Entropy-Augmented Forecasting and Portfolio Construction at the Industry-Group Level: A Causal Machine-Learning...

Area of Science:

• Quantitative Finance
• Machine Learning
• Information Theory

Background:

• Traditional financial models often struggle with complex market dynamics.
• Machine learning offers advanced pattern recognition capabilities for financial forecasting.
• Information-theoretic measures quantify system complexity and predictability.

Purpose of the Study:

• To investigate the efficacy of information-theoretic complexity measures in enhancing machine learning-based industry-group return forecasting.
• To evaluate the impact of these measures on portfolio construction strategies.
• To assess the economic performance and interpretability of entropy-augmented models.

Main Methods:

• Utilized daily return data for 25 U.S. GICS industry groups over three decades.
• Augmented gradient-boosted decision tree models with Shannon entropy and fuzzy entropy.
• Employed a strictly causal rolling-window design for weekly, monthly, and quarterly estimations.
• Developed maximum-profit and minimum-risk allocation strategies.

Main Results:

• The weekly maximum-profit model with Shannon entropy achieved over 30,000% accumulated return, significantly outperforming baseline and fuzzy entropy models.
• Entropy and fuzzy entropy provided smaller, robust improvements on monthly/quarterly horizons, reducing volatility and enhancing downside protection.
• Industry allocations were stable and interpretable, with profit strategies favoring cyclical/growth industries and risk strategies favoring defensive ones.

Conclusions:

• Entropy-based complexity measures demonstrably improve economic performance and interpretability in financial forecasting.
• These measures yield industry-rotation strategies that are more profitable, stable, and transparent.
• The findings support the integration of information-theoretic concepts into machine learning frameworks for enhanced portfolio management.