New KQSP method fixes K-line forecast crossings without retraining

A team of researchers (Yu, Tao, Chen, Wang, and Bunn) have introduced KQSP (K-line–Quantile Sequential Projection), a novel method that resolves two persistent consistency problems in probabilistic K-line forecasting: quantile crossing (when a higher-quantile forecast falls below a lower-quantile forecast) and K-line crossing (when the forecast high/low violates the open/close relationship). The method is parameter-free and training-free, meaning it can be applied as a post-processing step to forecasts generated by any model—including deep learning systems and pretrained foundation models—without modifying the model or retraining.

Existing solutions typically address only one type of crossing and rely on output reordering, specialized architectures, or penalized training objectives, each with drawbacks in accuracy or generality. KQSP, by contrast, operates sequentially on quantiles and K-line components to enforce consistency while making the smallest possible corrections to the original forecasts. In experiments across diverse forecasting models, KQSP reduced both quantile and K-line crossing rates to zero on all test datasets, without degrading predictive performance. This suggests that consistency can be enforced independently of the forecast generation process, opening the door to simpler, more reliable financial forecasting pipelines.