Predictive Extrema, Unprofitable Policies: An AI-Assisted Audit of Candle-Based Binance Spot Timing Models

Battle Count: 70

This paper is directly relevant to quantitative trading as it rigorously tests the prediction-to-policy translation gap in cryptocurrency trading. It demonstrates that high event-ranking performance (AUC up to 0.974) does not guarantee positive net policy returns after transaction costs. The paper provides detailed experimental contracts, cost assumptions, execution semantics, adaptive search budgets, and forensic audit methodology that are highly applicable to any quantitative trading strategy evaluation. The NO_TRADE conclusion across all tested policies serves as a cautionary framework for strategy promotion. The predict-then-optimize failure mode, coverage-destroying-sparse-edge insight, and concentration-sensitive positive control are directly actionable for quant practitioners.

The paper implements five distinct experimental campaigns with family-specific simulators, each requiring different execution semantics (next-open entry, stop-first tie rules, multiplicative vs. flat cost application, sleeve-based portfolio management). The mandatory daily selector uses ExtraTrees regression; local extrema campaigns test 14 architectures across 3 windows and 14 thresholds; the paired daily adaptation uses expanding CV folds with threshold calibration; the rotation control compares 240 policies. The forensic audit pipeline, fail-closed artifact reconciliation, and AI-assisted evidence synthesis add further complexity. However, the core ML models are standard tabular/sequence classifiers, and the simulators are deterministic scripts rather than production trading systems.

The paper explicitly acknowledges reproducibility gaps: raw candles and large checkpoints are excluded from the arXiv bundle; the companion GitHub repository is mutable with no archival DOI; the One4All campaign's raw result directories are absent; earlier rotation plot inputs are missing; the exact AI service build was not recorded. The audit script (audit_empirical_evidence.py) has no numeric fallbacks and reconciles counts, but full end-to-end reconstruction requires a tagged immutable repository, environment lock, public-download manifests, configuration hashes, and all cited result summaries. Fixed-seed runs and deterministic simulators aid partial reproducibility.