Tools
Forecast accuracy calculator
How good were your forecasts? Enter what you forecast and what actually happened — numbers such as sales or demand, or chances of yes/no events such as a prediction market price — and get the standard accuracy measures, with what each one means.
Forecast accuracy calculator
- Brier score
- 0.140 (0 is perfect; always saying 50% scores 0.250)
- Skill against a coin flip
- 43.8% better than always saying 50%
- Log loss
- 0.440 (lower is better; 0.693 for always 50%)
The formula
error e = actual − forecast · MAE = mean |e| · RMSE = √mean(e²) · MAPE = mean |e ÷ actual| · bias = mean(e) · Brier = mean (p − outcome)²
How to calculate forecast accuracy for numbers
A forecast error is the difference between what happened and what was forecast: actual minus forecast. Every accuracy measure for number forecasts summarises those errors in a different way.
MAE (mean absolute error) is the average size of a miss, in the data’s own units: an MAE of 8 on monthly sales means the forecasts were 8 units off on average. RMSE (root mean squared error) squares the errors before averaging, so one big miss counts for more than several small ones. MAPE (mean absolute percentage error) divides each error by the actual value, which makes forecasts of different-sized things comparable: a MAPE of 6% means the forecasts were 6% off on average.
MAPE has known weaknesses. It is undefined when any actual value is zero, gets extreme when actuals are close to zero, and penalises negative errors more heavily than positive ones. “Forecast accuracy = 1 − MAPE” is a common shorthand in business reporting, not a separate measure: 94% accuracy is just a MAPE of 6%.
Forecast bias
Accuracy measures say how big the misses are; bias says whether they lean one way. The simplest measure is the mean error — the average of actual minus forecast, with signs kept. Zero means misses on both sides cancel out; a negative mean error means forecasts ran high on average, and a positive one that they ran low.
A forecast can be accurate on average and still biased, and the fix for bias is simple: if errors average m, adding m to future forecasts removes it. That is why bias is tracked separately from MAPE.
Accuracy for probability forecasts: the Brier score
A forecast that says “70% chance” is never simply right or wrong, so it is scored with a proper scoring rule. The Brier score is the mean of (forecast probability − outcome)², with the outcome 1 if the event happened and 0 if not. It runs from 0 (perfect) to 1, and forecasting 50% on every question scores exactly 0.25 — the benchmark a useful forecaster has to beat.
Check which version a published score uses. Glenn Brier’s original 1950 formula sums the error over both outcomes, so it runs from 0 to 2 and the always-50% forecaster scores 0.5; the Good Judgment Project reported scores on that scale. This calculator uses the 0–1 form that Polymarket and Kalshi report.
Log loss is the other common rule: the average of −ln(the probability given to what actually happened). It punishes confident misses far harder than the Brier score — a forecast of 0% for something that happens scores infinitely badly.
Calibration
A forecaster is well calibrated when events they call 70% happen about 70% of the time. A reliability diagram plots how often events actually happened against the forecast probability; a calibrated forecaster’s points sit on the diagonal. Calibration alone is not enough — always forecasting the long-run average is calibrated but useless — so the aim is forecasts that are as sharp as possible while staying calibrated.
How accurate are prediction markets?
The longest record is the Iowa Electronic Markets. Across 964 polls in five US presidential elections from 1988 to 2004, the market was closer to the final vote share than the poll 74% of the time, and it beat the polls in every election when forecasting more than 100 days out (Berg, Nelson & Rietz, 2008).
Markets are not equally good everywhere. Studies find a favourite–longshot bias: cheap contracts on unlikely outcomes are overpriced, so longshots win less often than their prices suggest. In InTrade data, prices were reasonably well calibrated close to expiry but biased for events further away (Page & Clemen, 2013); a working paper on Kalshi contracts to April 2025 finds low-price contracts “win far less often than required to break even after fees” (Bürgi, Deng & Whelan).
The platforms publish their own figures. Polymarket’s accuracy page gives a Brier score of 0.0595, without saying at what horizon, and notes that 78% of its markets resolved No, which makes low scores easier. Kalshi’s August 2026 paper reports its Brier score falling from about 0.087 three months before close to below 0.02 at close, across 2.2 million markets. Manifold’s calibration page shows 0.174, measured on trades rather than over time. These are self-reported, use different samples and methods, and cannot be compared with each other or with the figures above.
- Berg, Nelson & Rietz (2008), “Prediction market accuracy in the long run”, International Journal of Forecasting 24(2)
- Page & Clemen (2013), “Do prediction markets produce well-calibrated probability forecasts?”, Economic Journal 123(568)
- Bürgi, Deng & Whelan, “Makers and Takers: The Economics of the Kalshi Prediction Market” (working paper)
- Polymarket, accuracy page (read 30 Oct 2026)
- Kalshi Research, “Calibration in Prediction Markets: Theory and Evidence” (Aug 2026)
- Manifold, calibration page (read 30 Oct 2026)
Frequently asked questions
- How do you calculate forecast accuracy?
For number forecasts, work out each error as actual minus forecast, then average them: the mean absolute percentage error (MAPE) is the average of |error ÷ actual|, and “forecast accuracy” is often quoted as 1 − MAPE. For probability forecasts, use the Brier score: the average of (forecast probability − outcome)².
- What is the forecast accuracy formula?
The most quoted is MAPE = average of |actual − forecast| ÷ actual, with accuracy = 1 − MAPE. MAE (average absolute error) and RMSE (square root of the average squared error) measure misses in the data’s own units instead, and work when actual values are zero.
- What is forecast bias?
A tendency for forecasts to miss in one direction. It is measured by the mean error — the average of actual minus forecast with signs kept. A negative mean error means forecasts ran high; a positive one, low.
- What is a good Brier score?
On the usual 0–1 scale, 0 is perfect and forecasting 50% every time scores 0.25, so anything below 0.25 beats a coin flip. What counts as good depends on how hard the questions are: scores are only comparable on the same set of questions.
- How accurate are prediction markets?
In the longest-running study, the Iowa Electronic Markets were closer to the final result than polls 74% of the time across five presidential elections. Accuracy improves as the event approaches, and studies find longshots tend to be overpriced.
Not financial advice. These calculators do arithmetic on the numbers you enter; they do not know a platform’s current prices or whether it is open to you. Check both in the platform reviews.