Throwing machine learning at the order book to predict the next move

We trained directional models on the live HYPE order book across 8 time-horizons (5 min to 1 day) and 7 trigger thresholds, with the pass bars locked before touching the data. Not one of the 56 combinations cleared both bars. Short horizons had thousands of cases but the typical move (~31 bps) sat under the ~33 bps you need to cover a 10 bps round-trip; long horizons paid 90 to 180 bps but collapsed to 5 to 40 test cases. A cost-versus-payoff geometry problem, not a tuning problem.

Pre-registered with two bars locked pre-data: model break-even probability <= 0.65 AND >= 500 held-out cases (data kept aside from training, so the model is judged only on moves it never saw). Closest misses: the 15 min horizon fell 7 cases short of the floor (493), and 5 min came in at 0.659, over the 0.65 cap by 0.009. Five tempting rescues (loosen the cap to 0.70, drop the floor to 480, re-split the data, add more features, lock the near-miss anyway) were all rejected as moving the goalposts after seeing the data. Honest read: at HYPE's scale a 10 bps round-trip eats most of the short-horizon move, and there aren't enough days to evaluate the long horizons where the math would work. Killed at the first feature set, not forever.