Crossing-Free Probabilistic K-Line Forecasts Without Retraining
创建于 更新于
摘要
This work introduces K-line–Quantile Sequential Projection (KQSP), a parameter-free, training-free reconciliation method that enforces both quantile monotonicity and OHLC consistency for probabilistic K-line forecasts (OHLC). KQSP operates in two sequential stages (K-line projection followed by quantile projection) and is model-agnostic, requiring no retraining or extra hyperparameters. Empirical results across an ANN and eight zero-shot foundation models show KQSP reduces both quantile crossing and K-line crossing to zero without degrading predictive accuracy, and often improves metrics such as Average Quantile Loss (AQL), MAE, and RMSE. Comparisons with multiple baselines indicate KQSP achieves equivalent or better correction performance while applying smaller adjustments to the original forecasts. The method demonstrates robustness across architectures and forecast sources, highlighting that probabilistic K-line consistency can be achieved independently of the underlying forecast model [page::1][page::2][page::3].
