Improving the robustness of binary classifiers in cryptocurrency exchange rate forecasts

Abstract Binary direction classifiers can appear stable on one chronological split while offering little predictive or economic value out of sample. We revisit this problem using three relative-value cryptocurrency price ratios constructed from 28,157,374 public Binance USDS-M futures trades observed from March to May 2025. The original experiment combined event-time bars, LightGBM, a trend-continuation rule, and a rolling Hurst-regime condition. The revision retains the original fixed-split results as a descriptive record and adds an independent audit with three expanding walk-forward folds, four model benchmarks, rule-only controls, component ablations, three random seeds, uncertainty intervals, signal coverage, and 0/10/20-basis-point cost scenarios. The original table showed a smaller validation-to-test precision gap in seven of nine cells, but the exact sign and Wilcoxon tests were not significant at 5%. In the independent audit, LightGBM alone obtained mean precision 0.511, recall 0.213, F1 0.251, and 20.8% positive-signal coverage. The full logical AND-gate obtained precision 0.473, recall 0.014, F1 0.026, and 1.34% coverage. Its unlevered accepted-signal return averaged -0.16 basis points before costs and -10.16 basis points after a 10-basis-point round-trip cost. These results do not establish a general or cost-robust trading edge. They show instead why generalization gap, discrimination, coverage, and economic value must be reported separately when a regime gate suppresses predictions.