Topic 19 of 20 · Part V : Advanced Decision Intelligence

Causal inference

Even when causal certainty is impossible, the framework forces the product to state which assumptions connect the evidence to the conclusion.

The control pattern

Data asset

Observational and experimental data

Knobs

Treatment, Control, Outcome window, Confounders, Matching, Identification assumptions

Outcome

Better distinction between correlation and cause

Measurement

See how to measure this below

What it is

Causal inference asks whether an intervention caused an outcome, not merely whether two variables moved together. The data asset consists of observations or experiments that create a credible comparison.

Knobs include treatment definition, control group, outcome window, confounders, matching method, segmentation, and the assumptions required for identification. In markets, causal claims are difficult because events overlap, expectations are incorporated before announcements, and the same macro forces affect many securities at once.

Even when causal certainty is impossible, the framework is valuable because it forces the product to state which assumptions connect the evidence to the conclusion.

The knobs in detail

Each row is one adjustable property of the data asset, and what moving it tends to do.

KnobWhat you adjustLikely effect
TreatmentWhat counts as the interventionDefines the question being asked
ControlWhat the treated group is compared againstDetermines credibility of the comparison
Outcome windowOver what period effects are measuredChanges the answer more than most expect
ConfoundersWhat else could explain the resultThe main threat to any market claim
MatchingHow comparable units are pairedBalances treatment and comparison groups
Identification assumptionsWhat must hold for the estimate to mean anythingShould be stated, not implied

Applied: Stocks Assistant

Stocks Assistant can use causal methods carefully for research questions such as whether a guidance cut changes abnormal returns, whether a dividend initiation affects volatility, or whether a product feature improves user decision quality. An event study might compare returns around an announcement with a market or sector benchmark, but it must define the event time, estimation window, comparison model and contamination rules. Language models can help extract treatments and outcomes from documents, yet the statistical design should remain explicit and reproducible.

How to measure it

Evidence that the knob produced the intended behaviour, rather than shifting the problem elsewhere.

  • Balance between treatment and comparison groups
  • Confidence intervals, not point estimates alone
  • Placebo-test results
  • Robustness across alternative windows, peer groups and confounder controls

Common mistakes

Presenting factor correlations or backtested associations as causal findings.

Reporting a single specification without sensitivity analysis.