Black-Litterman Model

Equilibrium Returns (Prior) in the original model

In the model as Black and Litterman published it, the equilibrium expected returns are reverse-engineered from market capitalization weights using the CAPM. Given the risk aversion parameter , the covariance matrix , and market cap weights :

These implied returns represent the market's consensus view. If an investor has no private information, the optimal portfolio is simply the market portfolio.

What this platform uses as the prior

Our implementation does not perform that reverse optimization. It sets the prior mean to the historical mean return of each asset, annualized as a compound growth rate, the same vector the Mean-Variance method estimates, and passes it straight into the posterior update:

The only quantity taken from the market is the scalar , estimated from the benchmark price series and clipped to . When no benchmark is supplied, or the benchmark has fewer than 60 observations, defaults to 1.0. The prior covariance is the same shrunk covariance matrix the other frontier methods use.

This matters for how you read the result. A sample mean historical return carries the estimation error that equilibrium anchoring was designed to remove, so this method does not inherit that particular robustness property. What it does provide is the entropy-tilt view construction and the Bayesian blend described below.

Investor Views

Views are expressed as linear constraints on expected returns. An absolute view “Asset A will return 12%” or a relative view “Asset A will outperform Asset B by 2%” is encoded via the pick matrix , view vector , and confidence matrix :

where controls how confident the investor is in each view. Lower diagonal entries in mean higher confidence.

Posterior Distribution

The BL posterior combines the prior with investor views using Bayesian updating. The posterior expected return is:

and the posterior covariance is:

where is a scalar controlling the uncertainty in the prior (typically small, e.g. 0.05). The posterior returns are then used in standard mean-variance optimization to obtain the final portfolio weights.

Measure-Theoretic Foundation

Rather than requiring explicit investor views, our implementation constructs implicit views by tilting the empirical return distribution. Following Meucci's Entropy Pooling framework, we define a new probability measure over historical scenarios by exponentially reweighting the reference (empirical) measure :

where is a score for each scenario and is a tilt intensity parameter calibrated to a target KL divergence budget. The run uses one signal: each asset's exponentially weighted mean daily return (half-life 63 trading days) divided by its volatility over the window, clipped to . The score of a day is that day's asset returns weighted by this signal, so days on which the high-signal assets did well receive more probability.

KL Divergence Budget

The tilt intensity is chosen so that the KL divergence between the tilted and reference measures stays within an information budget:

where is the KL budget. A larger budget allows stronger tilts (more opinionated views) while a smaller budget keeps the posterior closer to the empirical distribution. This provides a principled, information-theoretic control over how aggressively the model departs from market consensus.

The budget is stated at the horizon the returns are quoted at, which is one year: nats per year. The scenarios are daily, and relative entropy adds across independent periods, so each daily scenario is tilted to . Spending the whole budget on every daily scenario and then annualizing the tilted mean would turn a small daily tilt into a view of several annualized Sharpe units, which is a statement no momentum signal supports.

View Extraction

The tilted distribution moments are converted into BL-compatible views. With (identity pick matrix for absolute views on all assets), the view vector is the prior plus the annualized shift the tilt causes in the mean return:

The view is written as a shift from the prior, not as the raw tilted mean, so that both sit on one basis. The prior is a compound annual growth rate and a tilted mean is arithmetic, and the two differ by roughly half the variance. With no tilt at all the view equals the prior exactly, so a view never carries a change of basis in place of information.

Confidence levels for each view are determined using the Idzorek (2007) method, which converts a percentage confidence into the corresponding diagonal entry. Each confidence grows with the achieved KL divergence and with the size of the view shift relative to the asset's volatility, between a floor of 5% and a ceiling of 90%. At the default budget most views sit at the floor, so the posterior stays close to the prior.

End-to-End Flow

The full pipeline proceeds as follows:

1. Build the prior: set to the sample mean historical return, and estimate the scalar from the benchmark price series.
2. Score historical scenarios with the volatility-scaled momentum signal described above.
3. Tilt the empirical measure via exponential reweighting with calibrated to the KL budget.
4. Extract implicit views (tilted mean returns) and Idzorek-style confidences from the tilted distribution.
5. Run BL posterior update to blend the prior with the extracted views.
6. Optimize: solve maximum Sharpe ratio on the posterior returns and the posterior covariance. Per-name weight bounds, asset exclusions, and group sum limits enter this solve as extra rows. There is no minimum-volatility fallback. An infeasible program raises an error and the run fails with the reason.

One fallback does exist, and it sits earlier in the chain. If the posterior update itself fails, the optimizer solves over the prior moments instead of the posterior moments, and records fallback: prior_moments_frontier in the diagnostics. Check that field before you read a result as a view-tilted portfolio.

Diagnostics

The optimizer returns several diagnostics with each result: the achieved KL divergence, Effective Sample Size (ESS) of the tilted measure, per-asset confidence levels, and the top signal-ranked assets. These diagnostics help assess how far the tilted measure departed from the prior and which assets drove the tilt.