Tree-based classifiers often output overconfident probabilities, especially in noisy financial labels, so accuracy can look acceptable while probability estimates are systematically too extreme. That miscalibration feeds directly into bet sizing and Kelly-style sizing, creating oversized positions, deeper drawdowns, and weaker geometric growth. Calibration is evaluated with reliability diagrams plus Brier score (overall usefulness), ECE (average gap), and MCE (worst-case tail risk), with bootstrap confidence bands to reflect limited independent samples. Two practical calibrators are compared: isotonic regression (non-parametric, rank-preserving, best with enough data) and Platt scaling (sigmoid, steadier with small samples but less flexible). The key engineering constraint is avoiding temporal leakage. Calibration is fit on out-of-fold predictions p... #MQL5 #MT5 #AITrading #Strategy https://t.co/YzqMoyifVC