Causality Is Optional: Rethinking Portfolio Efficiency Through Predictive Lenses

TL;DR for operators

A portfolio does not care whether your signal has a beautiful causal origin story. It cares whether the signal points in roughly the right direction, ranks assets usefully, and is scaled well enough not to produce absurd weights.

That is the useful, slightly impolite message of Alejandro Rodriguez Dominguez’s paper, Is Causality Necessary for Efficient Portfolios?¹ The paper challenges a strong claim in recent causal factor-investing work: that causal factor models are necessary for investment efficiency. Its answer is narrower and more operational. Within static mean-variance and related quadratic optimisation frameworks, causal identification is not the necessary condition. The necessary operating conditions are geometric.

The three checks are:

1. Directional alignment: does the predictive signal have positive alignment with the true expected-return direction?
2. Ranking preservation: does it preserve the relative ordering of assets well enough to stay inside the useful efficient set?
3. Calibration: are magnitudes scaled well enough to size positions efficiently?

The paper does not say causality is useless. That would be a much louder, dumber article. It says causality is not automatically required at the optimisation stage. Causal models remain valuable for attribution, scenario design, regime diagnostics, and counterfactual stress testing. But once a predictive signal is passed into a static quadratic optimiser, the optimiser sees geometry, not metaphysics.

For asset managers, robo-advisers, and quant platforms, the practical implication is clear: before insisting that every return signal must be causally identified, audit the signal’s out-of-sample alignment, ranking stability, calibration, covariance sensitivity, and constraint behaviour. Causal modelling may improve these properties. It is not identical to them.

The optimiser sees geometry, not intentions

Portfolio teams often talk about models as if the optimiser has moral preferences. Causal signals are clean. Associational signals are suspect. Machine-learning signals are powerful but possibly untrustworthy. Factor signals are acceptable if everyone has been saying the factor’s name for at least twenty years.

The optimiser is less sentimental.

In a mean-variance setting, the optimiser receives a return vector, a covariance matrix, and constraints. It does not know whether a signal came from a causal graph, a regression, a neural network, a ranking model, or an analyst who had too much coffee. It only receives an implied risk-return geometry.

This is why the paper’s mechanism-first framing matters. It does not merely say “predictive models can work.” That would be harmless but vague. It decomposes what “can work” means.

A predictive signal can be wrong in several different ways:

This separation is the paper’s main contribution. It replaces a binary question — “is the model causal?” — with an operational diagnostic stack: “does the signal preserve the geometry needed by the optimiser?”

That is a better question because it can actually be tested.

Misspecification usually bruises the signal before it kills it

The paper begins by engaging with a strong causal-necessity claim: omitted variables and structural misspecification can distort factor exposures, causing inefficient portfolios. This part is not controversial. Omitted-variable bias is real. Misspecified factors can create wrong exposures. Anyone who has watched an optimiser turn a small expected-return error into a heroic leveraged bet already knows that expected returns are dangerous little objects.

The dispute is about necessity and failure mode.

The causal-necessity argument often implies that structural misspecification generically causes signal inversion and therefore frontier collapse. Dominguez’s counterpoint is sharper: misspecification can induce bias without generic inversion. In many settings, omitted-variable bias attenuates the signal. It shrinks the magnitude. It does not necessarily flip the sign.

That distinction matters because attenuation and inversion lead to different business decisions.

If misspecification usually causes inversion, then non-causal signals are structurally unsafe. The only responsible gate is causal identification. If misspecification often causes attenuation, then the right gate is empirical: measure whether the signal remains directionally aligned, whether the ranking survives, and whether calibration is good enough for the intended risk budget.

The paper’s structural cancellation example demonstrates this directly. When omitted confounding terms partially cancel, the estimated coefficient can remain sign-correct but smaller. The practical translation is simple: a biased predictor may still point in the right direction while overstating or understating the intensity of the bet.

That is not a free lunch. It is also not a corpse.

Direction is necessary, ranking is membership, calibration is money

The most useful part of the paper is its decomposition of portfolio efficiency into three layers.

First, directional alignment is the minimum condition. If the surrogate return vector is positively aligned with the true expected-return vector, the optimiser can still produce a coherent frontier. If the signal is orthogonal, the realised return disappears. If it is anti-aligned, the model has created a wonderfully efficient way to lose money. Finance has many of those already.

Second, ranking preservation determines whether the signal keeps assets in the right relative order. The paper connects this to the “portfolios from sorts” intuition: if the signal ranks assets correctly, it can preserve membership in the efficient set. A ranking model can therefore be useful even without exact causal interpretation.

Third, calibration determines how much efficiency is actually captured. This is the part many dashboards hide behind a pleasing backtest. A model can rank assets correctly and still size them badly. If magnitudes are distorted, portfolio weights are distorted. Sharpe ratios are attenuated even when signs and ranks look respectable.

In the paper’s formulation, realised Sharpe scales with cosine alignment between the true signal and the surrogate signal. That is an elegant result because it turns “model misspecification” from a fog machine into a measurable geometry problem.

The operational version:

It is closer to:

The approximation sign is doing work. The paper is not claiming these checks solve every market problem. It is saying these are the relevant checks for static mean-variance and closely related quadratic optimisation, conditional on the chosen risk model and signal representation.

The evidence is layered, not decorative

The empirical and simulation sections are best read as a set of targeted tests, not as one giant proof-by-chart. Each part answers a different question.

This is an important distinction. The S&P 500 example is deliberately modest: five stocks over 1000 trading days. It is an empirical illustration, not a universal market theorem. The broader empirical muscle sits in the appendix: a global bond universe of 1,350 instruments over 1,061 trading days, with rolling 252-day estimation windows and one-step-ahead out-of-sample evaluation.

That appendix is doing robustness work. It tests whether the geometry survives under higher dimensionality, covariance estimation problems, constraints, nonstationary volatility regimes, and bootstrap validation. It is not changing the claim. It is stress-testing the claim.

The bond appendix says estimation discipline matters

The rolling bond results are the most practically relevant part for operators because they look less like a classroom example and more like a portfolio process.

The paper compares sample covariance versus Ledoit-Wolf shrinkage, and unconstrained versus long-only constrained portfolios. The best-performing specification is shrinkage covariance with constraints: mean rolling Sharpe of 1.48, with statistically significant Sharpe differences in 51.7% of rolling windows. Sample covariance without constraints performs badly: mean rolling Sharpe of -0.65, with significance in only 23.9% of windows.

That is not a minor footnote. It is the paper quietly admitting something every allocator should appreciate: signal geometry is not enough if the risk model is unstable.

The regime split makes the point more sharply. In high-volatility regimes, sample covariance specifications lose almost all statistical significance. The unconstrained sample covariance version shows significance in 1.1% of high-volatility windows; the constrained sample covariance version shows 0.7%. By contrast, shrinkage plus constraints preserves significance in 41.4% of high-volatility windows and 62.0% of low-volatility windows.

The business lesson is not “use non-causal models and relax.” It is “separate signal validity from estimation hygiene.” The paper’s geometry says causal identification is not necessary for optimiser viability. The bond evidence says viability still depends heavily on covariance regularisation, constraints, and rolling validation.

Translation: the optimiser does not need a causal sermon, but it does need adult supervision.

What this changes for quant model governance

The paper’s practical value is less about declaring a winner in the causality-versus-prediction debate and more about improving model governance.

A useful governance checklist would separate four questions that are often blurred together:

This is a healthier process than forcing every alpha model through a causal-identification gate and pretending that the resulting graph has domesticated the market. In adaptive markets, causal structure can be unstable, latent, reflexive, and regime-dependent. Causal analysis can still be valuable, but it should not be confused with a guarantee of portfolio efficiency.

For a robo-adviser, this means predictive signals should be judged by portfolio-level consequences, not by standalone model elegance. For an asset manager, it means causal research can inform signal design, but the approval committee should still ask whether the signal’s ranking and calibration survive out of sample. For a quant platform, it suggests tooling opportunities: alignment dashboards, calibration stress tests, shrinkage comparisons, constraint sensitivity, and rolling bootstrap diagnostics.

The revenue implication is not glamorous, which is usually a good sign. The value is cheaper diagnosis. Instead of debating whether a signal is philosophically pure, teams can measure whether it remains geometrically usable under the optimiser that will actually trade it.

What the paper directly shows, and what Cognaptus infers

It is worth separating the paper’s direct claims from the business interpretation.

This distinction matters because a sloppy reading would turn the paper into a slogan: “causality is optional.” The more precise reading is better: causality is optional for a specific optimisation necessity claim, not optional for understanding markets.

Where causality still earns its keep

The paper is careful on this point, and the article should be too. Causality remains useful when the business question is not simply “what portfolio weights maximise this quadratic objective?”

Causal models help when a team needs to know why a signal works, whether it is likely to survive a policy shock, how exposures respond to macro drivers, or what happens under a counterfactual regime. They are also useful for stress testing, attribution, and communicating investment logic to stakeholders who do not enjoy being told that the model works because a high-dimensional vector points in a nice direction.

The boundary is therefore not:

The useful middle ground is intellectually less dramatic and operationally more valuable. Causal analysis belongs upstream and around the optimiser: designing better signals, interpreting exposures, testing regimes, and building scenarios. The optimiser itself still needs geometry.

The limitation is the framework, not a disclaimer pasted at the end

The paper’s results live inside static mean-variance and related quadratic optimisation. That boundary should not be waved away.

Many investment processes are not clean static quadratic problems. They include turnover penalties, liquidity constraints, transaction costs, tax effects, drawdown objectives, options convexity, path dependence, capacity limits, and client-specific constraints. In those settings, alignment, ranking, and calibration still matter, but the sufficiency conditions may no longer transfer directly.

There is also a measurement problem. The true expected-return vector is not observable. Out-of-sample realised returns are noisy. Alignment diagnostics are themselves estimates. Ranking stability can look persuasive until a regime changes. Calibration can be overfit with impressive precision and no economic content. Yes, the optimiser is geometric. No, the geometry is not handed down from Mount Markowitz on stone tablets.

The global bond results reinforce this limitation rather than weakening it. Performance improves materially when covariance shrinkage and long-only constraints are used. That means production success depends not only on signal direction but also on risk estimation and regularisation. A beautifully aligned signal paired with a fragile covariance matrix can still produce nonsense.

Finally, the paper does not prove that causal modelling is commercially unnecessary. In many institutions, causal reasoning is part of model explainability, risk governance, regulatory comfort, and crisis response. Those are not side quests. They are part of running money without surprising the board in ways that require lawyers.

The operator’s takeaway: audit geometry before demanding causality

The paper’s strongest contribution is a change in diagnostic order.

Do not start by asking whether a signal is causal. Start by asking whether it is geometrically usable:

1. Does it preserve positive out-of-sample alignment?
2. Does it rank assets reliably across rolling windows?
3. Does calibration survive rescaling tests?
4. Does the efficient frontier degrade smoothly under noise?
5. Does covariance shrinkage improve stability?
6. Do long-only, leverage, turnover, and exposure constraints regularise or destroy the result?
7. Does performance remain meaningful in high-volatility regimes?

If the answer is no, causal language will not rescue the portfolio. If the answer is yes, lack of causal identification is not automatically disqualifying within the paper’s optimisation setting.

That is the business relevance of this work. It gives investment teams a practical replacement for a blunt causal purity test: a geometry audit. Less romantic, more useful. Markets will survive the disappointment.

Causality is still valuable. It is just not the bouncer standing at the door of every efficient frontier.

Cognaptus: Automate the Present, Incubate the Future.