Correlation Is Not Alpha

You have spent entire courses learning to predict returns honestly: deflate the Sharpe, purge the cross-validation, distrust the factor zoo. And it was all necessary. But here is an uncomfortable truth that none of those tools address: a perfectly honest, fully out-of-sample, multiply-deflated prediction is still not a reason to allocate capital. Prediction tells you the signal and the return move together. Allocation bets that turning the signal on will make the return appear. Those are different claims, and the second one — the causal one — is the only one your P&L actually cares about.

This is the leap that quietly demolishes more strategies than overfitting ever did, because it survives every overfitting test. Your edge can be real, stable, and out-of-sample, and still evaporate the instant you trade it — not because you data-mined it, but because a third thing was driving both your signal and the return, and that third thing just changed.

This lesson installs the single mental upgrade the whole course is built on: learning to see a backtest as an observational study, not an experiment.

A backtest measures association, not effect

Analogy. A backtest is a security camera that recorded, for 15 years, that every time the ice-cream van parked outside, more people drowned at the beach that day. The footage is real, out-of-sample, and replicable. Banning ice-cream vans will not save a single swimmer, because a third thing — hot weather — drives both ice-cream sales and swimming. Your backtest is that camera. It faithfully records co-movement; it cannot, on its own, tell you which arrow points where.

Definition. A causal effect of a signal $X$ on a return $Y$ is the difference between the return you would get if you acted on $X$ and the return you would get if you did not, holding everything else fixed. A backtest never observes both worlds for the same period — it only observes the world that happened. What it estimates is the association $\text{Corr}(X, Y)$, which equals the causal effect only when no third variable drives both. That “only when” is the whole subject.

Worked example. Suppose a regime variable $Z$ (say, “risk-on vs risk-off”) drives both your signal and next-month returns. Within any single regime the signal does nothing — its slope on returns is flat. But the risk-on regime has both a higher average signal and a higher average return, and the risk-off regime has both lower. Pool the two regimes together and the scatter tilts upward:

The pooled slope of $+0.8$ is your backtest “edge.” It is composed entirely of the between-regime difference — the confounder — and none of it is a causal effect of the signal. Trade it, and the day the regime mix shifts, your edge inverts.

When to use it

Apply the “observational study” lens before every capital-allocation decision, not just exotic ones. The question is always the same: if I knew nothing changed except that I switched this signal on, would the return still appear? If the honest answer is “only if regime / liquidity / crowding stays as it was in-sample,” you have a correlation, not an effect — size accordingly, or go find the confounder.

Allocating capital is a causal bet

Analogy. A weather forecaster predicts rain; a rain dancer claims to cause it. Both might have an impressive track record of “danced, then it rained.” Only one of them is making a claim that survives being tested by intervention — actually dancing on a dry day. When you allocate capital you stop being the forecaster and become the rain dancer: you are no longer reading the co-movement, you are acting and demanding the consequence.

Definition. An intervention (Pearl’s $do(\cdot)$ operator, which the next lesson formalises) is a deliberate setting of the treatment, breaking whatever process usually determines it. The quantity you actually care about is $\mathbb{E}[Y \mid do(X{=}1)] - \mathbb{E}[Y \mid do(X{=}0)]$ — the return when you force the signal on versus off. A backtest gives you $\mathbb{E}[Y \mid X{=}1] - \mathbb{E}[Y \mid X{=}0]$ — the return when the signal happened to be on versus off. The $do$ and the plain conditional are equal only under causal assumptions you must argue for explicitly.

Worked example. Two desks each report a signal correlated with next-day returns at the same strength.

Identical backtests, opposite fates. The only thing that separates them is the mechanism, and the backtest is blind to mechanism. You cannot tell A from B by staring at the equity curve harder; you can only tell them apart by reasoning about what causes what — which is the causal-inference toolkit.

When to use it

Frame every strategy as an intervention with a named counterfactual at the research-proposal stage, before a single backtest is run. Writing down “treatment = X, counterfactual = not-X, claimed mechanism = …” forces the causal question into the open while it is still cheap to answer, instead of discovering after deployment that you were a rain dancer all along.

The confounder zoo: what fakes an edge

Analogy. Confounders are the stagehands of a backtest — invisible from the audience, but moving all the scenery. The play looks like the lead actor (your signal) is driving the plot; really, four stagehands behind the curtain are pushing both the actor and the props into place. Learn their names and you stop being fooled by the performance.

Definition. A confounder is a variable that causally influences both the treatment (your signal) and the outcome (the return), creating a non-causal “back-door” path between them. Because the back-door path carries association, the observed correlation overstates (or invents) the causal effect. The four that haunt trading research:

Worked example — crowding. A value signal earns 4% per year in backtest over 2002–2014. Decompose it honestly: 1% was a genuine risk premium, and 3% was the price impact of cumulative inflows as the value factor became fashionable — a confounder (popularity) driving both the factor’s measured returns and its apparent strength. Out-of-sample, with inflows exhausted and crowding now causing outflows, the 3% turns to $-2\%$. The realised edge collapses from 4% to roughly $1\% - 2\% = -1\%$. Nothing was overfit; a confounder simply ran out of fuel and reversed.

When to use it

Run the confounder zoo as a checklist on every candidate signal: could regime, liquidity, crowding, or selection alone produce this correlation? For each “plausibly yes,” you owe either a control (later lessons), a robustness test (does the edge survive within each regime, in liquid names only, before the signal was popular?), or an honest haircut to your expected return. Treat an unexplained edge as a confounder until proven otherwise.

Prediction can still be useful — just don’t confuse it for an effect

Analogy. A confounded edge is a rental, not a deed. You can absolutely live in a rented house and profit from it — but you’d be a fool to renovate it as if you owned it, because the landlord (the confounder) can end the lease without notice. The catastrophe is treating the rental as a deed: levering up, concentrating, and assuming permanence in something you never structurally owned.

Definition. The practical taxonomy is three-way, not two-way:

Worked example. Two managers both run the same momentum signal. Manager A believes it is structural and runs it at 3x leverage with no regime overlay. Manager B knows it is a confounded-but-stable edge (it works in trending regimes, dies in choppy ones) and runs it at 1x with a regime filter that cuts exposure when realised correlation spikes. When the regime turns, A is down 40% and B is down 6%. Same signal, same correlation, same backtest — opposite causal beliefs, and the causal belief is what sized the position and the overlay. The money was made and lost in the causal model, not the predictive one.

When to use it

Classify every live edge into the three buckets and re-classify it on a schedule. A “confounded but stable” edge can migrate to “confounded and fragile” silently as crowding builds or a regime ages — so the classification is a monitoring task, not a one-time label. The position size and hedge should follow the bucket, and the bucket should follow the evidence.

Recap

You came in able to predict returns honestly and left knowing that honest prediction is not enough. A backtest is a security camera, not an experiment: it records that signal and return co-moved, never that the signal caused the return. Allocating capital is a rain dance — an intervention demanding a consequence — and the consequence only arrives if no stagehand (regime, liquidity, crowding, selection) was secretly moving both. Out-of-sample testing and Sharpe deflation, your old weapons, fight overfitting and are blind to all of it. The fix is not a better split but a different question: what is the treatment, what is the counterfactual, and is a back door open? You can still trade a confounded edge — as a rental, sized for the landlord’s whims — but never renovate it as a deed.