Machine Learning-Assisted Validation of an Adaptive True Renko Trading Strategy: An Empirical Study on crypto

Abstract

This paper presents an empirical evaluation of a proprietary adaptive True Renko trading strategy in which machine learning is used directly as a signal-validation component. The True Renko engine first constructs directional bricks using an ATR-based brick mechanism and applies a moving-average filter to generate candidate trade signals. The candidate signal is then represented through Renko, trend, momentum, volatility, market-regime and risk/reward features and evaluated by an XGBoost-based supervised validator. The ML model does not generate an independent trade signal and is not a Meta ML layer; it only determines whether a True Renko candidate is accepted for execution. The study uses a seven-year BTCUSDT historical dataset, with four years used for development/training and three years reserved as unseen evaluation data. The empirical design evaluates ATR-based True Renko configurations with moving-average filtering and six brick sizes from 100 to 600. The unseen-test stage is subsequently evaluated with ML confidence thresholds of 0.50, 0.53, 0.55, 0.57, 0.60, 0.65 and 0.70. The supplied registry contains 52 experiment runs: six development runs, six no-ML unseen-test runs and forty ML-assisted unseen-test runs. Across the ML threshold experiments, increasing the confidence threshold consistently reduces the number of accepted trades, while profitability and win percentage do not change monotonically. The results demonstrate that ML validation can materially change trade selection and exposure while preserving the deterministic True Renko signal-generation architecture. The findings support further forward testing of the combined True Renko and direct ML-validation framework, while also highlighting the need for transaction-cost modelling, statistical testing and independent forward validation before live deployment.

Keywords