Base Rates in Trading: What They Are and How to Read Them

What a base rate is

A base rate is the frequency with which an event occurred, measured over all comparable cases in a period. It is not a prediction. It is a count: how many times a condition occurred, and in how many of them what we are measuring happened.

An example. If there were 4,104 sessions with an opening gap, and in 58.1% of the upward ones price touched the prior close again within the same regular session, that is the upward gap fill base rate for that instrument and that period.

It is useful for calibrating expectations and for spotting when a belief in the trading community does not match what actually happened. It is not useful for knowing what the next session will do: the past frequency describes a population, and tomorrow is a single observation from that population.

Its usefulness depends on three things that must always accompany the figure: the definition of the event, the instrument, and the sample with its period. If one is missing, the number is decoration.

The definition comes before the number

Two people can measure how often the gap gets filled and get very different results without either making an arithmetic mistake. What changes is the definition.

Three things have to be fixed: the reference the gap is measured against, the window in which the fill is allowed, and what to do with zero gaps. In these tables the gap runs from the prior regular close, at 16:00 ET, to today's open, at 9:30 ET; the fill requires price to touch that close again within the same regular session; and zeros are left out.

Change one piece and the number changes. Using the last overnight price instead of the regular close gives a different gap. Allowing the fill on later days raises the frequency. Including zeros raises it even more, because a zero gap is filled by definition. That is why comparing two sources only makes sense if both publish their definition.

Without the instrument, the figure means nothing

The most common objection to a loose statistic on social media is not that it is miscalculated. It is that it does not say which market it comes from. The table applies the same measurement to five indices.

The four US indices cluster in a reasonably narrow band, while NKD sits clearly below, with the lowest fill frequency of the five on both sides. Between the instrument with the highest upward fill frequency and the one with the lowest there are more than twenty percentage points.

That gap is the reason a figure without an instrument is unusable. Someone can quote a true percentage and still mislead you, just by omitting where it was measured.

The table shows something more. Look at the sample in each row: the RTY series covers little more than half the sessions of the ES one. The overall date range is the same, but not every instrument contributes the same history.

The sample rules: why the n sits next to every figure

A frequency is an estimate, and every estimate carries uncertainty. That uncertainty depends mainly on the number of observations. With few, the percentage swings a lot by chance; with many, it stabilizes.

The contrast is easy to see. A 60% over 5 cases is three hits: change one and the percentage jumps twenty points. A 60% over 3,000 cases would need hundreds of different cases to move the same amount. Same number, two different things.

Hence the publishing rule: no figure appears without its n and its date range. It is not a methodological ornament, it is what tells you whether the percentage deserves attention. And the n is checked by row, not just in the total: a table with 18,794 sessions can contain rows of only a few hundred. The total is reassuring; the row is what governs.

Degenerate cells: when the cut leaves the box empty

Every filter splits the sample. Instrument, day of the week, gap size, day type: four cuts turn thousands of sessions into handfuls. The boxes with so few cases that the percentage no longer informs are degenerate cells.

We have one in our tables and we do not hide it. In the day type classification, the Normal category appears in 0.1% of sessions: 5 cases in the whole history. Its continuation column reads 60.0%, which is three sessions out of five. That is not a base rate; it is an anecdote formatted as a percentage.

The same problem produces the clean extremes that draw so much attention. In the Initial Balance size table, the 1.5% to 2.0% band shows 0.0% of sessions without a break over 57 cases. That zero does not mean impossible: it means it did not happen in 57 sessions.

The consequence is uncomfortable but clear. The more specific and attractive a statistic sounds, the more likely it comes from a small cell. Specificity is paid for with sample.

An average is not a distribution

The second most costly mistake, after ignoring the sample, is treating a mean as if it were the usual case. The mean summarizes; it does not describe the variety underneath.

The table shows it with ES's daily range. Comparing each session with the instrument's recent average range, a little over half of days fall in the normal band, between 70% and 130%. The rest split between compressed and expanded, in similar proportions. Almost half of sessions do not look like the average day.

The same goes for aggregate percentages. The ES gap fill gives 58.1% on the upside, but when split by gap size the figure runs from 90.8% for the smallest to 17.8% for those larger than 1%. The aggregate is true and, at the same time, describes neither extreme well.

How to check whether a figure still holds

A base rate does not expire on a date, but it can stop describing the current market. It is worth checking with any figure, your own or someone else's. First, look at the date range. If it ended years ago, the figure describes a different market. These tables start in 2010 and run through the latest loaded sessions; that range appears at the foot of each one.

Second, split the sample and recompute. If the first half of the history and the second half give similar results, the figure is stable. If they differ a lot, the aggregate averages two different regimes and represents neither.

Third, check whether the definition is still valid: a change in session hours or market structure breaks the continuity of a series without the calculation raising an error. And fourth, recompute when new data arrives, instead of quoting a frozen figure forever.

The limits of what is measured

These tables measure historical frequencies of precisely defined events. They do not measure profitability. Between price touching a level and a trade ending in a gain there are commissions, slippage, the prior adverse excursion and the decision of when to exit. None of that is in the table.

Nor do they measure causes: a frequency describes a repeated coincidence, not a mechanism. And they do not allow filters to be combined freely. Each cut is shown separately because multiplying percentages from two tables produces a joint probability that nobody measured.

Search bias is the last limit and the hardest to see: if you test enough combinations, some will give an eye-catching number by chance. The defense is to fix the definition before the result and show the full sample, not just the nice-looking cuts.

Past results do not guarantee future results. This article is informational and educational material, not personalized investment advice or a recommendation to buy or sell. Trading futures involves risk of loss.

Frequently asked questions

What is a base rate in trading?

It is the historical frequency with which a precisely defined event occurred, measured over all comparable cases in a period. It describes what happened in that sample; it does not predict the next session or indicate a direction to take.

How much data is needed for a statistic to be reliable?

There is no magic number, but uncertainty falls as independent observations grow. A figure based on a few dozen cases swings a lot by chance. Always look at the sample of the specific row, not just the table total.

Why is a historical probability not a buy signal?

Because it measures the frequency of an event, not the result of a trade. It does not include costs, slippage or the prior adverse excursion, and it describes an entire population, not tomorrow's case. It is context for sizing expectations.

Can two statistics be multiplied to combine filters?

No. Multiplying two percentages from different tables assumes the events are independent, and they almost never are. The result is a joint probability that nobody measured. Combining filters requires recalculating on the data, accepting that the sample shrinks.

These numbers, instrument by instrument

• es
• nq
• rty
• ym
• nkd
• cl
• gc
• 6e

Where this is used in Perfiltrade

• Base-rate reports — 100+ report types across 30 instruments, every figure with n, period and caveats.
• Discovery — Ranking of patterns by historical edge and sample confidence: where to look first.
• Seasonality and correlations — Seasonality, VIX/BTC correlations and a bias explorer by event or weekday.
• Cerebro AI — AI assistant that answers with platform data and cites the report, n and caveats.
• Trading Journal — Trade log with emotion, plan and alignment with the setup's base rate.
• Backtest Lab — Monte Carlo, chronological validation and funded-account simulation on setups.

The calculation and its limits are in the methodology.