Why Forecasts Fail
6 steps · one page
In short
The strongest evidence on this question does not come from finance. It comes from a study designed to score predictions properly, which is the thing published financial forecasts almost never permit.
Scope. This article reports what has been measured about forecasting accuracy, including the part that cuts against the popular summary. The evidence does not show that prediction is impossible — it shows something more specific and more useful. MarketClue publishes no forecasts, targets or projections of any kind. No forecaster, institution or publication is named beyond the researchers whose published work is cited. Findings are reported with their populations and periods stated, verified 17 August 2026.
The foundational study
Between 1984 and 2003, Philip Tetlock collected predictions from 284 people who made their living commenting on or advising about political and economic trends — academics, policy analysts and journalists. He gathered roughly 28,000 predictions, recorded each forecaster's stated confidence, and waited.
| Finding | Result |
|---|---|
| Average accuracy | Barely better than random chance |
| Tetlock's own comparison | A dart-throwing chimpanzee would have done about as well |
| The most confident forecasters | Performed worst of all |
| Forecasters with one large organising theory | Did worse than those who held several partial ones |
The finding that matters most is not the average, and it is routinely dropped from summaries. The forecasters who performed worst were the ones who appeared on television, wrote opinion pieces and possessed a single large explanatory theory. Confidence and public prominence were associated with lower accuracy, not higher. Which connects directly to the argument in Reading Financial Media Critically: coverage is selected on being interesting, and a confident unqualified forecast is more interesting than a hedged one. So the selection mechanism that determines which forecasts a reader encounters is positively correlated with the trait associated with being wrong. The reader is not receiving a random sample of forecasts. They are receiving one filtered toward the characteristic the evidence identifies as a warning sign.
Why most financial forecasts cannot be scored at all
Scoring a prediction requires three things, and published financial forecasts frequently supply none of them.
A probability. "Likely" cannot be scored; "65%" can.
A resolution criterion. A statement of what observation would count as the prediction having come true, and what would count as it having failed.
A deadline. Without one, a prediction is never wrong — only not yet right.
A forecast missing any of the three is not a claim about the world in the sense set out in the previous article — and it is worth noticing that this is not primarily a failure of rigour. An unscoreable forecast is a professionally safer product than a scoreable one, and the incentive to produce it is straightforward.
The part that cuts the other way
The popular summary of this literature is that nobody can predict anything. That is not what the evidence shows, and reporting it that way would be as inaccurate as the confident forecasting it criticises.
Tetlock and colleagues later ran a large forecasting tournament, funded by a United States intelligence research agency, recruiting volunteers to forecast world events. A subset of those volunteers — the group that became known as superforecasters — outperformed intelligence community analysts by around 30%, without security clearances, classified data or institutional resources.
And the performance was persistent rather than lucky. Roughly 70% of superforecasters retained the status from one year to the next, and across all forecasters the correlation between one year's performance and the next was 0.65 — which is high, particularly for online volunteers whose engagement varied.
Worked example
What distinguished them, reported as findings rather than as instructions. They integrated base rates with case-specific information rather than reasoning from the case alone. They updated frequently in response to evidence — a great deal, though not without limit. Answers that explicitly used comparison classes scored materially better than the next-best category of reasoning. And they were less prone to standard cognitive biases than ordinary participants, though not immune: some still exhibited them, just less often. Also measured: forecasters who inclined toward the view that outcomes were meant to happen were significantly less accurate. This portal reports these as observed correlates of accuracy in a tournament setting. It does not present them as a method, because the conditions that produced them are specific and are set out below.
Why that does not transfer straightforwardly to investing
The tournament conditions are the reason the result exists, and investing lacks all three of them.
The questions resolved. Every forecast had a defined outcome and a date, so accuracy was computable.
Feedback was scored and returned. Participants learned their calibration, repeatedly, against a numerical measure.
The questions were selected to be answerable. They were hard but resolvable, which is not the same as the open-ended questions investors actually face.
An investor deciding what to hold receives none of that. Outcomes are confounded by everything else happening simultaneously, there is no scoring, and — as Overtrading computes — the sample of decisions is far too small to separate skill from chance in the timeframes people actually operate over. So the honest position is that forecasting skill demonstrably exists, is trainable under measurement, and the conditions under which it was demonstrated are largely absent from private investing.
Frequently asked
8 questions
What did the foundational study find?
Between 1984 and 2003, Tetlock collected roughly 28,000 predictions from 284 professional commentators and advisers. Average accuracy was barely better than random chance — Tetlock's own comparison was a dart-throwing chimpanzee.
What is the most important part of that finding?
Not the average. The forecasters who performed worst were the ones who appeared on television, wrote opinion pieces and held a single large explanatory theory. Confidence and prominence were associated with lower accuracy.
Why does that matter for a reader?
Because coverage is selected on being interesting, and a confident unqualified forecast is more interesting than a hedged one. So the mechanism determining which forecasts a reader encounters is positively correlated with the trait associated with being wrong.
Why can't most financial forecasts be scored?
Scoring needs a probability, a resolution criterion and a deadline. Published forecasts frequently supply none. Without a deadline a prediction is never wrong, only not yet right — and an unscoreable forecast is a professionally safer product than a scoreable one.
So is forecasting impossible?
No, and saying so would be as inaccurate as the confident forecasting this article criticises. In a large tournament, a subset of volunteers outperformed intelligence community analysts by around 30% without clearances or classified data.
Was that luck?
The evidence says not. About 70% of superforecasters retained the status year to year, and across all forecasters the correlation between consecutive years' performance was 0.65 — high, particularly for online volunteers with varying engagement.
What distinguished them?
Integrating base rates with case-specific information rather than reasoning from the case alone; frequent updating on evidence; explicit use of comparison classes, which scored materially better than the next-best reasoning category; and lower susceptibility to standard biases, though not immunity. Forecasters inclined to think outcomes were meant to happen were significantly less accurate.
Does that transfer to investing?
Not straightforwardly. The tournament had resolving questions, scored feedback and questions selected to be answerable. An investor gets none of those — outcomes are confounded, there is no scoring, and the sample of decisions is far too small to separate skill from chance. Forecasting skill exists and is trainable under measurement; the conditions that demonstrated it are largely absent from private investing.
References
- Tetlock (2005; new edition 2017) — Expert Political Judgment: How Good Is It? How Can We Know?, Princeton University Press (the 284-forecaster study) —
- Mellers et al. (2014) — Psychological Strategies for Winning a Geopolitical Forecasting Tournament, Psychological Science 25(5) (the tournament findings and correlates of accuracy) —
- Mellers et al. (2015) — Identifying and Cultivating Superforecasters as a Method of Improving Probabilistic Predictions, Perspectives on Psychological Science 10(3) (persistence of performance) —
- Tetlock and Gardner (2015) — Superforecasting: The Art and Science of Prediction —
Educational and informational only — not investment advice, a recommendation, or an offer to buy or sell any security. Investing involves risk, including the possible loss of principal. Worked examples use fictional companies and figures.