View Fusion Vis-à-Vis a Bayesian Interpretation of Black–Litterman for Portfolio Allocation

Trent Spears, Stefan Zohren, Stephen Roberts

This paper, published in early 2023, provides a rigorous bridge between classical portfolio theory and modern machine learning by introducing "information fusion" to the Black-Litterman framework.

TL;DR

The authors extend the famous Black-Litterman portfolio model to handle multiple, potentially conflicting predictions (views) from different sources. By applying "information fusion" techniques from the world of robotics and sensors to financial factor models, they demonstrate a "model diversification" benefit that leads to more robust and higher-performing investment strategies.

What Problem Does This Paper Address?

In quantitative finance, the Markowitz Modern Portfolio Theory is the gold standard for allocation, but it is notoriously sensitive to small errors in expected returns. The Black-Litterman (BL) model improved this by allowing investors to inject their own "views" (predictions) into the math.

However, a modern portfolio manager doesn't just have one view; they might have three different machine learning models (e.g., a Transformer, a Gradient Boosted Tree, and an ARIMA model) all predicting returns for the same assets. Standard BL doesn't tell you how to combine these. Simply averaging them ignores the fact that these models might be highly correlated or have very different levels of uncertainty. This paper tackles the challenge of "view fusion"—how to mathematically merge multiple predictive sources into a single, consistent input for portfolio optimization.

Methods

The authors frame the problem using a Bayesian hierarchical model based on Arbitrage Pricing Theory (APT). Instead of predicting individual stock returns, they predict "factor" returns (like Value, Momentum, or Size).

The core novelty lies in their use of four specific Information Fusion algorithms to combine views:

  1. Precision-Weighted (PW) Fusion: The simplest method, which assumes the different models are completely independent (zero correlation).
  2. Covariance Intersection (CI): A conservative method that works even when the correlation between models is unknown. It ensures the resulting uncertainty estimate isn't overconfident.
  3. Inverse Covariance Intersection (ICI): A more modern, less conservative version of CI that accounts for shared information between models.
  4. Covariance Union (CU): A method designed for when models disagree wildly, ensuring the fused estimate is consistent with at least one of the sources.

They also incorporate a realistic quadratic transaction cost model to ensure their results aren't just theoretical "paper profits" that would vanish in real-world trading.

Datasets

The authors used a massive dataset to test their theories:

  • Source: CRSP and IBES databases (via Wharton Research Data Service).
  • Universe: The top 2000 US stocks by market capitalization.
  • Timeframe: 29 years of daily data (1992–2021), notably including the 2020 COVID-19 market crash and recovery.
  • Factors: They constructed 5 primary risk factors (market beta, size, volatility, momentum, and value) and 70 industry-specific factors.

Key Insights and Ideas

  • Model Diversification is a "Free Lunch": Just as diversifying across assets reduces risk, "fusing" different predictive models provides a performance boost. The authors call this "model diversification."
  • Uncertainty as a First-Class Citizen: Most financial models treat uncertainty as a "tweakable" parameter. This paper treats it as a direct output of the ML models, distinguishing between aleatoric (random noise in the data) and epistemic (uncertainty in the model itself) components.
  • ICI is the Sweet Spot: Among the fusion methods, Inverse Covariance Intersection (ICI) consistently provided the best balance—it wasn't as naive as PW fusion nor as overly cautious as CI.

Key Results

The authors compared their fusion methods against three standalone "view" models (ARIMA, Boosted Regression, and Gaussian Processes) and the S&P 500 Total Return index.

  • Superiority of Fusion: The fused models generally outperformed the individual models.
  • Sharpe Ratio Boost: Over the 29-year period, the ICI fusion method achieved a median annual Sharpe ratio of 0.48, significantly higher than the Precision-Weighted (0.11) or the individual ARIMA/GP models.
  • Robustness: The fusion methods remained profitable even after accounting for transaction costs, proving that the "model diversification" benefit is large enough to survive real-world frictions.

Limitations and Future Work

  • Static Priors: The authors re-initialize their "prior" beliefs at every time step. They suggest that a "dynamic" state-space model that carries information forward through time could be even more powerful.
  • Drawdown Control: While the model optimizes for risk-adjusted returns (Sharpe ratio), it doesn't explicitly include mechanisms to prevent large peak-to-trough drawdowns, which is a major concern for real-world fund managers.
  • Simple Input Models: The "view models" used (ARIMA, etc.) were relatively basic. The authors note that using more sophisticated, state-of-the-art deep learning models as inputs to the fusion process would likely yield even better results.

Summary updated

Abstract

The Black–Litterman model extends the framework of the Markowitz modern portfolio theory to incorporate investor views. The authors consider a case in which multiple view estimates, including uncertainties, are given for the same underlying subset of assets at a point in time. This motivates their consideration of data fusion techniques for combining information from multiple sources. In particular, they consider consistency-based methods that yield fused view and uncertainty pairs; such methods are not common to the quantitative finance literature. They show a relevant, modern case of incorporating machine learning model-derived view and uncertainty estimates, and the impact on portfolio allocation, with an example subsuming arbitrage pricing theory. Hence, they show the value of the Black–Litterman model in combination with information fusion and artificial intelligence–grounded prediction methods.
BibTeX
@article{spears2023view,
  author = {Trent Spears and Stefan Zohren and Stephen Roberts},
  title = {{View Fusion Vis-à-Vis a Bayesian Interpretation of Black–Litterman for Portfolio Allocation}},
  journal = {The Journal of Financial Data Science},
  year = {2023},
  doi = {10.3905/jfds.2023.1.132},
}

Loading related content...

Figure