Equities, FX, futures, options and market microstructure. Every run is logged and replayable, every finding is written up with its limits, and every figure is tied to the run that produced it.
Six stages, in a fixed order. The order matters more than any single test: nothing is allowed to look good before it has been allowed to fail.
Every study starts from a mechanism: a rule, constraint or incentive that forces prices to behave a certain way, and a statement of what would prove us wrong. Patterns without a reason are what any large dataset hands you for free.
best = max(corr(c, target) for c in candidates[:36])
Hover a thumbnail to inspect it. The best match so far is outlined.
Waiting for the search to run.
Estimation happens on a fixed window under one pre-stated specification. The more flexible a model is, the better it explains history, which is exactly why explaining history proves nothing.
A straight line. Misses the structure, and misses it the same way everywhere.
A partition the research environment cannot reach is held back. It is opened once, after the specification is frozen. If the idea is adjusted after the look, the look is spent and the result no longer counts.
The same pipeline runs on versions of the idea that should have no edge: signal inverted, timing shuffled, noise standing in. A real effect stands clear of every one of them. Drift, leakage and bugs do not.
Running permutations.
Spreads, commissions, impact, financing and execution delay are subtracted, not footnoted. A signal formed at one close is traded at the next open. What remains is the edge that scales to the size being asked of it.
Most of the edge survives here.
A candidate must clear every gate; one failure ends it. Failures are not deleted. Each is recorded with its mechanism and screened against every future hypothesis before a line of code is written.
| Check | Evidence required | Result |
|---|---|---|
| Data admission | Timestamps as known, next-open execution | |
| In-sample study | One specification, fixed in advance | |
| Blind holdout | Single look at the partition | |
| Stability | Holds outside the study window | |
| Placebo controls | Stands clear of every control | |
| Parameter neighbourhood | Neighbouring settings agree | |
| Costs and capacity | Positive net of cost, at size | |
| Drawdown budget | Within the loss budget | |
| Sample size | Enough independent events | |
| Independence | Not a copy of existing exposure |
Each desk keeps its universe, its calendar, its cost model and its window on the same run as the figure. The ground differs; the discipline does not.
The configuration a live book would have faced.
Corporate actions and borrow travel with the signal. Cross-section and event study in one contract.
A universe that quietly drops the names that failed, or a short side nobody could borrow.
Rigour is not a tone. It is a short list of specific habits, each one aimed at a way empirical finance is known to fool itself.
A partition kept out of reach of the environment where ideas are developed. It is read after the specification is frozen. Reading it twice turns it into more study data.
AskCould anyone have seen this partition before the idea was fixed?
Researchers and agents work on separate machines at once. A frontier can be held by one of them at a time, every run lands in a single immutable history, and a failure found on one desk is known to all of them.
Claim a frontier that is already held, then release one as falsified.
Quantitative Finance
The published work this discipline leans on. Not endorsements, only where to start.
Quantitative workspaces are provisioned per mandate, with lead approval on every write.
Request access