Forward Validation of a Frozen Bitcoin Price-Barrier Prediction Model: Why 75% Accuracy Is Not Predictive Edge

I built a machine-learning pipeline that predicts whether Bitcoin will move 2% up or 2% down within 24 hours, using separate upside and downside models trained on 34,311 time-series rows (2021 to 2024). I froze both models, locked the boundary between historical and forward data, and evaluated them on unseen forward data without refitting or recalibration. On 1,076 matured forward predictions, accuracy at a 0.5 threshold was about 74.9%, but overall ROC AUC was about 0.483, which is no better than chance (upside about 0.553, downside about 0.399). The case shows that headline accuracy can look strong while a model has no demonstrated ability to separate outcomes, and that freezing, integrity controls and plain reporting make that conclusion trustworthy. I report the result as a negative finding. Code for the monitoring component: https://github.com/jinguism-arch/btc-model-sql-monitoring