Automation of Black-Litterman Asset Allocation Model Using Machine Learning for Algorithmic Stock Trading

Abstract

Markowitz's mean-variance portfolio optimization model, while foundational in modern finance, is known for its sensitivity to input parameters and tendency toward estimation error maximization, often resulting in overly concentrated portfolios. The Black-Litterman model presents a refined alternative by leveraging Bayesian inference to incorporate an investor's subjective views with prior return estimates, yielding a posterior distribution that supports balanced, intuitive portfolio allocations. This paper advances the Black-Litterman framework by automating the generation of investor views using time-series forecasting models, including ARIMA, Kalman filter, and Long Short-Term Memory (LSTM) networks, thereby reducing subjectivity and enhancing suitability for algorithmic trading. Through empirical analysis, we demonstrate that daily optimization using these automated views produces better performance than conventional procedures based on mean-variance optimization. The findings suggest that this dynamic approach can extend portfolio optimization beyond asset allocation and serve as the foundation of an effective algorithmic trading strategy.