Predictions are only meaningful once a calibration has been fitted against real history and checked on held-out bars. A directional accuracy near 50% is the expected null result, not a bug.