Information Sets and Learning Models for First Exits from Dynamic Price Zones: An Empirical Competing-Risk Study

We evaluate the information and model complexity needed to forecast first closing-price exits from pre-issued dynamic zones. Established competing-risk principles encode no exit within 60 min and first upper/lower exits in four intervals as nine outcomes. Annual historical analysis uses 502,443 DZ-LSTM-zone and 474,591 causal moving-quantile (MQ60) test origins for BTC, ETH, and XRP from 2020 through June 2026. Multinomial logit, XGBoost, survival forest, and an auxiliary MLP receive 22 non-zone predictors, 2 added distances, or all 32 inputs. Exact-calendar reconstruction repeats all 504 learner/information-set fits on unchanged historical targets and samples. Full inputs have lower Brier loss than non-zone inputs in 4/4 DZ-LSTM and 4/4 MQ60 learners, with signed relative reductions of 10.02–11.85% and 13.58–14.53%, respectively. Two distances recover most historical input gains. A locked follow-up applies stored 2026 MQ60 models and training scalers without tuning to 19,056 unused origins from 30 June through 30 September 2026; 4/4 learners have lower full-input Brier, and 4/4 reductions remain Holm-significant under all tested date-block choices. The contribution is empirical evidence about boundary information and learner choice. These retrospective checks do not establish a new output layer or universal neural superiority. Existing volatility benchmarks and negative external-market results delimit interpretation. Episode alerts remain optional illustrations under assumed policies and costs.