Conformal Prediction Intervals for Portfolio Sizing
New research integrates conformal prediction with fractional Kelly sizing to dynamically scale positions based on forecast uncertainty.
Conformal prediction intervals can serve as a dynamic scale for fractional Kelly position sizing. By using a 75% conformal interval to determine the range of a forecast, the strategy shrinks positions as the interval widens and grows them as it narrows. This approach treats uncertainty not just as a descriptive tool for a forecast, but as a direct input for capital allocation.
In a development window from 2016 to 2021, the method achieved 28.5% annualized net log growth with a 1.34 Sharpe ratio, outperforming a passive S&P 500 hold of 15.9%. The strategy beat the textbook standard deviation by 2.1 points at matched leverage. These results suggest that conformal intervals provide a more effective proxy for risk in position sizing than traditional volatility measures.
Stability in the interval width is more valuable for sizing than local sharpness. Attempts to adapt intervals faster to market conditions reduced annual growth by 0.7 to 5.3 points. The most effective configuration relied on slow, unweighted, per-asset rolling quantiles. This indicates that for capital allocation, a steady estimate of uncertainty is superior to a highly reactive one.
Risk control can be further tightened by monitoring the historical miss rate of these intervals. When downside misses exceed the historical rate, cutting leverage reduced the maximum drawdown from 27.7% to 20.3% in the development window. This mechanism creates a systematic circuit breaker based on the failure of the prediction model itself.
Despite the strong development results, the strategy struggled in a pre-registered evaluation window from 2022 onward. Annual growth dropped to 7.0% and 8.5%, falling below passive benchmarks. While the calibration of the intervals remained stable, the growth did not translate, highlighting the difficulty of maintaining an edge across different market regimes.
This work shifts the focus for quantitative managers from the accuracy of a point forecast to the reliability of the uncertainty bound. It suggests that the primary lever for improving risk-adjusted returns may be the method of scaling positions rather than the underlying signal. The failure in the 2022 window underscores the risk of over-optimizing configurations via autonomous search.