Home/Psychology/Uncertainty and randomness

Trading under uncertainty – how to judge decisions whose outcome cannot be known

Five losses in a row can occur even under a rule that wins more often than it loses. In a model with a 55% win rate and independent trades, the probability of such a streak occurring anywhere within 100 decisions is about 65%. The streak alone therefore cannot tell you whether the rule has stopped working. Assessing that requires separating a chance outcome from decision quality, understanding the limits of your data and recording the conditions that would change your view. This article explains how to make and assess decisions when the outcome of an individual trade cannot be known in advance.

The same set of rules produces different endings – what gets judged is the record written before the decision, not the path that happened to fall out.

Define the event before writing down your confidence

“I expected the price to rise” leaves too much room for reinterpretation. Did you mean a higher next close, a target reached before an exit, or a positive account result? Those are different events. A probability becomes checkable only after the event, observation time and treatment of equality are fixed. Changing the event after seeing the chart changes the forecast being assessed.

For an invented exercise, each forecast concerns whether the next completed M15 Bid candle closes strictly above its own opening Bid. A higher close is coded 1; a lower or equal close is coded 0. Record the probability before that candle opens, using the same declared clock and price feed. The four observations below are stipulated and fully resolved. They describe candle direction, not executable fills or net trading profit. No entry decision or position is assumed.

Give the forecast a score that retains its probability

Four hypothetical forecasts of the same defined event
CaseProbability pObserved event ySquared error (p−y)²
A0.6010.16
B0.6000.36
C0.8010.04
D0.8000.64

Use the mean of the four squared errors: (0.16+0.36+0.04+0.64)/4=0.30. Lower is better under this stated binary scoring convention. Case D contributes more than B because a non-occurrence was assigned a higher probability. Case C contributes less than A because the event occurred and received a higher probability. The score retains information that a simple count of correct directional calls would lose.

This is the one-event mean-square convention described in NOAA’s forecast-verification memorandum. Its weather application is not evidence of a trading advantage. State the convention when reporting a score: a formulation summing both complementary event categories would double this number. Neither version has currency units.

Compare on the same events without rewriting the forecasts

A constant probability of 0.50 on each of these same four events produces an error of 0.25 on every row and a mean of 0.25. The recorded forecasts therefore score worse by 0.05 on this tiny set. That is an exact descriptive comparison. It is not a demonstration that the constant forecast is superior in future market conditions.

The realised event frequency is 2/4=50%. That frequency does not make every earlier probability above 50% an error in itself. Probabilistic forecasts allow events not to occur. Nor does the observed frequency establish calibration of the 0.60 and 0.80 groups: each group contains only two cases. Retain the individual probabilities rather than replacing them afterward with the observed frequency. If an event cannot be resolved from the available record, mark it unresolved and disclose its treatment instead of quietly choosing a convenient outcome.

Keep the small calculation separate from a claim about skill

The table shows what happened to four pre-recorded forecasts under one definition. It supplies no representative sample, independence guarantee or account return. Repeating a similar calculation on overlapping candles can add rows without adding independent information. A longer evaluation needs a prespecified sample, comparison method and record of changes to the forecasting rule.

The research distinction between different forms of overconfidence also matters here. A probability score does not measure whether someone believes they outperform other people, and four outcomes do not diagnose a personal trait. Use this worksheet to make a statement auditable, then use the detailed lesson to examine sample uncertainty, streaks and the limits of calibration. The practical change is modest: replace a remembered feeling of certainty with a forecast whose target and timing can actually be checked.

Three exercises with worked answers

The exercise data are hypothetical. Choose an answer, then check the calculation.

1. What is the mean squared error under the stated convention?

Four row errors: 0.16, 0.36, 0.04 and 0.64.

Show the answer and explanation

C. Average all four row errors under the same binary convention. The result describes forecasts of the specified event.

  • A. That is the sum, not the mean.
  • B. That is the separate constant-probability comparison score.
  • C. Correct: the errors sum to 1.20; dividing by 4 gives 0.30.
  • D. A probability score has no currency unit.
2. How is an equal opening and closing Bid coded in this exercise?

The event is a next completed M15 Bid close strictly above its own opening Bid.

Show the answer and explanation

A. Apply the declared event definition. Do not substitute account profit for candle direction.

  • A. Correct: equality is a non-occurrence of the stated event.
  • B. The event is a strict price comparison, not absence of a loss.
  • C. The coding rule was fixed before the result.
  • D. No trade or cost record was supplied.
3. What does the lower constant-forecast score establish here?

Recorded forecasts score 0.30; a constant probability of 0.50 scores 0.25 on the same four cases.

Show the answer and explanation

D. Keep the exact comparison and its limited scope. More ambitious claims need an appropriate evaluation design.

  • A. No future distribution, execution or return model was established.
  • B. A four-case score does not diagnose a person.
  • C. Overwriting the original forecast destroys the record being evaluated.
  • D. Correct: the comparison describes the specified observations only.

Continue to the full treatment of uncertainty

The detailed sections retain the illustrations, losing-streak models and calibration discussion. The four forecasts above are a separate scoring exercise with their own fixed event definition.

1. Risk is not the same thing as uncertainty

In Risk, Uncertainty, and Profit (1921) the economist Frank Knight separated two situations that everyday language glues into a single word[1]. Risk is the state in which the distribution of possible outcomes is known in advance: a roulette wheel has thirty-seven pockets and nobody has to estimate that from a sample. Uncertainty is the state in which the distribution has yet to be estimated, and in which its shape may change over time.

Retail trading involves uncertainty, rather than just a known probability distribution. In a casino game, the odds follow from the rules; a trader estimates the hit rate from a finite sample collected under conditions that may change. These are different kinds of knowledge. A market estimate is not simply a less precise version of a fixed probability specified in advance.

The consequence bears on the sentence a trader utters most often. "My setup has a 55% win rate" is shorthand. The full version runs like this: over some number of trades I recorded 55% hits; at that number of observations the data are also consistent with values clearly lower and clearly higher; and beyond that I do not know whether the conditions under which those observations arose still hold. The second sentence is far less comfortable, but it is the one that describes the actual state of knowledge.

This entire article assumes that the rule in question has some justification beyond having looked good in history. The question of whether a set of rules has positive expectancy after costs belongs to section 8 and is settled there. Psychology does not create an edge where there is none, and no arithmetic in this text substitutes for that answer.

The share of losing accounts is a share, not an explanation. The obligation to publish the percentage of retail accounts that lose money on contracts for difference was introduced by the 2018 product intervention of the European Securities and Markets Authority and kept in place by national rules[10]. ESMA material from that period quoted a range of 74–89% of accounts, depending on the provider and the period studied. Providers covered by the applicable CFD disclosure rules publish their own current figures for the relevant clients and reporting period. Requirements differ by jurisdiction, so check the rules and current disclosure applicable to your provider. A single percentage is not a universal quantity and expires along with the period it covers. The percentage itself says nothing about causes: it does not separate costs, leverage, the quality of the rules and behaviour. Position as of August 2026; the regulatory detail is covered in section 13.

2. Win rate is an estimated quantity, not a known one

The win rate of a rule is not a parameter read off a manual; it is the result of a measurement on a sample. Every measurement on a sample has a width. The table below shows how wide that uncertainty is if a trader consistently records 55% hits and differs only in the number of trades over which that was computed.

Observed win rate and 95% consistency interval by sample size
Number of tradesObserved win rateInterval consistent with the data (95%)Width of the interval
2055.0%34.2% – 74.2%40.0 pp
5056.0%42.3% – 68.8%26.5 pp
10055.0%45.2% – 64.4%19.1 pp
20055.0%48.1% – 61.7%13.7 pp
30055.0%49.3% – 60.5%11.2 pp
50055.0%50.6% – 59.3%8.7 pp

Method: Wilson confidence interval for a proportion, 95% level, assuming independent trials and a constant probability of a hit. The number of hits is rounded to a whole number, hence the 56.0% in the row for fifty trades. Values computed independently; the interval describes the uncertainty of the measurement, not the range of future outcomes.

The practical reading is uncomfortable. After fifty trades the data are consistent both with a win rate clearly below 50% and with one well above it. Win rate alone does not determine profitability, which also depends on gains, losses and costs. Only at 385 trades does the lower bound of the interval at exactly 55% cross 50%, and with the number of hits rounded to a whole number – as in the table above – it stops falling back below only from four hundred onwards. Even then this is a statement about a past sample, not a promise about the future. Formally testing whether a 55% win rate differs from 50% requires a still larger sample – on the order of eight hundred trades; the sample-size calculation itself, together with its assumptions and the variant for comparing two rules against each other, is handled in the article on judging the process. All that is needed here is the order of magnitude: judging a rule is a matter of hundreds of trades, not a dozen or so.

The interval in the table measures only one thing: the uncertainty that comes from the number of observations. It does not measure the second one, often more important in trading – whether the conditions under which those observations arose still hold. If the win rate were a fixed property of the rule, a longer sample would simply bring the estimate closer to the truth. If it changes over time, however, a longer sample mixes different periods together, and mixes them the more, the longer it is. That is the practical meaning of the distinction from the first section: under uncertainty, collecting data does not converge on certainty, only on an ever more accurate description of the past.

The tendency to draw strong conclusions from small samples is not a beginner's failing. Amos Tversky and Daniel Kahneman handed a questionnaire to 84 professional psychologists and showed that even researchers who work with statistics professionally expect a small sample to reproduce the distribution of the population, and expect a result obtained on a small sample to repeat in the next one[2]. The paper appeared under the title "Belief in the law of small numbers" in Psychological Bulletin in 1971.

[EXTRAPOLATION] The subjects were scientists, the task took the form of a questionnaire, and the answers carried no real financial stake. What I carry over from this paper is the mechanism – systematically overrating how much a small sample says – not the effect size and not a claim about traders. I have not established the replication status of the 1971 questionnaire itself and know of no registered repetition of it – this is a single measurement, not an independently confirmed regularity.

3. Which losing streaks are consistent with the model

The question "is a streak like this normal" only makes sense once you have established which of two different quantities is meant. Confusing them is the source of most of the misunderstanding on this topic.

The first quantity: the probability that k trades designated in advance will all turn out to be losses. With independent trials and a probability of loss q it is simply:

P = qk

At a 55% win rate, that is q = 0.45, eight losses in a designated batch of eight trades is 0.458 = 0.168%. That number answers a question asked before the streak: "what is the chance that the eight trades I am about to place will all end in a loss".

The second quantity: the probability that a streak of at least k losses appears anywhere within a run of N trades. That is the question a trader actually asks – and the answer is an order of magnitude larger. You cannot get it by multiplying the first quantity by the number of windows: within 250 trades there are 245 overlapping six-element windows, but neighbouring windows share five trades, so they are not independent events. Treating them as independent inflates the result by anything from a few per cent to almost double, and the more so the longer the streak in question.

The correct calculation is simple, though it does not fit into a single multiplication. The state of the system is the current length of the losing streak: zero, one, two and so on. Every loss moves the system one state up, every hit brings it back to zero, and reaching state k is absorbing, because the streak has already occurred. Stepping through N steps gives the exact value. The numbers below I computed with that method and checked against a simulation of 200,000 repetitions with seed 20260810 – the repeated NumPy check differs by at most 0.21 of a percentage point.

The method can be checked on an example small enough to be done by hand. Take a coin, that is a 50% win rate, a streak of three losses and a run of just five decisions. A designated batch of three is 0.53 = 12.5%. There are, on the other hand, 32 five-element sequences and all of them are equally likely; of these, eight contain three losses in a row somewhere. The probability that such a streak occurs anywhere within five decisions is therefore:

8 / 32 = 25%

That is exactly twice as much as for a designated batch of three – and the state method described above gives the same value. For a run of 250 decisions, writing out every case is impossible, but the principle does not change. I checked the agreement of the two ways of counting for several other combinations up to twenty decisions long as well; the results are identical to the thirteenth decimal place.

Probability of losing streaks under independent model assumptions
Win rateSequence designated in advanceAnywhere in 100 decisionsAnywhere in 250 decisions
3 losses5 losses8 lossesstreak ≥5streak ≥8streak ≥5streak ≥8
45%16.6%5.03%0.84%92.0%30.7%99.8%61.4%
50%12.5%3.13%0.39%81.0%17.0%98.6%38.4%
55%9.11%1.85%0.168%64.7%8.4%93.0%20.3%
60%6.40%1.02%0.066%45.9%3.6%79.2%9.2%

Illustrative values, computed for independent trials and a constant win rate. The horizons N = 100 and N = 250 correspond roughly to half a year and a year of trading at somewhere between a dozen and twenty-odd trades a month; at a different frequency the numbers in the last two column groups change, while the first three do not.

For one value of the win rate it is worth seeing the whole picture. Below is a 55% win rate and a year covering 250 decisions.

Cumulative probability of a losing streak within 250 decisions
Streak of at leastIn a designated sequence of that lengthAnywhere within 250 decisions
≥ 39.11%practically certain
≥ 44.10%99.8%
≥ 51.85%93.0%
≥ 60.83%68.6%
≥ 70.37%40.0%
≥ 80.168%20.3%
≥ 100.034%4.4%

Six losses in a row within a year is, at this win rate, an event that happens more often than it does not. Eight – in roughly every fifth year of trading. Ten – in about 4.4% of such annual sequences. Under independent, identical years, that corresponds to an average wait of about twenty-three years, rather than a prediction of when the streak will occur or a claim that it will not.

What this arithmetic does not say. It does not say that the win rate is 55% – that was a number substituted into a formula, not a measured quantity; with a real estimate from the second section, the whole table carries an uncertainty interval of its own. It does not say when a streak will occur, or whether one is under way right now. It assumes that trades are independent of one another, and that is a simplification: market volatility clusters in time, so a rule designed for one type of conditions can generate losses in runs, and real streaks are sometimes longer than the model implies. It also assumes a two-valued split: every trade is either a hit or a loss, that is q = 1 − p. The definition of a hit in the sixth section, meanwhile, allows a third ending – a close on a time condition, which is not a hit – so before you count streaks on your own journal, assign that category to one of the two and write down which. Without that, the same journal will give two different results, and the difference will be the larger, the longer the streak the question is about. Finally, it does not say the thing a trader cares about most: whether this particular streak comes from a rule that still works or from a rule that has stopped. The length of a streak alone cannot settle that – settling it requires a condition written down beforehand, which is the subject of the fifth and sixth sections. And one last thing: the arithmetic says nothing about the consequences for capital, because those depend on position size, and that is set in section 9. What this number is good for further on: a streak that falls out in most years of trading is a poor trigger for a break from trading – why, is explained by the article on reset protocols.

4. The same rule, six different paths

The table in the previous section describes a single number. It is easier to see what it means by looking at several complete paths at once. The chart below shows six independent draws from the same rule: the same 55% win rate, a hundred consecutive decisions, no difference in how it operates. All that separates them is the order in which the hits and the misses fell.

Six paths of the same rule: 55% hit rate, 100 decisions vertical axis: cumulative hits minus misses −15 −10 −5 0 +5 +10 +15 +20 +25 +16 +10 +6 +4 −2 +24 0 20 40 60 80 100 decision number
One rule, the same 55% win rate, six independent draws of a hundred decisions each (seeds 101–106, independent trials). The paths end at +24, +16, +10, +6, +4 and −2. The longest losing streaks across these six paths are four on three occasions, five on two and six once.

Five paths end above the line, one below it. The trader who drew the path ending at minus two has, after a hundred decisions, exactly the same rule as the one who drew the path ending at plus twenty-four. He also has a full set of arguments for concluding that the rule does not work: six losses in a row still in the first third of the path, a long flat stretch, a result below zero after half a year of work.

The differences between the paths build up not at the end but in the middle. Up to roughly the thirtieth decision all six lines fit inside a narrow band and none of them supports any conclusion. That is the practical meaning of the second section – the first few dozen observations do not separate paths that will ultimately end up very far apart.

One decision, four endings

The same problem shows up in a single trade. Illustrative example. The condition written in the plan is met, the protective order sits where the plan says it should, and so does the target. The possible endings:

  1. price reaches the target and the position closes at a profit;
  2. price reaches the protective order and the position closes at a loss;
  3. price approaches the target, turns back and reaches the protective order – a loss after the position had been clearly in profit for some time;
  4. price does neither one nor the other, and the position is closed in line with the time condition written into the plan.

Which of these endings changes the assessment of the decision? None. The assessment covers two things: whether the condition was met at the moment of entry and whether the execution matched the plan. Both are settled before any of the endings arrives. Something else does change – one observation is added to the run, and the run needs hundreds of them before it starts to mean anything.

The third variant deserves separate attention, because it is the source of most of the later tinkering with the plan. A position that was close to the target and came back to the protective order looks like an avoidable loss and offers a ready-made conclusion: I should have closed earlier. The thing is that at the moment of the decision the third variant is indistinguishable from the first, and a rule of "I close earlier" applied consistently would also cut short some of the endings of the first type. Whether such a change pays off is settled on a run and on data, not on one painful case.

The vertical axis deliberately counts hits rather than money. Translating such a path into the state of an account requires two further things: the ratio of profit to loss, described in section 8, and position size, set in section 9. A tool for computing your own fan of paths together with the distribution of drawdowns is available in the Monte Carlo simulator; all that is needed here is the observation itself, that one rule produces many paths.

5. "I made money breaking the rules" – why profit does not justify a violation

The previous sections show why a loss does not prove an error. The reverse statement is equally true and far less often spoken: a profit does not prove that the decision was correct. Without that second half the whole line of reasoning works one-sidedly and turns into an excuse.

Illustrative example, illustrative values. The plan contains a condition: I do not open a position if less than thirty minutes remain until a scheduled data release. The condition was written earlier, calmly, for a reason the trader can reconstruct. One day the setup looks exceptionally good and twelve minutes remain until the release. The trader enters. The data come out on his side and the position closes with a profit equal to two units of risk.

The assessment of that decision does not change one iota. At the moment it was taken a written condition was in force and it was not met. Everything that happened afterwards is information the trader did not have at the time – and a decision is judged by the information available before it. What the profit does change is something else: it makes it harder to recognise that decision as a violation later, because what stays in the record is a positive result, while the condition that went unmet has to be dug out of it.

[PRACTICE WITHOUT RESEARCH] I am not claiming here that a profit increases the frequency with which the same person repeats the violation – I know of no measurement that establishes this. The sentence above concerns only what is visible in the trade record after the fact, and the conclusion in the next paragraph rests on nothing else.

That is why, in the trade record, the "condition met" column has to be filled in from the condition, not from the outcome. The trade in the example above is a violation and is recorded as one, despite the positive result. A trade that met the condition and ended in a loss is not a violation and is recorded the same way. How to construct such a record is described in the article on the trading journal.

The four combinations of decision quality and outcome quality – good decision with a good outcome, good with a bad one, bad with a good one, bad with a bad one – have their own discussion in the article on separating outcome from process. It is enough here to mark which cell the example above belongs to and why that particular cell is troublesome: it is the only one in which the market pays for behaviour the trader does not want to repeat.

The opposite excess is an error too. The sentence "the outcome does not matter" is false. The outcome does matter – as one observation in a run from which, after a few hundred repetitions, something can be computed. It does not matter as a verdict on a single decision.

A counterexample to the main thesis. From the fact that a losing streak fits inside the distribution it does not follow that reacting to a losing streak is always a mistake. Sometimes it is justified – but what settles that is something other than the length of the streak. If the trader wrote down beforehand under what observable circumstances he would treat the rule as out of date (for example: the condition the rule rests on has stopped occurring at all; transaction costs for this instrument have risen in a way that changes the arithmetic; the instrument has been replaced by another), then meeting such a condition is grounds for a change – regardless of whether a losing streak happens to be under way. If no such record exists, the losing streak remains the only premise, and on its own it does not distinguish a rule that works from a broken one. The difference lies in when the criterion came into being, not in how much it hurts.

6. Stated confidence and how to assess it

If an individual outcome cannot be controlled and the hit rate must be estimated, one useful check concerns what the trader records before placing the order: stated confidence and how closely it matches the later frequency of hits. This assesses calibration, rather than every aspect of decision quality. You can also examine whether the available information was used and the prior plan was followed.

Instead of the question "do I enter or not", the question runs: how convinced am I that the condition I judged to be met really is met and that the scenario will play out. The answer "65%" carries different information from "85%", and both are something other than "I am sure". The goal is not high certainty. The goal is calibration: that decisions where the trader declared 70% end in a hit in roughly seven cases out of ten.

Overconfidence is not one single phenomenon. Don Moore and Paul Healy separated three distinct things: overestimating one's own performance, believing oneself better than others, and excessive precision of estimate, that is declaring narrower intervals than knowledge warrants[5]. The paper combines a review, a theoretical model and an original experiment in which 82 students completed paid quizzes of varying difficulty. What is carried over here is the conceptual distinction, rather than an effect size or a finding about traders. For a trader it is mainly the third form that matters, together with its counterpart in calibration: the problem is not that somebody is highly convinced, it is that declared certainty systematically exceeds the win rate.

The record made before the decision

Calibration cannot be checked from memory, because the recollection of one's own belief is formed after the outcome is known and is distorted by that outcome. What is needed is a record made beforehand. The minimal version has five fields and takes less than a minute.

Pre-decision record fields and verification purpose
FieldWhat to enterWhat for
TimeHour and minute of the entry, before the order is placedLets you check later that the entry was made before the decision, not after the outcome
ConditionWhich written condition I judge to be metSettles later whether the trade complied with the rule, independently of the outcome
DirectionThe scenario I am betting onWithout it there is no way to decide whether the forecast came true
BeliefOne of the bands: 50–60%, 60–70%, 70–80%, 80–90%The only field that allows anything to be calibrated
What would convince me I am wrongOne sentence about an observation that argues against the scenarioForces you to consider contrary evidence before the decision rather than after it

Bands instead of single percentages are deliberate. The difference between 63% and 66% is not a quantity a human being can estimate reliably, and the apparent precision only makes later analysis harder. Four bands are enough. Leaving no entry at all is also an acceptable answer: if, after reading the condition, no sensible degree of belief can be attached to the scenario, then not taking the trade is a correct output of the procedure, not a failure of it.

The check after a run

Before the first entry is made, one thing has to be settled: what counts as a hit. In this scheme a hit is the realisation of the scenario written in the "Direction" field – price reaches the recorded target before it reaches the protective order. That is not the same as a positive result: a position closed on a time condition at a small profit is not a hit, and a hit can be a negative result once costs are subtracted. The definition is a convention and you are free to choose a different one, on one condition: it must be fixed before the first entry and identical for all of them. Three possible readings – reaching the target, a positive result, realisation of the directional scenario – give three different frequencies, and mixing them invalidates the whole comparison.

After that it is mechanical: group the entries by declared band and compute the share of hits in each group. Illustrative example, illustrative values:

Post-run calibration check by declared belief level
Declared beliefNumber of entriesHitsShare of hitsReading
50–60%341956%consistent with the declaration
60–70%412561%consistent with the declaration
70–80%281657%declaration higher than the win rate
80–90%11655%a discrepancy large enough to be visible even at eleven entries

Reading this table calls for the same caution as the rest of the article. At thirty entries in a group the interval consistent with the data is more than thirty percentage points wide – that is, between the values the table in the second section gives for twenty and for fifty observations. This means such a summary detects only a gross discrepancy, for example a declared 85% against a real 55%. It will not detect a subtle one and there is no point pretending it does. The last row shows, however, that a small sample does not automatically invalidate the reading. Eleven entries give an interval consistent with the data fifty percentage points wide – from 28% to 79% – and yet that entire interval lies below the declared 80–90% band; the probability of six or fewer hits at eleven entries and a real win rate of 85% is 0.016. This is a reason to check calibration, rather than proof that this particular trader overstates confidence. The value 0.016 tests the specific 85% hypothesis, not the whole 80–90% band; selecting groups after seeing the data and comparing several bands also matter.

Procedure: recording belief before the decision.
Trigger: every order that opens a position. No entry means no order – the sequence is part of the procedure, not decoration on it.
Action: the five fields from the table above, in one place, before the order is placed.
Stopping criterion: the review takes place only once at least two belief bands have accumulated twenty-odd entries each. Before that no review is done, because there is nothing to count – and that is exactly the condition whose absence turns the procedure into a ritual.
Auditability: the time of the entry must be comparable with the time of the order in the account history. If that pair cannot be matched up, the record is not evidence, it is a recollection.
How you will know it is not working: after a month more than a fifth of orders have no entry – the form is too long or sits in the wrong place in the process. Or: all the entries carry the same belief band – the scale is measuring nothing and has to be narrowed to two levels. Or: after two reviews the gap between declaration and win rate does not move despite deliberate correction – then the problem is probably not calibration but the rule itself, and that is a question for section 8.

7. The illusion of control and the limits of what depends on the trader

In 1975 Ellen Langer described a series of experiments in which participants behaved as though choosing for themselves increased their chances in a purely random game[4]. In the most frequently cited of them – study two, N = 53, of whom 27 were in the self-choice condition and 26 were assigned a ticket – office workers on Long Island bought one-dollar tickets in a lottery with a pool of about fifty dollars: some could pick their own ticket, some were handed one. When they were later offered a buy-back, those who had chosen for themselves asked for an average of 8.67 dollars, and those with an assigned ticket for 1.96 dollars. The probability of winning was identical in both groups.

The evidentiary status of this study has to be reported alongside the result. It is a single experiment from half a century ago, on a small sample, with a trivial stake and in one task; the effect size computed from the reported t statistic is large (d ≈ 1.2), but one study with a large effect size and one replicated regularity are not the same thing. I have not established a registered replication of this particular experiment and say so plainly – the later literature on the illusion of control is uneven, and the effect shows up more clearly there in measures of declared confidence than in behaviour itself. The same rigour that the next section applies to the revision of the hot-hand research applies here too.

Chosen and assigned tickets. Mean asking resale prices in Langer (1975), study 2: USD 8.67 and USD 1.96. Winning probability was identical.
Langer (1975), study 2, N = 53. The asking-price difference concerned lottery tickets; it did not measure trader behaviour or results.

[EXTRAPOLATION] The study was run on office workers in the United States in the 1970s, with a trivial stake and in a lottery task. What I carry over from it is the mechanism – choosing for yourself raises the sense of influence over an outcome you do not control – not the effect size and not a claim about traders.

The trading counterpart is easy to recognise. Time spent on analysis alone does not give control over the subsequent price move. New, relevant information can nevertheless change trade selection or the estimated distribution of possible outcomes; that should be distinguished from a sense of control arising solely from involvement. Whether it also changes execution is a question your own record answers: count how many modifications you made to an open position and check whether that number rises with the time spent on analysis before entry. Five indicators measuring essentially the same thing give one confirmation, not five – and they look like five. Actions on an open position may increase a sense of agency; counting them alone, however, measures neither that feeling nor an improvement in the result.

[PRACTICE WITHOUT RESEARCH] I know of no measurement establishing, among retail traders, a relationship between time spent on analysis before entry and the number of later modifications to the position. The sentence above is a hypothesis testable in your own record, not a conclusion from Langer or from Barber and Odean – the first study measured a ticket buy-back price, the second measured turnover and net results.

The large dataset on real accounts that comes closest to this observation concerns a different market and measures something else – the relationship between turnover and net result, not the illusion of control. Brad Barber and Terrance Odean analysed the accounts of 66,465 households at a US discount brokerage, common stocks, 1991–1996. The highest-turnover households achieved about 11.4% a year net, the average household about 16.4%, the least active about 18.5%, and the market about 17.9%; the difference was for the most part due to costs[8].

[EXTRAPOLATION] This is the US equity market in the 1990s, investors rather than leveraged currency traders. What I carry over is the cost mechanism – trading more frequently weighs on the net result – not the effect size and not the conclusion that overtrading in itself destroys accounts. The relationship is cross-sectional: it describes the association between turnover and result, it does not prove that increasing turnover for a particular person will worsen that person's result.

Factors under the trader’s control versus outside the trader’s control
Under the trader's controlOutside the trader's control
Whether to take a trade at allThe outcome of a single trade
The entry condition written before the sessionThe order in which hits and losses fall
The size of risk fixed in the plan (parameters: section 9)The moment at which market conditions change
How and when the order is placedThe liquidity available at the moment the order fills
The number of decisions taken in a sessionWhether today turns out to be a typical day
Recording the decision before the outcome is knownWhat the trades he did not take would have looked like

The last row is the easiest one to trip over. Settling accounts over skipped trades – "if I had entered then, I would have had two hits" – uses information that only came into existence after the decision. That is not a tally of losses, it is a story built out of knowledge that was not there at the moment of the decision.

Procedure: a counter of actions on an open position.
Trigger: every position close. The entry is made together with the close, not in the weekly review, because after a week the number of modifications gets invented rather than reconstructed.
Action: add to the entry the number of actions performed on the already open position – moves of the protective order, changes of target, additions to the position – along with the time spent on analysis before entry, rounded to the nearest quarter of an hour.
Stopping criterion: review only after several dozen trades have accumulated. Before that the counter has nothing to show, and comparing a handful of entries gives a result that changes with every further trade.
Auditability: every action counted must be reconstructible from the order history in the account. A number taken from memory measures a recollection, not behaviour.
How you will know it is not working: the counter does not differentiate between trades – they all read zero or they all read the same. Then this measure says nothing about you and something else has to be measured, for instance the time between opening a position and the first action on it. A second form of not working: the number of actions rises, but only on days with clearly different volatility – then the market explains it, not a sense of influence, and adding a cognitive mechanism on top is premature.

8. Patterns in noise and simpler explanations

Cognitive shortcuts that work efficiently in most situations fail systematically in one particular environment: one in which feedback arrives rarely, with a delay, and full of noise. The market is exactly that kind of environment. That is enough to explain the mechanism and does not require adding a story about what those shortcuts supposedly served in the species' past.

How many hits a 50/50 coin gives in a run of 100 flips share of hits in a run of a hundred flips frequency of runs 35% 40% 45% 50% 55% 60% 65% 99% of runs fall between the lines (37–63% hits) "broken rule" "an edge"
The binomial distribution for p = 0.5 and a hundred trials, computed exactly. The central 99% of runs fall between 37% and 63% hits; results of 35% or less, or 65% or more, appear in about 0.35% of runs, that is roughly three to four times in a thousand runs. The extreme bins of the distribution are marked in red. Those extreme runs differ from the rest in nothing but the order of the draws – and they look like a broken rule or like an edge.

The same effect is visible on a price chart. A randomly generated path produces stretches that look like a trend, levels that look like support, and bounces that look like a reaction of the market – even though they arose purely from the order of the draws.

Two illustrative paths on identical axes: index 90–110 and observations 0–20. Similar shapes do not establish a repeatable mechanism.
Two explicitly illustrative paths on the same scale. Resemblance is not evidence of an edge; separate validation is required.

Human intuition about what randomness looks like is, moreover, systematically shifted in one direction. In 1972 Willem Wagenaar summarised the body of research on the generation of random sequences by people: asked to produce a "random" sequence, subjects switch the outcome to the opposite too often and produce runs shorter than genuine randomness delivers[3].

[EXTRAPOLATION] Wagenaar's review concerns the generation of sequences, that is what subjects produce themselves, not how they judge sequences shown to them from outside. All I carry over from it is the direction of the discrepancy – the mental picture of randomness has shorter runs than real randomness – and carrying that over to how losing streaks are perceived is an extension the review does not establish.

With that expectation of randomness, four or five consecutive losses may look like a breakdown, even though they can be highly probable under the model in section 3. The probability depends on the hit rate, the number of decisions and the length of the streak. Without those assumptions, the streak cannot be described as inevitable.

Two mirror-image false conclusions come out of that shift. The first: after a losing streak the next trade "must" be a hit. It must not; with independent trials the chance is the same as it was the first time. The second: after a winning streak you can raise the stake, because "it is running". Both rest on the same assumption, that a random process remembers its own past.

The history of research on the second of them is a separate lesson in caution towards "debunked myths": the classic 1985 debunking of the "hot hand"[6] was itself called into question in 2018 by an analysis of the method it used[7]. A full discussion of that revision, of both false conclusions and of the situations that are not instances of them, is in the article on cognitive traps.

A separate variety of pattern-fitting is the story built after the fact. Once a candle has closed the move looks obvious, and before it closed there were equally convincing arguments for the opposite scenario. The control is the record from the sixth section: the forecast and the degree of belief noted down before the outcome. The mechanism itself – hindsight certainty – has its own discussion in the article on cognitive traps, and its effect on judging trades in the article on outcome and process.

Not every explanation needs to be psychological. A loss described as “stop hunting” does not itself establish deliberate action by other participants. Orders may cluster near popular levels, while liquidity, volatility and execution can affect the outcome. Before attributing a cognitive bias, check simpler explanations: noise, costs and market conditions. If one of these is sufficient, a psychological label adds no evidence about the cause of that loss.

Procedure: the simpler-explanation test.
Trigger: every sentence in the journal that explains a loss by the behaviour of the market or by somebody else's intent – hunting for protective orders, aiming at a particular level, the market reacting to your position.
Action: write three simpler explanations next to it – noise, transaction cost, order execution conditions – and mark for each whether it suffices on its own to explain what happened.
Stopping criterion: the entry is closed when one of the three suffices, or when none of them does and that too has been written down. Recording the absence of a simpler explanation is a correct ending to the procedure, not a failure of it.
Auditability: at review time you can count what share of psychological explanations survived this test, because both columns – the original explanation and the simpler explanations – stay in the record side by side.
How you will know it is not working: the share is 100% or 0%. In the first case the test is a sham performed for form's sake, in the second you have stopped recording explanations at all, so there is nothing to test. In both cases the procedure has stopped settling anything and it is better to go back to simply noting the circumstances of the trade, without explaining it.

9. When uncertainty moves beyond the arithmetic

Every number in this article rests on one silent assumption: that the set of possible outcomes is known and probabilities can be attached to it. Knight's distinction from the first section is a reminder that this assumption is sometimes false – and then it is not the estimate that fails, but the model itself.

On 15 January 2015 the Swiss National Bank removed, without warning, the minimum exchange rate of 1.20 francs per euro that it had maintained since September 2011[11]. For more than three years EUR/CHF had moved in a narrow band just above that level. After the announcement the rate fell below parity within minutes and closed the day slightly above one franc. Protective orders had nothing to lean on, because on the way down there were practically no bids – and for the same reason that session has no single low. The EBS platform confirmed a session low of 0.8500 francs per euro in a note to clients; other venues recorded extremes of their own, at some retail providers markedly lower, and Pilbeam's account gives "the region of 0.80 francs per euro"[12]. The spread of those figures is itself information rather than a data defect: when there is nobody on the other side, a single market price simply does not exist. The British broker Alpari UK filed for insolvency proceedings four days later[13], and FXCM had to reach for a 300 million dollar rescue financing line from Leucadia National[14].

Timeline: the SNB introduced the CHF 1.20 per EUR minimum on 6 September 2011 and discontinued it on 15 January 2015. A schematic of multiple venues explains why execution depends on available quotes.
The historical SNB decision and an illustrative quote diagram. Different data sources and execution conditions do not establish one universal low or a guaranteed stop price.

The practical conclusion for this article is narrow and concerns only the status of the arithmetic. Slippage of that magnitude may not be adequately described by a distribution estimated from the previous three years, especially after a regime change. The size of the event alone cannot establish that. The losing-streak table does not predict such days and is not meant to. What to do with that knowledge from the point of view of position size and exposure is settled in section 9.

A second case illustrates a similar limit of the model. On 20 April 2020 the May WTI crude oil contract opened at USD 17.73 and settled at minus USD 37.63 per barrel[15]. A model restricting that contract’s price to non-negative values would exclude this outcome. This does not mean that nobody had considered the possibility. It shows why the model’s limits and the instrument’s specification need to be checked, rather than inferred from an intuitive floor for prices.

In market writing such events are described with the term "black swan", popularised by Nassim Taleb[9]. That is a concept from essayistic literature, not the result of an empirical study, and it should be read as such.

The opening question — whether the rule still works — requires an assessment of its basis and data from a longer series. One trade cannot settle it. Assess an individual decision using the information available at the time, the plan and the execution record; a pre-order record helps prevent the rationale from being fitted to the result. The confidence intervals, streak table and six paths show the scale of uncertainty. They do not replace validation of the rule or rule out a change of view when new, relevant information becomes available.

The arithmetic in the third section answers the question of how much of the unusual fits inside the normal behaviour of a working rule. It does not answer the question of what happens to a person who is watching a sixth loss in a row and has a seventh decision in front of them. That is a separate mechanism – the moment at which state begins to change the execution of a settled process – and it is dealt with in the article on emotions and impulsiveness.

FAQ – Trading under uncertainty

What is the difference between risk and uncertainty?
Risk is a situation in which the distribution of possible outcomes is known and can be computed with – as in a game with fixed rules. Uncertainty is a situation in which the distribution has to be estimated from data and its shape may change. The distinction comes from Frank Knight (1921). For a trader it means that he does not know his win rate the way a casino knows its edge: he knows only an estimate from a past sample, carrying measurement error and uncertainty about whether the conditions still hold.
How many trades does it take to judge whether a rule works?
The order of magnitude is hundreds, not dozens. At an observed 55% hits, the lower bound of the 95% interval consistent with the data crosses 50% only at 385 trades, and formally testing whether 55% differs from 50% requires a sample on the order of eight hundred trades. Method: the Wilson interval for a proportion assuming independent trials; the sample-size calculation itself, with its assumptions and variants, is handled in the article on judging the process. At somewhere over a dozen trades a month that means years, not weeks – and it is one of the reasons why judging the decision process is achievable sooner than judging results.
How can you tell whether a losing streak signals a breakdown rather than normal variance?
The streak’s length alone cannot settle the difference: with a 55% hit rate and 250 independent trades, the probability of at least six consecutive losses is 68.6%. A change may be justified by a previously recorded invalidation condition, such as the disappearance of the rule’s premise or a material change in costs. It may also be justified by new, verifiable information that the plan did not anticipate. Record what changed and why it challenges the rule, separating that reason from the loss itself. A streak can justify a data review; without that review, it proves neither a breakdown nor continued effectiveness.
Why write down the degree of belief before a trade?
Because without a record made before the decision there is no way to check calibration, and the recollection of one's own certainty is formed after the outcome is known and distorted by it. A minimal record suffices: the time, the condition judged to be met, the direction, the belief band and one sentence about what would argue against. Once between a dozen and twenty-odd entries have accumulated in at least two bands, you compare declared certainty with the actual share of hits. Such an analysis detects only gross discrepancies – at thirty observations the interval consistent with the data is more than thirty percentage points wide.
Can randomness be removed from trading?
No. The outcome of a single trade depends on factors the trader neither observes nor controls. What can be done is to move the assessment from a place where it cannot be settled to a place where it can: from the outcome of a single decision to the decision's compliance with a condition written before it, and to the calibration of declared belief. This is neither consolation nor a substitute for an edge – a rule with no justification after costs remains a rule with no justification, however carefully it is executed.

Sources and references

  1. Knight, F.H., Risk, Uncertainty, and Profit, Houghton Mifflin, 1921. The distinction between measurable risk and unmeasurable uncertainty.
  2. Tversky, A., Kahneman, D., "Belief in the law of small numbers", Psychological Bulletin, 76(2), 1971, pp. 105–110. A questionnaire among 84 professional psychologists; overrating the representativeness of small samples and expecting a result to repeat in the next small sample.
  3. Wagenaar, W.A., "Generation of random sequences by human subjects: A critical survey of literature", Psychological Bulletin, 77(1), 1972, pp. 65–72. A review of research on the generation of random sequences by people: switching the outcome too often and producing runs that are too short.
  4. Langer, E.J., "The Illusion of Control", Journal of Personality and Social Psychology, 32(2), 1975, pp. 311–328. A series of six experiments. Study two: N = 53 (27 people choosing a ticket themselves, 26 with an assigned ticket), office workers on Long Island, 1973, ticket price 1 dollar, pool about 50 dollars; buy-back prices of 8.67 against 1.96 dollars with an identical chance of winning (t = 4.33, p < 0.005). Scan of the original paper (hosted outside the publisher’s service; study two checked on pp. 316–317).
  5. Moore, D.A., Healy, P.J., "The trouble with overconfidence", Psychological Review, 115(2), 2008, pp. 502–517. A review and theoretical model with an original experiment involving 82 students and paid quizzes of varying difficulty; distinguishes three forms of overconfidence.
  6. Gilovich, T., Vallone, R., Tversky, A., "The hot hand in basketball: On the misperception of random sequences", Cognitive Psychology, 17(3), 1985, pp. 295–314. The classic analysis calling the "hot hand" into question.
  7. Miller, J.B., Sanjurjo, A., "Surprised by the hot hand fallacy? A truth in the law of small numbers", Econometrica, 86(6), 2018, pp. 2019–2047. A demonstration of bias in the method used by Gilovich and co-authors, plus a re-analysis of the same data. The reversal of the conclusion concerns the controlled experiment on 26 Cornell players (study four in Gilovich and co-authors), not the Philadelphia 76ers game data.
  8. Barber, B.M., Odean, T., "Trading Is Hazardous to Your Wealth: The Common Stock Investment Performance of Individual Investors", The Journal of Finance, 55(2), 2000, pp. 773–806. 66,465 households, a US discount brokerage, common stocks, 1991–1996; highest-turnover households about 11.4% a year net, the average household about 16.4%, the least active about 18.5%, the market about 17.9%; the difference mainly due to costs.
  9. Taleb, N.N., The Black Swan: The Impact of the Highly Improbable, Random House, 2007. Essayistic literature; the source of the concept, not the result of a study.
  10. European Securities and Markets Authority (ESMA), decision on a temporary product intervention concerning contracts for difference, May 2018, together with subsequent extensions and the national rules introducing the obligation to disclose the share of losing retail accounts.
  11. Swiss National Bank, press releases: "Swiss National Bank sets minimum exchange rate at CHF 1.20 per euro" of 6 September 2011 and "Swiss National Bank discontinues minimum exchange rate and lowers interest rate to −0.75%" of 15 January 2015. The primary source for the dates on which the minimum rate of 1.20 francs per euro was introduced and removed.
  12. Pilbeam, K., section "Intervention in the Foreign Exchange Market: Rationale, Effectiveness, Costs and Benefits" in: "Currency Interventions: Effective Policy Tool or Shortsighted Gamble?", Intereconomics, 50(2), 2015, pp. 64–81. A forum of six authors (Pilbeam, Bris, Alcidi, Barslund, de Groen, Gros); the page range given covers the whole forum, not Pilbeam's section alone. Description of the course of 15 January 2015: a temporary fall of the rate to the region of 0.80 francs per euro and stabilisation slightly above one franc. The reference value for that session's low – 0.8500 francs per euro – comes from the EBS platform's note to clients; quotes published by other providers differed from it, which is the very point at issue here.
  13. Financial Conduct Authority, notice to clients of Alpari (UK) Limited. The company formally entered insolvency proceedings under the Special Administration Regime on 19 January 2015; the administrators appointed were Richard Heis, Samantha Bewick and Mark Firmin of KPMG LLP.
  14. FXCM Inc., press release of 16 January 2015 and current report 8-K with event date 16 January 2015, filed on 23 January 2015 (SEC EDGAR, CIK 0001499912). Credit agreement with Leucadia National: a two-year term loan of 300 million dollars, concluded after the losses of 15 January; net proceeds were about 279 million dollars. The 8-K filed on 16 January 2015 concerns a separate matter – negative balances on client accounts.
  15. Commodity Futures Trading Commission, Interim Staff Report on Trading in NYMEX WTI Crude Oil Futures Contract Leading up to, on, and around April 20, 2020, November 2020. The May WTI contract fell on 20 April 2020 from 17.73 dollars and settled at −37.63 dollars a barrel – below zero for the first time in the history of that contract.
  16. NIST: Engineering Statistics Handbook: binomial proportions and confidence intervals.
  17. ESMA: Empirical dispersion of retail speculative outcomes under transaction costs.

The streak probabilities, the confidence intervals for the win rate and both charts in the text were computed independently: the streaks by a Markov chain with a state corresponding to the current length of the losing streak, checked by simulation over 200,000 repetitions with seed 20260810; the intervals by the Wilson method; the paths on the chart in the fourth section were generated with seeds 101–106. All values marked as illustrative serve to show the mechanism and do not come from any trader's account.

Your learning progress

Finished this lesson? Mark it yourself. Opening or scrolling a page never marks it complete.

Progress requires JavaScript and browser storage.

Saved only in this browser. No account or cross-device sync. Clearing site data can erase your progress.

Jarosław Wasiński

Editor-in-chief, ForexMechanics.com • Macro & FX analyst

Independent analyst and practitioner with over 20 years of experience in financial markets. Actively involved in Forex since 2007, with a focus on fundamental analysis, OTC market structure, and disciplined capital risk management.

  • Creator and editor-in-chief of MyBank.pl, providing reliable financial education since 2004.
  • Author of hundreds of in-depth market commentaries, structural analyses, and educational materials for traders and investors.
  • Advocate for market transparency, promoting education grounded in hard data and institutional reports (BIS, ESMA, CFTC).

All content is educational and informational only – it does not constitute investment advice or a recommendation. Remember: leveraged trading (Forex/CFD) carries a high risk of rapid capital loss.