Base Rates Before Stories

The outside view, and where it applies

The outside view is the discipline of classifying a situation into a reference class and asking how the class fared, before considering the particulars. It matters because the inside view — the story about this company, this management team, this pattern — is composed of the facts that are available and vivid, which are not the same as the facts that are decisive. Most acquisitions destroy value for the acquirer. Most turnarounds take years or fail. Most high-growth names mean-revert. A specific case can genuinely escape its reference class, and the burden of proof sits with the claim of escape, supported by evidence rather than enthusiasm. The practical refinement is that the reference class should be as narrow as the facts allow, because a rate that averages over the conditions you can observe is a rate that hides them. In the worked case the pattern is identical in both groups, and the payoff is identical, so the only difference is location — and the location doubles the expectancy. That makes the location part of the setup rather than a contextual detail, and it changes what to collect: trades tagged by condition, so that the rate can be conditioned on the part of the setup that is observable before the entry. This is also where the blended figure becomes a trap rather than a summary. Sixty per cent of the trades were taken at the level and forty mid-range, so the book produced +0.62R a trade, which is a respectable-looking number. But it is the expectancy of a mixture, and its value depends on how the mixture is composed rather than on the method: shift twenty mid-range trades to a level and both the win rate and the expectancy rise without a single new pattern being learned. The blended average hides the lever because it averages over exactly the condition the trader can act on. Same pattern, two locations — At a level that mattered: 18 of 30: Win rate 60%; expectancy 0.60 × 2 − 0.40 = +0.80R ← · Mid-range, no level: 9 of 20: Win rate 45%; expectancy 0.45 × 2 − 0.55 = +0.35R ← · Overall: 27 of 50: Win rate 54%; expectancy +0.62R · What the blend depends on: The mix of locations, not the pattern Raising the overall win rate is not the lever here. Moving trades from the mid-range to a level raises both the win rate and the expectancy, and it costs nothing but patience — which is why the condition belongs in the definition of the setup.

Turning a rate into a decision

A base rate becomes useful when it is attached to a payoff and a size. Expectancy in R answers whether a setup is worth taking at all; the number of opportunities answers whether it is worth the time; and the size rule answers how much each one should carry. A pattern with a +0.35R expectancy taken twenty times a month is a business; the same pattern taken twice a month, with the same attention and costs, is a hobby. That arithmetic is the reason the conditioning matters — a rate that is twice as good is worth waiting for, and waiting is only defensible if the payoff for waiting is on the page. The honest limitation of every base rate is that it describes a class and the trade is a member of it. A 60% rate at a level does not mean the next trade taken at a level wins; it means the collection of them does, which is why the sample size behind the estimate has to be recorded alongside it. Fifty trades is enough to see a difference of the size in the worked case and not enough to be confident of the exact figure, so the reading is directional: trade the level, avoid the mid-range, and keep collecting until the numbers tighten. That is the same discipline the previous two lessons established, applied to probabilities rather than to behaviour. It is worth naming the failure that follows from ignoring base rates, because it is the most expensive one in this subject. A trader who reasons from the inside view alone has no way to distinguish a setup that works from a story that sounds like it should: every trade has a reason, and the reasons are always compelling, because they were assembled to justify the order. The base rate is what stands between a plausible narrative and a positive-expectancy action, and conditioning it is what turns a general claim about a pattern into a specific claim about the trades actually being taken. • Expectancy in R decides whether a setup is worth taking at all. • The frequency decides whether it is worth the attention. • Condition the rate on what is observable before the entry — location first. • Record the sample size, and read a small sample directionally. • A compelling reason is not evidence, because every trade has one. The blend is the trap this lesson is built around: a respectable overall expectancy can sit on top of a good condition and a bad one, and the average will keep the bad one alive because it dilutes the evidence against it.

The test that is 95% right and still usually wrong

The reason a base rate dominates even a good signal is worth working through once, because it is the single most useful piece of probability arithmetic a trader can carry. Take a rare pattern — the kind that fires on a hundred setups, of which only **ten** are the real thing and ninety are look-alikes. Now apply a filter that is genuinely **ninety percent accurate**: it correctly passes nine of the ten real setups, and it wrongly passes ten percent of the ninety false ones, which is another nine. The filtered list is eighteen candidates of which nine are real. A ninety-percent-accurate filter on a ten-percent base rate has produced a coin flip, and the reason is structural rather than a defect in the filter: the errors are drawn from a base nine times the size of the signal, so a small error rate on the many swamps a small error rate on the few. The lesson generalises and it is the reason this lesson starts with the outside view. **The rarer the event, the more accurate a test must be to say anything useful about it.** A pattern that occurs rarely cannot be rescued by a moderately accurate confirmation, because the false positives scale with the size of the base they are drawn from. Applied to setups, that means a filter is worth adding when it removes a large fraction of the failures at the cost of a small fraction of the wins, which is a different and much stronger requirement than “it is usually right”. There is a practical version of this that costs nothing to run. Instead of asking whether a filter is accurate, ask **how many of your historical losers it would have removed and how many of your winners it would have removed with them**. Both numbers are in your own record, and the ratio between them is what the filter is worth — not its accuracy in the abstract. A filter that eliminates sixty percent of losses while eliminating ten percent of winners is worth using even if it is “wrong” often, and a filter that is right nine times out of ten but removes winners and losers in the same proportion is decoration. A 90%-accurate filter on a 10%-base-rate setup (100 signals, 10 real) — Real setups that pass: 9 of 10 — the filter is working as advertised · False setups that pass: 9 of 90 — a small error rate on a much larger base · Candidates after filtering: 18, of which 9 are real — the hit rate is now a coin flip ← The mistake this prevents is treating a confirmation as a transformation. On a rare pattern, a filter that is right nine times in ten can leave you roughly where you started — because most of what it sees is not the thing you are looking for.

Choosing the reference class, and the sample you can actually get

The previous reads treat the reference class as something to be found, and in practice it is something to be chosen — under a constraint that has a precise shape. A narrow class is a better match to the trade in front of you and contains fewer observations; a broad class has many observations and describes a situation that is only loosely like yours. A rate estimated from ten trades carries a standard error so wide that almost any value is consistent with it — with ten observations, a 60% win rate is statistically indistinguishable from 40% — while a rate estimated from five hundred trades is tight and describes a mixture of conditions that may not include yours. The honest resolution is not to pick one but to state both, weighted by how much the conditioning actually changes the answer: if the rate is similar across locations, the narrow class adds nothing and the broad one is preferable; if it differs sharply, as it does in this lesson’s worked case where location doubles the expectancy, the narrow class is worth its small sample and should be reported with the count beside it. That is the practical form of the discipline: **a base rate without its sample size is not a number, it is a claim.** “Sixty percent at a level” has a different meaning at twelve observations and at two hundred, and the meaning is captured by the interval rather than the point. The interval for a proportion is easy to carry in the head: the standard error is the square root of p(1−p) ÷ n, so a 60% rate on 25 trades has a standard error of about ten points, and on 100 trades about five. That is enough to decide whether a difference between two groups is worth acting on — the difference from this lesson’s worked case, where one location doubled the expectancy, is large enough to survive a small sample, while a five-point advantage between two classes of thirty trades each is not. The last piece is the one that makes the exercise compound: the class is conditioned **before** the outcome is known, and the conditioning variables are recorded at entry. A base rate built by splitting the journal after the fact into whichever groups look best is the same multiple-comparison failure the evidence lesson describes, applied by hand. The defensible procedure is to write down, in advance, the two or three observable conditions a setup belongs to — location, frame, volatility regime, time of day — and then to accumulate trades within those groups until the counts are large enough for a comparison. The payoff is that the resulting number is not an estimate borrowed from someone else’s study but a measurement of the trader’s own behaviour, and it is the only kind of base rate that can be trusted to describe the trades actually being taken. Same trade, three reference classes — All pattern trades: n = 200 · a tight estimate of a broad, mixed class · Pattern at a repeated level: n = 48 · the conditioning that doubles the expectancy, with a usable sample ← · Pattern at a level, in a weekly uptrend, after a pullback: n = 9 · a better match and an interval too wide to act on ← The pairing with the habit this lesson installs is exact: condition the rate, but log the count. A rate whose sample size is missing will be revised by the next bad month, and the trader will conclude the market changed rather than that they were reading an interval as a point.

Sources

• The outside view and reference classesKahneman, “Thinking, Fast and Slow”
• Conditioning on the setup rather than the storyMauboussin, “The Base Rate Book”
• Expectancy and win rate with R multiplesVan Tharp, “Trade Your Way to Financial Freedom”