MetricUpdated: 2026-08-28
What is P90 relative error?
P90 relative error means about 90% of the comparable errors in the backtest set were at or below this level.
Plain meaning
Relative error = |predicted lots ÷ official lots − 1|. If the official result is 10 lots and the prediction is 12 lots, relative error is 20%. P90 is the level that about nine-tenths of those errors do not exceed. It is counted across comparable application tiers, not once per IPO.
Why it matters
Median error describes a typical case. P90 describes harder cases. Averages or medians alone can hide the tail.
How to read it
Lower is better, but read it with sample size and market regime. It is not the maximum error and not a promise for a future IPO.
How it appears here
The allotment prediction page shows time-weighted relative error and P90 together so overall level and higher-error cases can be compared.
Common mistakes
- Treating P90 as the maximum error.
- Assuming it is calculated once per IPO.
- Assuming historical P90 guarantees future single-IPO accuracy.