The problem with raw return forecasts

Part 4 sidestepped return forecasts entirely, and Part 5 discovered structure without them. But sometimes you genuinely do have a view — you think value will outperform, or that a particular sleeve is mispriced. The naive approach is to plug that view into mean-variance optimization as an expected return. The result is almost always a disaster: the optimizer, being an error-maximizer, piles into whatever asset has the highest forecast, magnifies tiny estimation errors into huge allocation swings, and produces portfolios no sensible manager would hold. Express a mild preference and MVO hands you an all-in wager.

The issue isn't that you have views — it's that dumping point forecasts into the objective is a terrible way to express them. The Bayesian family solves this elegantly. It starts from a sensible prior — what the market already implies — and lets you nudge it with specific, confidence-weighted views. You express conviction only where you actually have it, and the math blends your views with the prior in proportion to how sure you are, so a mild view produces a mild tilt. Output remains reviewed research routed through the gallery's gates.

Black-Litterman: start from the market, nudge with views

The Black-Litterman model is the family's workhorse, and it works in two moves:

1. Reverse-engineer the prior. Rather than asking you to forecast returns, it infers the market-implied equilibrium returns — the expected returns that would make today's market-cap weights optimal. That equilibrium is your prior, and it is a far more stable starting point than any hand-built forecast.
2. Blend in your views. You state views — absolute ("I expect emerging-market equity to return 8%") or relative ("tech will outperform staples by 3%") — each with a confidence. The model produces a posterior that shifts the equilibrium toward your views, more where you're confident and less where you're not, then optimizes against that posterior.

The beauty is restraint: where you have no view, the portfolio stays near the sensible market prior; your conviction moves the weights only as far as your stated confidence justifies. The engine implements the full Black-Litterman machinery — market-implied equilibrium, the tau scaling, the pick matrix and view vector, and a regularized blend that stays numerically stable. The review focus is each view, its confidence, and how far the posterior shifts from the prior.

The factor variant and augmented views

Rather than opining on individual securities, you often have views on factors — value, quality, momentum, duration, a sector. Two extensions deepen Black-Litterman:

• Black-Litterman Factor Model expresses both the prior and your views in factor space rather than asset by asset. If your conviction is really "I want more quality exposure," this is the honest way to say it, and it propagates that tilt coherently across every holding that loads on the factor. (This connects directly to the factor risk models in Part 7.)
• Augmented Black-Litterman folds richer inputs into the view structure — additional signals or model outputs — while keeping the confidence-weighted blending discipline so nothing gets double-counted.

The review question is always whether a tilt is a deliberate, defensible decision or an artifact of noisy estimation.

Entropy pooling: views as constraints on scenarios

Black-Litterman blends views into expected returns under Gaussian assumptions. Entropy Pooling takes a more general route: it reweights an entire set of return scenarios so that they satisfy your views, while staying as close as possible (in an information-theoretic sense, minimum relative entropy) to the original distribution. This is powerful because your views can be about almost anything — not just expected returns, but volatilities, correlations, tail probabilities, or the likelihood of a specific scenario. You impose a view as a constraint on the scenario distribution and entropy pooling finds the minimally distorted distribution that honors it.

This makes entropy pooling the natural bridge to stress testing (Part 10): a "stress view" — suppose inflation spikes — is just a constraint that reweights scenarios toward that world, and you can then optimize against the stressed distribution.

Opinion pooling: combining multiple voices

Real investment processes have several opinions — a strategist, a quant model, a committee. Opinion Pooling combines multiple view distributions into one coherent posterior, weighting each source by reliability. Where opinions conflict, pooling resolves them in a principled way rather than letting the loudest voice or the last edit win. The review focus is the source weights and how conflicting opinions were reconciled.

What to review

The discipline with view-based methods is to keep every view honest:

• Is each view evidence-backed? A confident view with no support is just a high-variance bet wearing a Bayesian costume. Review the rationale per view.
• Are the confidences calibrated? Overstated confidence pulls the posterior too far; the prior/posterior comparison chart shows exactly how far each view moved the allocation.
• Are you double-counting? If the same conviction enters as both a direct view and a factor tilt, it gets counted twice. Augmented BL is powerful but demands this check.

When to reach for views

The Bayesian family is the right home whenever you have genuine, articulable conviction — a house view, a research signal, a committee's positioning — and want to express it without destabilizing the portfolio or abandoning the discipline of a sensible prior. When you have no real views, don't manufacture them; risk parity or hierarchical methods are the better default. Views are a feature to use when earned, not a box to fill.

The takeaway

The Bayesian family lets conviction and humility coexist: Black-Litterman anchors on market-implied returns and moves only as far as your confidence warrants, the factor variant expresses views where they actually live, and entropy and opinion pooling handle richer and multi-source views. Keep every view evidence-backed and calibrated. Next in the series: factor models and covariance estimation — the unglamorous engine room where every optimization's inputs are actually built.