switching, volatility clustering, nonlinear price responses, and rapidly changing liquidity. Although machine-learning models can detect nonlinear dependencies, their direct use for intraday trading decisions can be unstable because of transaction costs, drawdown, and exposure risk. This study develops Swarm-Calibrated Ordinal Forests (SC-OF), a hybrid soft-computing architecture for risk-aware forecasting and execution control in non-stationary intraday markets. Continuous OHLCV streams are transformed into a technical-ordinal feature space in which ordinal symbolic patterns preserve local temporal ordering while reducing sensitivity to amplitude distortions. A validation-weighted ordinal forest combines random-forest and extremely randomized tree learners to produce probabilistic directional signals. Particle swarm optimization then calibrates the probability threshold, trailing-stop boundary, and exposure level using a transaction-cost- and drawdown-aware objective. The framework is evaluated on publicly available 5-minute Binance spot klines for BTCUSDT, ETHUSDT, and BNBUSDT using a purged walk-forward procedure. Comparisons include single-tree and ordinal-forest ablations, technical-feature-only models, logistic regression, histogram gradient boosting, buy-and-hold, transaction-cost sensitivity analysis, and Wilcoxon signed-rank testing. SC-OF outperforms static ordinal decision trees and passive exposure under moderate transaction costs and exceeds the single-tree SC-ODT ablation across all tested assets. Results indicate that its advantage arises less from exceptionally high directional classification accuracy than from combining moderate probabilistic prediction with adaptive risk calibration. Performance deteriorates as transaction costs increase, identifying a practical execution limit for intraday soft-computing strategies. These findings position SC-OF as a computationally efficient and interpretable framework for risk-aware forecasting in non-stationary financial environments.

How to Cite

Salama, Ahmed S. and Nagm, Ahmed (2026) "Swarm-Calibrated Ordinal Forests for Risk-Aware Forecasting in Non-Stationary Intraday Financial Time Series," Sustainable Machine Intelligence Journal: Vol. 14: Iss. 3, Article 4.
DOI: https://doi.org/10.63689/3005-3617.1094