Online Factor Correction with Delayed Conformal Feedback

Fixed financial factors can lose predictive value as market conditions change, motivating online adjustment of their influence on stock selection. Prediction intervals offer information for this adjustment, but coverage alone specifies neither an allocation rule nor its effect on investment performance. Delayed returns further complicate the link between calibration and decisions. We propose a factor-correction layer that connects these steps through an executable prediction target: each factor's standalone strategy return after transaction costs. A forecast of this target rewards expected payoff, while an online prediction radius discounts uncertainty before factor scores are combined. Three interleaved feedback sequences accommodate the return delay; a signed universal-portfolio update yields explicit coverage bounds that retain the calibration state inherited from warmup. The contribution is a specified connection between delayed calibration and factor allocation, evaluated with matched controls that distinguish forecasting gains from interval effects. On 157 public price and volume factors in Chinese and US equity universes, the correction retains positive net-return estimates relative to fixed coefficients and reduces the mean loss on the worst 5% of days relative to forecast-only allocation in both markets. Return intervals include zero, so these findings support an observed return–risk tradeoff rather than a confirmed return gain. The framework makes the financial role and statistical scope of online uncertainty explicit.