Control groups
An astronomical mask is only interesting if it differs from ordinary market days. · 8 min
Why controls exist
Many days in a 25-year sample will look “special” if you only report the event mean. The Lab always asks: different from what?
The four built-in baselines
- Unconditional — every eligible bar in the window.
- Random timestamps — permutation of event times (the default p-value).
- Same calendar month — seasonality check.
- Matched trailing volatility — similar recent range, different sky flag.
Caveat
Demo market data is a seeded GARCH generator with no sky input. The correct result on it is a null. Those runs are labelled DEMO and are never promoted.
Key points
- Effect size is Cohen’s d versus the unconditional baseline.
- The UI defaults to unconditional + random; at least one baseline is required.
- A large raw mean with no baseline is not evidence.
References
- MacKinlay, Event Studies in Economics and Finance — MacKinlay A.C., Journal of Economic Literature 35(1), 1997, 13–39
- RAINBOW baselines — src/domain/research/baselines.ts — unconditional, random, same-month, matched volatility