CI90 coverage is how often the real outcome lands inside the model’s 90% confidence band. The honest target is 90% — a 90% band should contain the truth about 9 times out of 10. It measures whether the uncertainty band is the right width, not whether the forecast pointed the right way.
Coverage tells you whether the width of the uncertainty band is honest — not whether the forecast pointed the right way. A model can be perfectly calibrated and still be a coin flip on direction: if it says the price will be somewhere in a wide range and the price stays in that range, coverage is fine even though the call gave you no directional edge. How often the model calls up-vs-down is a separate axis (directional hit rate). Never read good coverage as "the predictions are accurate."
On Quantustik’s committed TOP-20 backtest (the checked-in record behind the public calibration page, measured against the model as served, with the market and sector coupling attached), how much the bands covered depends heavily on when the forecast started. Across nine quarterly start dates from 31 Mar 2024 to 31 Mar 2026, the 90% bands covered 78–95% of realized prices at 3 months, 71–92% at 6 months and 86–96% at 1 year, against a 90% target. We publish every start date rather than one average of them, because the windows overlap at 6 months and 1 year and an average would claim a precision the measurement does not have. Individual windows are worse still: UNH started 30 Jun 2025 covered 10% at 3 months and NVDA started 31 Dec 2024 covered 2% at 1 year, while other windows sit near 99%. Check the per-ticker backtest accuracy for a name you actually trade, not just the average.
Coverage well below 90% means the model is over-confident — bands too narrow, reality escapes them more often than promised. Coverage well above 90% is also a failure: the bands are so wide they are uninformative, technically "always right" but useless for a decision. Honest calibration lands near 90%, not as high as possible.
No. Coverage only tells you whether the 90% band is the right width — whether reality lands inside it about 90% of the time. A model can be perfectly calibrated on coverage and still be no better than a coin flip on which direction the price moves. Directional accuracy is a separate metric.
On the committed TOP-20 backtest behind the calibration page, measured against the model as served, coverage is published per forecast start date rather than as one average. Across nine quarterly start dates from 31 Mar 2024 to 31 Mar 2026, the 90% bands covered 78–95% of realized prices at 3 months, 71–92% at 6 months and 86–96% at 1 year, against a 90% target. Individual windows are worse still, so no single figure is a guarantee for any one name or any one entry date.
No. Coverage far below 90% means the bands are too narrow (over-confident); coverage far above 90% means they are so wide they carry no information. The honest goal is to sit near 90%.
Browse all S&P 500 tickers to see this metric applied to individual companies.
Educational research only — not investment advice.