Overconfidence and the Illusion of Skill

Three things the word “overconfident” is doing

Overconfidence is not one bias but three, and they fail differently. Overestimation is believing your ability or your information is better than it is — the trader who thinks their read has an edge it does not have. Overplacement is believing you are better than other participants, which is measurable and absurd in aggregate: in survey after survey, well over half of drivers, students and professional investors rate themselves above the median of their own group, and about three-quarters of fund managers rate themselves above average at their job. In a market, where every trade has a counterparty who also thinks they are right, being above average cannot be the median outcome. The third form is the quietest and the most dangerous to a decision: overprecision. It is not that an estimate is too high, but that it is held too tightly — the confidence interval around it is too narrow. When people are asked for a range they are 90% sure contains the answer, the true value lands inside it a little over half the time; with 98% ranges, roughly six in ten. The number attached to the belief is systematically more precise than the belief deserves, and since position size is usually a function of that number, overprecision is overconfidence translated directly into risk. None of this means doubt is a virtue. Every position requires an estimate, and a trader who refuses to commit an estimate never acts. What separates confidence from overconfidence is not the size of the belief but whether the belief has been checked against a record. A confident trader can say what they expect, at what probability, and then find out whether they were right; an overconfident one cannot, because the probability was never written down. A “90% confidence” range, tested — Ranges stated at 90% confidence: 20 · Times the true value fell inside the range: 8 · Actual hit rate: 40% ← · Stated hit rate: 90% · Gap between stated and actual: 50 points ← The same person would have been right to describe the range as “a coin flip with a bias”. The gap between 90% and 40% is not a fact about the market; it is a fact about how the interval was built, and it is measurable in twenty questions if you write the number down before you find out.

What it costs, and what it does not

Overconfidence is not expensive because it produces one enormous mistake. It is expensive because it produces many ordinary ones, and the bills arrive as frequency. More trades means more spread paid, more slippage, more chance that an entry chosen for a reason becomes an exit chosen for a feeling, and — in the household study — a return about a third below the market the households were trading inside. The most active households were not less intelligent and did not use worse information; they acted on their information far more often than it justified. There is a second, larger cost that does not appear on a statement. A trader who believes their read is better than it is does not need a written plan, because the plan is the belief; they do not need position sizing, because conviction will decide the size; and they do not need a record, because the memory of the winners is evidence enough. Overconfidence therefore removes the instruments that would have caught it — the journal from a few lessons ahead is the thing an overconfident trader skips, and it is the thing that would have shown the confidence was unfounded. The countermeasures are the ones this subject keeps returning to, applied to belief rather than to feeling. Write the forecast down with a probability attached, then score it (which is calibration, and later lessons do the arithmetic). Start every new situation from the base rate for its class rather than from the specifics of this instance. Size positions from the plan rather than from the strength of the view, because the view is the variable that has already been shown to be miscalibrated. And keep the record, because a genuinely skilled trader is one whose written forecasts are right more often than the base rate — and that is a claim a record can settle and an argument cannot. • Write the forecast with a probability, on a date, and score it later. • Start from the base rate for the class of situation, not from the particulars. • Let the plan set the size; conviction is the miscalibrated input. • Count trades, because the cost of overconfidence arrives as frequency. A rising account is the least reliable evidence available: a bull market makes every method look skilful, and the first years of a record are dominated by it. The test of skill is not the return but the return relative to what the same exposures would have earned without the decisions.

Where overconfidence concentrates, and how to measure yours

Overconfidence is not spread evenly across people or across tasks; it clusters where two conditions hold. The first is that the task is **judged to be a matter of skill** rather than chance. When a decision feels like it depends on your ability — picking a stock, timing an entry — people report far higher confidence than when the same decision is presented as a lottery, even though the underlying uncertainty is identical. Trading is presented to its participants as a skill pursuit in every advertisement, every chart, and every course, which is exactly the framing that maximises the effect. The second condition is **familiarity**. Confidence rises with the number of times you have done something, whether or not accuracy has risen with it — the well-documented result that knowing a little about a topic produces more confidence than knowing nothing and less accuracy than knowing a lot. The most-cited evidence in trading comes from the accounts themselves. In the large retail study that underpins this lesson, the most active traders earned materially less than the market while the least active did better, and a companion analysis found that **men traded more than women and earned less as a result** — the gap in returns was almost entirely accounted for by the gap in trading volume. That is the shape overconfidence takes in a portfolio: not a wrong opinion, but a frequency, with the cost paid in turnover rather than in belief. Nobody blows up because they were confident about a direction; they erode because they acted on confidence more often than the situation warranted. Measuring your own is possible, and it takes one column in the journal. For each trade, record the **expected** return and the **range** you thought plausible at entry; later, compare what you wrote with what happened. Two numbers emerge: how often outcomes fell inside your range (usually far less often than the range implied) and whether your expected returns were systematically above your realised ones. Neither number requires a large sample to be informative, because both are about the calibration of your statements rather than about your returns, and unlike returns they are not swamped by market noise. That is the practical answer to “am I overconfident?” — it is not a question you answer by feel, and it is not a question you can answer from a profit and loss statement. • Overconfidence is strongest when a task is framed as skill rather than chance — which is how trading is framed. • Familiarity raises confidence faster than it raises accuracy. • Its portfolio footprint is turnover: the most active retail traders earned less, and the activity gap explains the return gap. • One journal column — expected return and plausible range at entry — measures calibration directly and needs no large sample.

When thousands try, some look skilled: the false-discovery problem in performance

The most active households underperformed by a wide margin, and the fund industry’s record is the same result at a larger scale with a different explanation available. If a thousand managers each run an independent strategy with no skill at all, the distribution of their three-year returns will still have a top decile: roughly five percent of them, or fifty managers, will sit more than two standard deviations above the mean purely by chance. Those are the managers with the marketing decks, the conference invitations and the rolling three-year track records. Nothing about their evidence is fabricated, and none of it is informative, because the sample being reported was selected by its outcome — which is the same selection the previous lessons describe in published research and in a trader’s own memory, applied to people rather than to studies. The test that distinguishes skill from that selection is not a good track record, it is **persistence**: does the performance of past winners predict the performance of future winners? The classic studies of mutual-fund performance found that it largely does not, once the size, style and expense differences are controlled for — the winners of one period are close to a random draw from the universe in the next, with the small exception of the very worst performers, who tend to persist in being bad, largely because high costs are a reliable drag. When persistence is measured, the number of managers who look genuinely and durably skilled is very small, and the fraction of those who are simply the tail of a large distribution is what remains. A related piece of arithmetic is worth carrying: if you screen a universe of ten thousand strategies on a single metric with a p-value of 0.05, roughly five hundred will pass by construction, and the honest question about any one of them is how many others were tried. What follows for a learner is not cynicism but a specific discipline about evidence. A track record is evidence in proportion to three things: the **length** relative to the strategy’s volatility, which is the ratio t-statistic from the risk lessons; the **number of attempts** behind it, because a manager chosen after a search is a maximum of estimates; and the **mechanism**, because a plausible reason for an edge that can be described before the returns are examined is worth more than the returns alone. The same three tests apply to your own record. A year of good results in a market that rose is one observation; five years across a full cycle with a stated mechanism is a body of evidence; and a good quarter after a change to the process is neither, because the process was selected by that quarter. The error overconfidence makes is taking the third as the first, and it is the same error whether the performance belongs to a fund or to a personal account. Three tests for a track record — Length against volatility: A t-statistic — the Sharpe ratio × √years from the risk lesson ← · Predictive persistence: Do the winners of one period predict the winners of the next? Usually not · Stated mechanism: A reason for the edge that could be described before seeing the returns ← · Number of attempts: A manager selected after a search is the maximum of that search The uncomfortable corollary for a personal account: the same arithmetic applies to your own history. One good year is a sample of one market, and the trader who attributes it to skill has made the industry’s error at retail scale — which is precisely the gap the personal-finance lessons quantify as the behaviour cost.

Sources

• Household trading and the cost of activityBarber & Odean (2000), “Trading Is Hazardous to Your Wealth”
• Do investors trade too much?Odean (1999), “Do Investors Trade Too Much?”, American Economic Review
• Calibration of subjective probability intervalsAlpert & Raiffa (1969); Russo & Schoemaker (1992), “Managing Overconfidence”