I built a directional model as well that came to the same conclusion. I realized the move was looking at the structural reads, not necessarily at any directional forecast. I decided to stop reaching for direction and read the structure. Particularly where price sits in its long run range, as well as who is currently buying and selling. A couple of questions: have you checked whether the 50% range really holds about half the time? And, did the 90% range hold up on days with larger moves not just the quiet ones?