The 50% confidence interval (CI50) is the forecast's tighter, coin-flip range. CI50 coverage measures how often the realised price actually landed inside it. A well-calibrated CI50 should contain the true price about half the time — no more, no less.
A forecast that gives you a single price is hiding something. Our model produces a distribution of possible outcomes, and from it we publish confidence intervals — ranges with a stated probability attached. The 50% interval is the narrow one: the model is claiming the price has a 50/50 chance of landing inside it.
Coverage is how we check whether that claim was true. Run the forecast across many stocks and many dates, then count what fraction of the time the realised price actually fell inside the band. If the model said 50% and reality delivered roughly 50%, the band is honest. That is CI50 coverage.
Around 50% is the target, and — this is the part people find counter-intuitive — being far ABOVE 50% is also a failure, not a triumph. Read it like this:
Well below 50% means the band is too narrow. The model is over-confident: it is telling you it has pinned the price down more tightly than it really has, and reality keeps escaping the range. This is the dangerous direction, because it invites you to size positions as though you knew more than you did.
Well above 50% means the band is too wide. The model is under-confident: the band is technically "right" more often than advertised, but only because it is so generous it barely rules anything out. A band that contains the price 90% of the time while claiming 50% is not a better forecast — it is a vaguer one wearing a confident label.
Both directions are miscalibration. The goal is not high coverage; the goal is coverage that MATCHES the number on the tin.
The 90% band and the 50% band test different parts of the same forecast distribution. CI90 coverage checks the tails — whether the model correctly anticipates how bad the bad cases get. CI50 coverage checks the middle — whether the model has the everyday, typical range right.
A model can pass one and fail the other, and that failure is informative rather than embarrassing. A model whose 90% band covers correctly but whose 50% band is far too wide is telling you something specific: it has the tail risk about right, but it is mushy about the central case. Publishing both is what lets you see that, and it is why our calibration page carries both columns instead of the flattering one.
Position sizing depends entirely on band honesty. If you size a trade assuming a 50% chance the price stays in a range, and that band is really only right 30% of the time, you have systematically taken more risk than you thought — on every trade, in the same direction. That kind of error does not average out. It compounds.
This is also why coverage figures are worth more to you than a headline accuracy claim. "Our model is highly accurate" is a sentence with no fixed meaning — accurate at what, over what horizon, measured how? "Our 50% band contained the realised price this often, at this horizon, on this committed backtest" is a claim that can be checked and can fail. Our numbers, including the horizons and tickers where they are poor, are on the calibration page.
Coverage is an average across many tickers, and averages hide dispersion. A model can post respectable aggregate CI50 coverage while being badly miscalibrated on individual names — typically the ones that went through a genuine structural break the model had no way to anticipate. This is exactly why our calibration page publishes a per-ticker breakdown and not only the aggregate: the aggregate is the flattering view.
Coverage also cannot tell you the model will keep being calibrated. It is measured on historical data, and if the band multipliers were tuned on that same data, good coverage is partly a result of the tuning rather than evidence of a well-understood market — see overfitting. Coverage on data the model never saw is the evidence that counts.
Finally, coverage says nothing about direction. A band can contain the realised price exactly as often as advertised while the forecast's central estimate points the wrong way. Whether the model got the direction right is a separate question, measured separately by the directional hit rate.
How often the realised price actually landed inside the model's 50% confidence band. A well-calibrated CI50 should contain the true price about half the time.
No. The target is about 50%, and being far above it is a failure too — it means the band is too wide to be useful, even though it is technically "right" more often. Good calibration means coverage that matches the advertised probability, not coverage that is as high as possible.
The band is too narrow and the model is over-confident — it claims to have pinned the price down more tightly than it really has. This is the dangerous direction, because it leads you to take more risk than you think you are taking.
They test different parts of the same forecast. CI90 coverage checks the tails — whether the model understands how bad the bad cases get. CI50 coverage checks the middle — whether it has the everyday range right. A model can pass one and fail the other.
Browse all S&P 500 tickers to see this metric applied to individual companies.
Educational research only — not investment advice.