Conditional Predictive Gains from Multiscale Networks for a Chinese Equity Opening

Chinese equity indices open after foreign markets have already supplied part of the day’s information. We ask whether multiscale dependence networks improve opening forecasts conditionally on states observable before the mainland opening. Daily prices from 2010 to July 2026 support 2081 evaluation dates per index. The design distinguishes incomplete local records, stale news, index launch availability, and accumulated foreign changes during domestic closures. A fixed LightGBM model adds 57 fluctuation order dependent DCCA features to the same feasible Raw information set. Six economic state partitions, matched benchmarks, and chronological rule assessment distinguish a useful representation from unrestricted model superiority. The clearest additional evidence concerns Shanghai’s absolute prediction error. On 1036 mornings with above median foreign news disagreement, mean absolute error falls from 25.90 to 24.83 basis points, a 4.12 percent reduction; the pointwise 95 percent block interval for the 1.07-basis-point gain is [0.64, 1.52]. This secondary absolute-loss comparison has Holm adjusted p=0.0094 across 468 comparisons and p=0.0296 in a wider 1482 test audit. Its later period estimate remains positive but does not pass the corresponding multiplicity adjustment. High-minus-low gain differences also do not establish unique state dependence. Squared-error and extreme probability comparisons do not show general network dominance, and forward applied gates do not deliver an adjusted significant advantage. The evidence supports a loss- and benchmark specific conditional increment, rather than an identified multifractal mechanism or a universally preferable risk model.