Each row simulates: "if you took a trade every time the model output ≥ X% (or ≤ 1-X%), risking $Y per trade with win-R and loss-R as specified, what would your P&L be?" Higher thresholds = fewer trades, higher win rate but missing setups. Sweet spot is usually around 60-70% — enough confidence to expect a winning hit rate, but not so picky that you miss most signals.
If a threshold is unprofitable, the model isn't yet good enough at that confidence level — either needs more training data or that confidence range needs more samples to calibrate.