Every training run recorded in the registry, what it was testing and how it scored. For breakout and breakdown signals the number that matters is precision: of the signals fired, how many became a real move.
Precision is the share of fired signals that became a real move. Lift compares that to how often the move happens anyway — 1.0× means the model adds nothing over guessing. Fires on is the share of ticker-days flagged. Walk-forward repeats the measurement across separate historical periods, and is the number to trust when a single holdout looks lucky.
Accuracy is deliberately not shown: a +10% week happens about 5% of the time, so a model that never fires scores 95% accuracy while being useless.