Thinking in probabilities: why one chart proves nothing
A setup is a conditional probability with a payoff. One win or one loss says little about it; a counted sample, after costs, says more.
Start in plain English
Scalping, also called scalp trading, is a style of very short trades: "scalps" or "base hits" that aim for a few cents and last seconds to minutes. This lesson explains why a single result, or a handful of them, tells you almost nothing about a setup. For a small-cap scalper, who may take dozens of trades a week in stocks that move fast and halt, reading single outcomes as proof is a quick way to drop a sound plan or keep a broken one.
The one idea: a setup is a conditional probability with a payoff. "Conditional" means given these conditions. "Probability" means some share of the time, not all of it. "Payoff" means how much is won or lost on each outcome, after costs.
So when someone posts a chart where a setup worked, the honest reading is: this is one draw from a distribution we cannot see in the picture. The same chart, with the same information at the decision time, could have ended in a loss or in nothing at all. One outcome, win or loss, is weak evidence about the setup that produced it.
Every example in this course is therefore drawn at least twice: the path people remember (Path A) and a path that was equally consistent with what was knowable at the time (Path B), plus a "nothing happened" path (Path C) when that is common.
Base rates and conditional probability
A base rate is how often something happened in a defined sample. A conditional rate is how often it happened in the part of the sample that met an extra condition.
Here is a measured example from FennEdge's own research.
That 65.3% is a base rate for that definition, that period, that archive. It is not a forecast for any one stock tomorrow, and it says nothing yet about whether any trade on these stocks makes money after costs.
Now a conditional rate from the same study. Among the names whose high of day printed at 10:00 ET or later (n = 200; same period, archive and as-of date), 69.5% closed at or above their open. Across the whole sample, only 34.7% did (233 of 671; FennEdge computation from the Gapper HOD study data, 2026-07-01 to 2026-10-02, as of 2026-10-02). Adding one condition moved the rate from about one in three to about two in three.
That looks like a powerful filter until you ask when the condition is known. You only learn that a stock's high came after 10:00 once the day is over. The study says so directly: this split uses information you only have after the close. A conditional rate is useful for decisions only if the condition was knowable at the decision time.
Outcome quality vs. decision quality
Two separate questions apply to every trade:
- Was the decision sound given what was knowable? Were the conditions met, was the risk planned, was the size correct, were costs considered?
- What happened? Win, loss or scratch.
Judging the first question by the second is called outcome bias. Psychologists Baron and Hershey showed in 1988 that people rate the same decision as better when they are told it turned out well. In trading this leads to two errors: keeping a careless habit because it won once, and abandoning a sound plan because it lost three times.
A useful review note separates the two: "Plan followed, conditions met, 1R planned, stopped for −1.1R after costs" is a different note from "Size doubled, no stop, +3R." The second trade made more money and was the worse decision.
Hindsight and survivorship in chart examples
Most chart examples you see, including on trading education sites and social media, have three filters applied before you see them:
- Selection. Someone chose which chart to show. Charts that worked get posted more often than charts that failed or went sideways.
- Hindsight labels. Captions such as "the low of the day" or "the breakout" are only knowable after later bars printed. Fischhoff's 1975 experiments found that once people know an outcome, they overestimate how predictable it was.
- Survivorship. Traders who had a bad run often stop posting, so the visible record skews toward lucky runs.
None of this means the poster is dishonest. It means a single chart cannot carry the weight of a rule.
Observe vs. infer
A caption on a fictional chart of Quillfeather Bio reads: "Perfect dip buy at the low of day, $4.10 at 09:41, buyers stepped in, ran to $4.60." Split it apart:
| What the chart shows (observed) | What the caption adds (inferred or hindsight) | Why it is not established at 09:41 |
|---|---|---|
| A 1-minute bar at 09:41 traded down to $4.10 | "The low of day" | Only known after the session ended; at 09:41 the next bar could trade lower |
| The 09:41 bar closed at $4.15, five cents above its low | "Perfect dip buy … ran to $4.60" | The $4.60 bars had not printed at 09:41; "perfect" is judged by that later result, and the same entry could have been stopped if the next bar broke $4.10 |
| Volume on the 09:41 bar was lower than the prior five bars | "Buyers stepped in" | Prints are anonymous; the tape does not show who or why |
The left column is what a trader could have used at 09:41. The middle column is what makes the chart persuasive.
Two paths from the same start
Setup X is a fictional scalp with fixed rules. Its outcome distribution below is Illustrative: chosen to teach the idea, not taken from a study.
That is a positive estimated edge. It still loses or scratches 60% of the time. We simulated 1,000 runs of 10 trades from this distribution (computed from the assumed distribution, Illustrative). The median run finished at +2.0R; one run in ten finished at −3.0R or worse; one in ten at +8.0R or better. Computed exactly, about 26% of 10-trade runs finish below zero.
Here are three of those simulated runs. We picked them to show the spread, which is itself selection: the very bias the previous section described.
Path AThe run people remember
Trades: +2, +2, +2, −1, 0, +2, −1, +2, +2, −1 (R, before costs). Net after costs: +8.0R over 10 trades.
The learner concludes "Setup X works." The conclusion is too strong: a run this good happens in about one run in ten under the same distribution.
Path BSame distribution, opposite story
Trades: −1, −1, −1, −1, +2, +2, −1, 0, 0, −1. Net after costs: −3.0R.
The learner sees four losses in a row and concludes "Setup X is broken." Nothing changed: same rules, same distribution, same +0.25R expectancy. Four losses in a row in the first four trades happens about 4% of the time at a 45% loss rate (0.45⁴).
Path CNothing much happened
Trades: +2, 0, +2, −1, −1, −1, −1, 0, +2, −1. Net after costs: 0.0R.
A flat run is easy to ignore, but it is common: in our 1,000 simulated runs, 23.1% finished between −1R and +1R. The costs still had to be paid.
All three learners traded the same edge. Ten trades could not tell them apart.
What shifts the odds
In this lesson the conditions are about your evidence, not the market:
- Sample size. The probability that Setup X shows a net loss falls from about 26% over 10 trades to about 9% over 50 and about 3% over 100 (computed from the assumed distribution, Illustrative).
- Independence. Five trades in the same stock on the same morning are less informative than five trades on five different days. They share one news event and one market mood.
- Definition before counting. If you decide which trades "count" after seeing their outcomes, the sample is no longer a sample.
- Knowable conditions only. A condition that is known only after the close (like "high came after 10:00") cannot shift your odds at the decision time, however large its effect looks in a table.
Cost of being wrong
Two different mistakes cost money here.
- Abandoning a positive-expectancy plan after a normal losing streak. At a 45% loss rate, a run of three or more losses appears somewhere in a 10-trade block about 41% of the time, and somewhere in a 50-trade block about 95% of the time (Illustrative, computed). Dropping Setup X after Path B gives up +0.25R per trade of estimated edge.
- Scaling up a negative-expectancy plan after a lucky run. If the true expectancy were negative, a Path A run would invite larger size into a losing process. At a fixed $50 risk per trade, a −0.25R process loses about $12.50 per trade on average; doubling size after a lucky streak doubles that rate.
In both cases the loss comes from treating 10 trades as an answer.
How many trades before this means anything
For Setup X, the uncertainty around the +0.25R estimate is wide. Using the per-trade spread of this distribution (about 1.39R), a rough 95% range for the average after 20 trades runs from about −0.36R to +0.86R, and after 100 trades from about −0.02R to +0.52R (Illustrative, normal approximation). Even 100 trades barely separate this edge from zero. How many trades before your numbers mean anything covers the method; Expectancy covers the formula.
What this lesson does not claim
- It does not claim any real setup has the Setup X distribution. Those numbers are Illustrative.
- The Gapper HOD figures describe one 66-session window and one definition. They are not a forecast and not a trading rule.
- Thinking in probabilities does not make a strategy profitable. It only stops you from drawing strong conclusions from weak evidence.
- Base rates from the market and base rates from your own trades are different numbers. Your fills, costs and timing are part of your distribution.
Go deeper
- FennEdge research: What time do small-cap gappers hit high of day? is a worked example of base rates with stated caveats, including the "only known after the close" split. In-Play Days: research a stock's trading history shows how to compare past active days while keeping what was knowable at the time separate from hindsight.
- FennEdge tools: the Gapper HOD Timeline shows when gappers printed their highs. Guests, Free and Plus see the latest completed session; earlier sessions need Pro or Max (as of 2026-10-05). It shows completed sessions, not live signals.
- Base-rate neglect. Tversky and Kahneman (1974) describe how people ignore base rates when given a vivid specific case: one striking chart against a dull sample count.
- Social-media screenshots. A screenshot of a P&L is one outcome with selection, hindsight and survivorship applied. Ask: what was the sample, what were the costs, and who did not post?
- Next in this course: Expectancy turns these ideas into a formula; A trading plan as a hypothesis writes the Path B response into your plan before the trade; VWAP reclaims and rejections applies the two-path format to a chart setup.
Exercises
Six exercises follow this lesson: three concept checks, one scenario that separates decision quality from outcome, one caption review (find the hindsight), and one calculation of a losing-streak probability. Exercises, answers and progress are available after signing in.
Primary sources
- Tversky, A. and Kahneman, D. (1974). Judgment under Uncertainty: Heuristics and Biases. Science 185(4157).
- Baron, J. and Hershey, J. C. (1988). Outcome bias in decision evaluation. Journal of Personality and Social Psychology 54(4).
- Fischhoff, B. (1975). Hindsight is not equal to foresight. Journal of Experimental Psychology: Human Perception and Performance 1(3).
- FINRA. Frequent Intraday Trading: Understanding the Basics.
Sources checked 2026-10-05. Setup X, Quillfeather Bio and their numbers are fictional and Illustrative. Measured figures come from the FennEdge archive as stated. This lesson is education, not a recommendation to buy or sell any security. Past patterns do not indicate future results.