Definition
The median deviation of the predicted oversubscription multiple from its realized value, measuring the model's overall accuracy in estimating retail subscription demand. The median is more robust than the mean against outlier predictions.
Methodology
For each stock compute relative error = |predicted oversubscription − actual oversubscription| ÷ actual, then take the median across all stocks. Oversubscription is anchored walk-forward; regression coefficients are fitted on the full history.
Current Value
76.4%
Sample Size
214
Methodology
This metric comes from a walk-forward backtest on stocks listed in 2025 and later — the anchors (market regime, index, sponsor heat, first-day P&L) use only data visible before each stock listed; the regression coefficients are fitted on the full history, matching production prediction.

Data as of 2026-09-22

View Full Backtest Back to Metrics

About this page

This page documents the definition, methodology and current value of the prediction metric "Oversubscription Median Relative Error". The value is aggregated from the prediction_history table and refreshed by the backfill script, for objectively assessing model accuracy — not investment advice.

Cibo — HK IPO Allotment Forecasts, Allottee Profiling & Stock Connect