baseedge_labs
← research

Searching 7,500 configurations for a prop-evaluation base, and what a high win rate actually costs

Sep 3, 2026 · BaseEdge Labs research

Tokyo Drift

The goal was a configuration suited to a funded-account evaluation rather than to maximum profit: high win rate, a stable reward:risk, drawdown small enough to size contracts against, and consistency across the year. Roughly 7,500 configurations were swept across entry modes, filter stacks, bracket geometry, sessions, budgets, anchors and the reversal engine. Every run is on Micro Gold futures (MGC). Results are hypothetical backtests.

Scoring treats the account the way an evaluation does: a $100,000 account, a $6,000 profit target and a $3,000 drawdown limit. Contracts are sized to the limit, which makes drawdown the binding constraint rather than a tiebreaker.

Method

What we found

  1. Win rate is bought with reward:risk. Across every bracket cell in round one, the two moved in opposition. The strongest edge in the sweep sat at the lowest win rate, and the weakest at the top of the win-rate range. Every high-win-rate configuration is the same shape: a tight target against a wide stop, which wins often and loses rarely but enormously. That is arithmetic, not a strategy property. The target and stop set the hit rate, and the edge is roughly what it is either way.
  2. The high-win-rate configurations all failed out of sample. All fifteen lost money out of sample, and their win rates fell well below training. Their drawdowns grew sharply, which is the more dangerous failure. A contract count sized on the backtest would have badly understated real risk. The MGC-tuned default fails the same test. Its simulated pass rate went from most attempts in training to almost none out of sample.
  3. Very little survived all four periods. Only a small handful of the 7,500 finished net positive in every period. Every one uses the Pullback (Counter-Candle) entry with the reversal engine on. The structure is more consistent than any parameter in it.
  4. Ranking survivors by their floor profit factor misleads. The survivor with the top floor, f111111 rev1 tp1.0/sl1.75, barely makes money. The shipped configuration is a different survivor, f110110 rev1 tp3.25/sl2.0. It is the Pullback (Counter-Candle) entry with RSI, MACD-velocity, extension and minimum-VWAP-distance filters, the reversal engine on, and a 3.25 ATR target against a 2.0 ATR stop. Its floor is lower, but its win rate was the most stable across periods in the entire sweep, and it nets far more over the sample.
  5. Even the shipped configuration earns most of its profit in the trailing year. It was net positive in all four periods, but only marginally in 2024 H1, 2024 H2 and 2025 H1. Nearly all of its net profit came from the trailing year.
  6. Contract size is not a free lever. In the pass-rate simulation, every added contract lowered the pass rate in both windows, and the out-of-sample pass rates sat below the trailing-year ones. Past three contracts the pass rate keeps falling while the time saved flattens. Nothing in the round-one search supported four or more contracts.
  7. Time to pass depends on the regime, not the configuration. Out of sample the edge is thin, so it accumulates the profit target slowly. At one contract, the simulated time to pass ran to years on 2024–25 data and months on the trailing year. The same configuration at the same size gave two very different answers.
  8. The win-rate trade-off was an artifact of excluding one entry. On the RSI Reset entry, a high win rate and a strong profit factor were available together. RSI Reset, a1.75/sl5.0 and RSI Reset, a1.0/sl4.0 both reached high win rates. Their profit factors held well above the best non-RSI configuration at a comparable win rate. The tight-target, wide-stop shape is only a trap without the RSI filter stack.
  9. % Price brackets cut drawdown sharply. At neutral untuned brackets, ATR 2.0 / 2.0 against % Price 0.15 / 0.15, win rates were about the same. ATR earned more per contract at a higher profit factor. But % Price had a much smaller drawdown, which allowed more contracts under the limit and a better result for the account. The likely mechanism is how each stop reacts to volatility. An ATR stop widens exactly when volatility spikes, which is when losses cluster, so its losses carry a fat tail. A percentage-of-price stop stays proportional to the instrument regardless of recent range. Tuned, % Price produced the only configurations in either sweep that could carry four contracts within the limit.
  10. Breakeven and trailing stops are null on median. Breakeven alone and trailing alone each beat no stop-move on fewer than half the base configurations, with no median lift. This matches eleven earlier research stages that scored profit per trade. The result holds under a drawdown objective too. Combining breakeven and trailing was worse than either alone.
  11. There is one narrow exception. A low trigger with a wide distance, on a wide-stop configuration, gave a modest real improvement. High-trigger trails never fire when the target sits 1 ATR away. Their results are identical to no stop-move at all.
  12. The reversal engine depends on the entry. In the round-two matrix it lost on the large majority of on/off pairings, at a large median cost. That is the opposite of the Pullback configurations above, where every four-period survivor had it on. It interacts with the entry rather than being independently good.
  13. Round two produced two configurations, and the balanced one is better. Both use RSI Reset with RSI, MACD-velocity, extension and candle-range filters and the reversal engine off. The balanced one adds a trail that arms at 0.5 ATR and follows 2.0 ATR behind. The high-win-rate one wins more often. The balanced one has a higher profit factor, a better reward:risk, a smaller drawdown and a shallower worst month, and it reaches the target sooner. Both had the same share of months positive and support the same contract count. The extra win rate is bought with a low reward:risk, which makes the average loss much larger than the average win. That is what ends an evaluation.
  14. Both sit on a plateau, not a spike. Many neighbouring cells cleared the same bar: 11 of 12 months positive, two or more contracts within the limit, and a profit factor at or above 1.2.

The limit worth stating plainly

Finding 7 is the regime dependence documented in the regime-versus-fit study, restated in evaluation terms. This configuration passes quickly if conditions resembling 2026 persist. It does not pass at all if 2024–25 conditions return. Nothing in this sweep removes that, and no amount of further tuning on the trailing year can.

The round-two configurations were selected on the trailing 365 days alone. They carry no out-of-sample validation.