This is the graded finale for Causal Inference for Alpha & Execution, and it runs the whole discipline back at you. You started by separating the two sentences that sound identical and cost fortunes apart: the signal predicts returns (a backtest’s association) versus the signal causes returns (the claim every dollar of capital actually makes) — and you met the confounder zoo (regime, liquidity, crowding, selection) that fakes an edge no train/test split can catch. From there you learned the grammar: potential outcomes Y(1) and Y(0) and the fundamental problem that you only ever see one of them; ATE versus ATT and how the naive difference-in-means equals the effect plus a selection-bias term; causal DAGs with their back-door paths, and the make-or-break contrast between a confounder (a common cause you must control) and a collider (a common effect that controlling for creates spurious association). You learned why “control for everything” is a trap — collider bias, bad controls, post-treatment mediators — and re-read the factor zoo as confounding plus multiple testing. Then the estimation ladder: natural experiments and difference-in-differences with parallel trends; instrumental variables (relevance plus the untestable exclusion restriction) and regression discontinuity around a sharp cutoff; double machine learning with Neyman orthogonality and cross-fitting to partial out high-dimensional confounders without poisoning the causal estimate. And finally the flagship: market impact as causal TCA, where your own order is the treatment, naive TCA is confounded because you trade hardest when the market is already moving, and the honest limit is that a true RCT is a luxury markets almost never grant. No hints are shown, each answer locks the moment you submit, and your score stays hidden until the very end.
Course Recap
Big picture
Causal Inference for Alpha & Execution — the whole arc
- Causal Inference for Alpha & Execution
- 1 · Correlation is not alpha
- Backtest = observational study (association)
- Allocation = causal bet (intervention)
- Confounder zoo: regime, liquidity, crowding, selection
- 2 · Potential outcomes & DAGs
- Y(1), Y(0); see only one — the fundamental problem
- Naive diff-in-means = ATT + selection bias
- Confounder (common cause) vs collider (common effect)
- 3 · Confounding & the control trap
- Back-door criterion: block back doors, open none
- Collider bias + bad controls (mediators, post-treatment)
- Factor zoo = confounding + multiple testing
- 4 · Natural experiments & DiD
- As-if-random: reconstitutions, regulatory shocks
- Event study: abnormal returns and CAR
- DiD = (treated change) minus (control change); parallel trends
- 5 · IV & RDD
- Instrument: relevance (testable) + exclusion (not)
- 2SLS / Wald = cov(Z,Y) / cov(Z,X); weak-instrument risk
- RDD: jump at a sharp cutoff; local validity
- 6 · Double machine learning
- Partial out confounders: residualize Y and T
- Neyman orthogonality + cross-fitting
- Controls only OBSERVED confounders; needs overlap
- 7 · Market impact as causal TCA
- Your trade is the treatment; impact is its effect
- Naive TCA confounded by why you traded
- Square-root law; ties to optimal execution; RCT is rare
- 1 · Correlation is not alpha
One run, one shot
This is a graded, irreversible exam. There are 24 questions, shown one at a time. The instant you submit a question it locks — there is no Back button, no retry, and no Restart. A wrong answer simply fails that question and the exam moves on; you cannot revisit it. Your running score is hidden until the final screen. The pass mark is 70%. Some questions accept more than one correct option — read every option before you commit, because once you submit you own the answer.
A single signal that was never data-mined posts a deflated, fully out-of-sample information coefficient of 0.06 over fifteen years, then loses money the moment it is deployed. Every overfitting audit was clean. What is the most likely explanation?
Select an answer to continue.
Where this leaves you on the quant ladder
Pass or fail, you now carry the rarest discipline in quantitative finance: the habit of reading a backtest as an observational study and asking, before a dollar moves, what is the treatment, what is the counterfactual, and which back doors are open? You can draw the DAG, tell a confounder from a collider, refuse the “control for everything” reflex, and reach for the right tool — a reconstitution or a difference-in-differences, an instrument or a discontinuity, double machine learning to partial out the factor zoo — to turn a correlation into a defensible effect. Most valuable of all, you can point that lens at the one trade you unambiguously cause, your own order, and stop letting naive TCA confound your execution. The honest close is the one the whole course earned: markets rarely grant the randomised trial, so the goal was never proof — it was a well-reasoned degree of belief, triangulated across imperfect designs, held humbly and sized accordingly. That judgment is what separates a quant who harvests structural edges from one who rents confounders and calls them alpha.