MetricUpdated: 2026-07-09
What is P90 relative error?
P90 relative error means about 90% of historical backtest samples had error at or below this level.
Plain meaning
It is the 90th percentile of the error distribution and helps describe model performance in harder, higher-error cases.
Why it matters
Average error can hide tail behavior. P90 helps show how large errors get for most non-extreme samples.
How to read it
Lower is better, but read it with sample size, sample period and market regime. It is not a promise that future error will stay below the value.
How it appears here
On the allotment prediction page, P90 relative error appears in backtest metrics to show a higher-error historical level.
Common mistakes
- Treating P90 as the maximum error.
- Ignoring instability when sample size is small.
- Assuming historical P90 guarantees future single-IPO accuracy.