Calibration

Investing glossary · Models and probabilities

The match between the probabilities a model states and how often those events actually happen.

Definition

A forecast is calibrated when its probabilities match reality: of all the times it says "70%", the event should happen about 70% of the time. Calibration is different from accuracy. A model could be right often but overconfident, or humble and well-calibrated.

You test it by collecting many forecasts, grouping them by stated probability and checking the share that came true. Calibration is what separates a probability from a guess, and the reason a published track record matters: without checking, a "90% range" is only a label.

Worked example

Illustrative numbers

Suppose a model makes 1,000 forecasts, each stating a 70% chance. If it is well calibrated, about 700 of them should come true. Seeing between 640 and 760 would be consistent with chance; far fewer would mean the model is overconfident.

This example is illustrative: the same test applied to ranges checks whether a 90% range contains the outcome nine times out of ten.

How ATLAS uses it

ATLAS calibrates its probabilities and ranges against past forecasts, and publishes a track record so anyone can check whether the stated numbers held up.

Related in ATLAS: Track record →Browse all ATLAS tools →

Frequently asked questions

What is the difference between calibration and accuracy?

Accuracy asks whether the forecast was right. Calibration asks whether the stated confidence was right on average. A model can be accurate and badly calibrated.

How can I check a forecaster's calibration?

Gather their past forecasts with stated probabilities, group them (for example 60-70%) and compare each group's hit rate with its stated probability.

Related terms

Put the idea to work on your own stocks

Open any company in ATLAS to see its fair value, its odds of rising and its risk numbers side by side.

Try ATLAS →

Not investment advice. ATLAS estimates are produced by statistical models from public data; they can be wrong and past performance does not predict future results. Consider your own situation before investing. Disclaimer

ATLAS Signals

ATLAS Signals, every week

The week in the market, three stocks on ATLAS’s radar and what comes next. Get it every Monday.

By subscribing you agree to receive ATLAS Signals by email every Monday. You can unsubscribe with one click at any time. Privacy

Calibration: definition, formula and live example | ATLAS