Base-rate glossary: what the words mean
Every entry has two halves: what the term means, and what it does not tell you. The second half is the one people need.
Base rate
How often something has happened before. It is counted across many past cases, and reported with the uncertainty that comes with a count. Think of an almanac entry: it has rained in Mumbai in 84 of the last 100 first weeks of July.
What it does not tell you. It says nothing about next Tuesday. Next Tuesday was not part of the sample the rate was measured on.
Matched control
Other securities from the same date as the event, picked to match it: same price band, same volatility band, same traded-value band. They did not meet the condition themselves, and they are measured in exactly the same way. They are the placebo arm, the group that shows what would have happened anyway.
What it does not tell you. It cannot control for anything it does not match on. Sector is not one of the matching keys here, and this dataset does not carry sector at all. A pattern that is really one whole sector moving together would not be caught.
Gap versus control (lift)
The event's frequency minus its controls' frequency, worked out event by event and then averaged. It is the only rate on this hub published as a headline. That is because it is the only one with the market's own direction taken out of it.
What it does not tell you. A gap is not an edge. It is one observation about one sample, measured with uncertainty, and it may not last. It also says nothing about whether it would survive costs, position sizing, or the discipline to hold on through a losing stretch.
Effective sample size
The number of genuinely independent observations a sample is worth. Events on the same day share one market move, so a raw count overstates how much you really have. Effective n is the raw count divided by the design effect.
What it does not tell you. It corrects for same-day clustering only. Windows that overlap across days are handled separately, by the block bootstrap, which redraws whole dates in blocks. That means effective n is not the whole story of how observations depend on each other.
Design effect
How much same-day clustering shrinks a sample. The formula is 1 + (mean events per date − 1) × the same-day correlation of the outcome. A design effect of 7.65 means 6,000 raw observations are worth 784.
What it does not tell you. It depends entirely on the correlation estimate behind it. That estimate is measured from the data, so it carries an uncertainty of its own. Every page publishes both the raw-design and the paired-design figure, so the two can be compared.
Confidence interval
The range the measurement is consistent with. On this hub it comes from a bootstrap: the data is redrawn many times to see how much the answer moves. Each redraw takes whole dates in contiguous blocks, so the interval carries both the same-day clustering and the overlap between windows.
What it does not tell you. It describes sampling uncertainty, not the risk that the world changed. An interval that excludes zero says the sample is inconsistent with no effect. It does not say the effect will still be there next year.
Break-even frequency
A target and a stop are set the same distance away on either side. That distance is the bracket. This is how often the target has to be reached first, rather than the stop, just to cover charges: (bracket + cost) ÷ (2 × bracket). It is arithmetic on the published charge schedule.
What it does not tell you. It leaves out slippage, market impact, the bid-ask spread, taxes on gains and the cost of capital. Every one of them moves the required frequency up, never down. Slippage here means the gap between the price you expected and the price you got.
Censoring
An observation whose security stopped trading before the window ended. It is kept, never dropped: it is closed off at the last traded price and marked as censored.
What it does not tell you. Closing at the last traded price is optimistic for a security that was suspended on its way to zero. Every page also publishes a version that counts every censored observation as a loss, so you can see what that assumption is worth.
Point-in-time universe
Eligibility judged using only what was known on the event date, from the file the exchange published that day. Securities that were later delisted, merged or suspended are still in the sample for every day they traded.
What it does not tell you. It is only as point-in-time as its sources. Surveillance-stage membership means whether the exchange had placed a security under extra surveillance that day. Index constituency means which index it sat in. Neither can be rebuilt here, so those two filters are not applied, and the methodology page says so.
Pre-registration
Writing a pattern, its full parameter grid and its horizons into a public dated log before any data is computed. The grid is every setting that is tried. The horizons are the time windows the result is measured over. The page that comes out of it is then published whatever the result.
What it does not tell you. It stops a result being chosen after the fact. It does not make the result correct, and it does not stop someone registering a pattern whose answer they already suspected.
False-discovery adjustment
A correction applied because testing many times is bound to throw up some findings by chance. Every page prints what the figure has been adjusted for: how many variants of that pattern, and how many tests across the whole site.
What it does not tell you. It controls the expected share of false findings among the ones called significant. It is not a promise that any single result is real.
Knife edge
A result that holds at one parameter setting and vanishes at the settings next to it. It is flagged when the headline sits far from the median of its declared grid. The page then shows a map of the whole grid rather than a single cell.
What it does not tell you. A flat plateau of similar results is more believable than a lone spike, but it is not proof. Neighbouring settings share most of the same events, so they were never independent tests of each other.
Placebo check
The check reruns the same securities through the same machinery. Each event is moved to an unrelated date on which the condition was not met. It should produce no gap.
What it does not tell you. A clean placebo says the method is not creating a result by itself. It says nothing about whether the result that survives is useful.
Insufficient sample
A sample that fails one of the quality checks: fewer than 100 effective observations, too few dates or securities, too short a span, or too much of it packed into one quarter or one security. When that happens, no rate is shown.
What it does not tell you. It does not mean the pattern is absent. It means this sample cannot tell, which is a different thing to say, and the honest one.
Back to the hub · The wider market glossary
Next stepSee these terms doing the work
The method page uses every word defined here in the place it matters. It shows the control every rate is measured against, and the effective sample the intervals are computed on. It also shows the quality checks a sample must pass before any figure is shown.
Read the methodology