Regime-Aware Trading Strategies with Machine Learning

Subscribe to newsletter

Regime detection is important in portfolio management and remains an active area of research, particularly in the age of machine learning and AI. Reference [1] proposes a trading strategy based on machine learning, combined with regime detection using a Hidden Markov Model.

Specifically, the machine learning technique used is LightGBM, a gradient boosting framework that employs histogram-based split finding for efficient training on high-dimensional tabular data. The study utilizes 63 features spanning technical, macro, and cross-asset data. The resulting strategy is compared against baseline approaches, including XGBoost classifier, logistic regression, SMA 50/200 crossover, and time-series momentum.

The author pointed out,

This study addresses the question of whether machine learning can generate statistically validated alpha in equity markets while adapting to changing conditions. The main contribution is a regime-aware LightGBM framework that provides three advances over prior work…

2. Feature importance hierarchy. The ablation study reveals that cross-asset features (Bitcoin) contribute more predictive value than traditional technical indicators, while SHAP analysis shows that macroeconomic features (yield curve, gold/equity ratio) dominate over stock-specific patterns for high-beta technology stocks. This challenges the conventional emphasis on technical indicators in equity forecasting.
3. Regime-adaptive decision logic. The model autonomously learns different strategies for different market conditions: mean reversion logic in bear markets (prioritizing distance from 200-day SMA) versus risk appetite monitoring in bull markets (prioritizing market beta and gold/equity flows). This adaptive behavior, revealed through regime-specific SHAP analysis, demonstrates that ML models can internalize economically sound reasoning.

The framework achieves a portfolio Sharpe ratio of 1.18 (95% CI: [0.53, 1.84]) and outperforms four baseline models (XGBoost, Logistic Regression, SMA crossover, momentum) under identical walk-forward evaluation. The consistent ∼17% alpha-positive rate across both NASDAQ-100 and S&P 500 universes suggests the approach generalizes beyond the specific training universe.

In short, the paper finds that cross-asset features, particularly Bitcoin as a leading indicator, provide the strongest predictive value; macroeconomic indicators outperform traditional technical indicators for high-beta stocks; and the model adapts its decision logic across regimes, shifting from mean reversion in bear markets to risk appetite monitoring in bull markets.

It is somewhat surprising to us that Bitcoin emerges as the strongest leading indicator, while it is less surprising that macro indicators outperform technical indicators.

Let us know what you think in the comments below or in the discussion forum.

References

[1] Antonio Pagliaro (2026), Regime-Aware LightGBM for Stock Market Forecasting: A Validated Walk-Forward Framework with Statistical Rigor and Explainable AI Analysis, Electronics, 15(6), 1334.

Further questions

What's your question? Ask it in the discussion forum

Have an answer to the questions below? Post it here or in the forum