Overconfidence Bias
Overconfidence Bias: People consistently think they know more than they do. When asked questions with 90% confidence intervals, they're right only 50β70% of the time. 80% of drivers think they're above-average. Expert forecasters are frequently more confident in their predictions than their track records warrant. Overconfidence is one of the most expensive cognitive biases in finance, medicine, and strategic planning.
What Is Overconfidence Bias?β
Overconfidence is one of the most replicated findings in psychology. In the most common test, people are asked questions and provide a confidence interval wide enough that they believe there's a 90% chance it contains the true answer. Across many studies, people are correct only 50β70% of the time β they're overconfident by 20β40 percentage points.
The bias has three distinct manifestations:
- Calibration overconfidence: confidence intervals too narrow (intervals that should contain the answer 90% of the time contain it 60%)
- Better-than-average effect: 80%+ of people rate themselves above average on desirable traits
- Illusion of control: overestimating the degree to which outcomes are under personal control
How It Worksβ
The calibration test (most common measure):
β Question: "What is the length of the Nile in km?"
β Give a range you're 90% confident contains the answer
β Calibrated person: range is wide enough to be right 90% of the time
β Overconfident person: range is too narrow; right only ~60% of the time
The better-than-average effect:
β 93% of drivers rate themselves as above-average drivers
β 90% of university professors rate their teaching as above average
β These distributions are mathematically impossible
Expert overconfidence:
β Philip Tetlock's 20-year forecasting study: expert forecasters with
high confidence were wrong more often than their confidence implied
β Doctors, lawyers, financial analysts all show overconfidence
in their specific domains
Three Real-World Examplesβ
Financial Markets and Overconfident Tradingβ
Terrance Odean's analysis of 78,000 brokerage accounts found that the more active traders (who traded because of high conviction) earned lower returns than less-active traders. The active traders were overconfident in their analysis β they believed they had an edge when they didn't. Each trade has transaction costs; overconfident trading generates costs without generating returns. The US mutual fund industry's persistent underperformance versus index funds is partly attributable to overconfident active management.
Medical Diagnostic Confidenceβ
Studies of physician diagnostic confidence find systematic overconfidence: physicians who say they are "certain" of a diagnosis are wrong approximately 40% of the time. Post-mortem studies comparing final diagnoses to actual causes of death find significant discordance even in cases where physicians expressed high confidence. Overconfidence in diagnosis leads to under-testing (no need to confirm what I already know), which contributes to diagnostic error.
Expert Forecasting (Philip Tetlock)β
Philip Tetlock's 20-year study of expert forecasters found that prediction accuracy correlated negatively with fame and confidence. "Foxes" (those who gathered many perspectives) outperformed "hedgehogs" (those with confident grand theories) β and both barely outperformed baseline random. The most confident experts showed the worst calibration. Tetlock's findings suggest that domain expertise produces overconfidence without proportionate accuracy gains in complex domains.
When to Recognise Itβ
π¨ Overconfidence is likely operating when:
- You're providing narrow ranges for uncertain quantities
- You haven't sought disconfirming evidence for a confident belief
- Your past prediction accuracy is lower than your confidence levels suggested
- You're certain about outcomes that are genuinely uncertain
β Countermeasures:
- Track your predictions and compare to outcomes
- Deliberately widen your confidence intervals
- Seek disconfirming evidence before acting on confident beliefs
- Reference class forecasting for project estimates
- Ask "what would change my mind?" β if nothing would, that's overconfidence
| Pairs well with | Why |
|---|---|
| Dunning-Kruger Effect | Dunning-Kruger is one mechanism of overconfidence |
| Confirmation Bias | Confirmation bias feeds overconfidence by filtering disconfirming evidence |
| Planning Fallacy | Planning fallacy is overconfidence applied to project outcomes |
Common Misuses and Limitationsβ
Conflating overconfidence with all confidence. Some domains with excellent feedback loops (chess, short-term weather forecasting) produce well-calibrated confidence. Overconfidence is most pronounced in complex, low-feedback domains (long-term forecasting, medical diagnosis, financial markets) where feedback is delayed or ambiguous.
Treating underconfidence as preferable. The goal is calibration β being as confident as the evidence warrants. Underconfidence is also a calibration failure. Research shows that novices are often underconfident, while experts become overconfident. The ideal is the well-calibrated middle: confidence that accurately tracks accuracy.
Related Modelsβ
| Model | Relationship |
|---|---|
| Dunning-Kruger Effect | One specific mechanism of overconfidence |
| Confirmation Bias | Feeds overconfidence by filtering out disconfirmation |
| Planning Fallacy | Planning fallacy is overconfidence applied to future projects |
Frequently Asked Questionsβ
How can you measure and improve your own calibration?
Track your predictions: for every confident claim, record it and its confidence level (70%, 90%, etc.), then review outcomes. Well-calibrated: 70% confident claims should be right 70% of the time; 90% claims right 90% of the time. Tools like Superforecasting's Metaculus platform provide structured calibration training. Research shows calibration can be meaningfully improved with practice and feedback β it's a trainable skill.
Are some people well-calibrated naturally?
Yes. Philip Tetlock's superforecasters β the top 2% of his 20,000-person forecasting study β showed excellent calibration: their 80% confident predictions were right approximately 80% of the time. The differentiating features: they actively sought disconfirming evidence, updated beliefs when new evidence arrived, expressed uncertainty in numerical terms rather than verbal hedges, and had extensive track records of feedback. Calibration is learnable, not purely innate.
Does overconfidence have any benefits?
Potentially in some social contexts. Research by Dominic Johnson suggests that some overconfidence may have evolutionary value as a commitment device: appearing confident induces others to cooperate, even if the confidence is not fully warranted. Entrepreneurs with overconfident beliefs about their ventures may persist through early difficulties that would rationally cause others to quit β occasionally succeeding as a result. But in analytical domains (medicine, finance, strategy), overconfidence consistently reduces decision quality.
Further Readingβ
- Lichtenstein, S., Fischhoff, B. & Phillips, L. (1982). "Calibration of Probabilities." Judgment Under Uncertainty (Kahneman et al.)
- Tetlock, P. (2005). Expert Political Judgment β the definitive study of forecasting accuracy
- Tetlock, P. & Gardner, D. (2015). Superforecasting β how to improve calibration
Apply with AIβ
π Test and improve your confidence calibration with MindMax β
This page is part of the MindMax Mental Models Knowledge Base.