Causal Inference for Alpha & Execution

Every backtest you have ever run measured a correlation: when the signal lit up, returns tended to follow. Every dollar you have ever allocated bet on something stronger and unproven — that the signal causes the return, so that turning it on makes the money appear. The gap between those two sentences is the most expensive gap in quantitative finance, and almost nothing in a standard quant curriculum is built to close it. Machine learning, factor models, deep nets, statistical arbitrage — all of them are world-class prediction machines, and prediction is exactly the thing that confounding can fake. This course is the capstone critique that the predictive courses earn the right to receive: a rigorous, finance-first treatment of why a predictive edge and a causal edge are different objects, and how to tell which one you actually have.

We assume you have already lived through the overfitting discipline of Deep Learning for Market Data and the factor-zoo, spread-trading machinery of Systematic & Statistical Arbitrage. You know that a backtested Sharpe must be deflated, that signals decay, that the factor zoo is mostly data-mined ghosts. Causal inference is the missing layer underneath all of that: the formal language for the question “is this relationship structural, or is a third thing driving both?” It is also the field with the sharpest, most under-taught application in all of trading — market impact and transaction-cost analysis as a causal problem, where your own order is the treatment, the price move is the outcome, and naive TCA is hopelessly confounded by why you chose to trade. The arc:

• Correlation is not alpha — the predict-versus-cause leap stated precisely: why a flawless out-of-sample backtest is still only evidence of association, why capital allocation is a causal claim, and the catalogue of confounders (regime, liquidity, crowding, selection) that fake an edge.
• Potential outcomes & causal DAGs — the Rubin potential-outcomes framework (treated vs untreated worlds, the fundamental problem of causal inference), causal directed acyclic graphs as the grammar of assumptions, and the all-important distinction between a confounder (a common cause you must control) and a collider (a common effect you must not).
• Confounding & the control trap — why “control for everything” is amateur advice: collider bias, bad controls (conditioning on a mediator or a post-treatment variable), the back-door criterion, and a clean re-reading of the factor-zoo overfitting problem as a confounding-and-multiple-comparisons problem.
• Natural experiments & difference-in-differences — markets rarely grant you a randomised trial, but they hand you quasi-experiments constantly: index reconstitutions, regulatory shocks, tick-size changes, index inclusions. Event studies done right, and difference-in-differences with its parallel-trends assumption.
• Instrumental variables & regression discontinuity — when you cannot block the back door, find a lever that moves the treatment but nothing else: the instrumental-variables logic (relevance + exclusion), two-stage estimation, and regression discontinuity around a sharp cutoff (index-membership rank, rating thresholds).
• Double machine learning — how to keep your gradient-boosted and deep models and get an unbiased causal estimate: Neyman orthogonality, cross-fitting, partialling out high-dimensional confounders with ML nuisance models, and why naive “throw the treatment into the feature set” plug-in estimates are biased.
• Market impact as causal TCA — the flagship application: your trade moves the price, so the realised cost is a treatment effect; naive TCA is confounded because you trade more aggressively exactly when the market is already moving; the link to optimal execution and the honest limits — unobserved confounders, untestable assumptions, and why a true RCT is a luxury markets almost never grant.

By the end you will read a backtest the way a causal inferencer reads an observational study: asking what the treatment is, what the counterfactual is, which back doors are open, and whether the “edge” survives once the confounders are blocked. A graded final exam runs the whole discipline back at you at the end, one locked question at a time.