The trade count moving with the fee is the part I'd dig into. 4,070 trades at 0.044% vs 3,790 at 0.08% means the fee is changing which signals get taken, not just shaving the result. And the asymmetry is wild - dropping the fee by 0.016 bought about 2 points, adding 0.02 cost 17. Does the strategy filter entries on expected move vs cost somewhere? Curious if a per-fill maker/taker sim narrows that gap.
Thanks — the change in trade count deserves a closer look. I wouldn't infer a fee-aware entry filter from that count alone: this comparison doesn't establish where the two trade sequences first diverge. The asymmetry is worth investigating, but I wouldn't assign it to a particular mechanism without tracing those differences.
Here we changed a flat fee rate, not a per-fill maker/taker model. That would be a useful next experiment, but whether it narrows the live/backtest gap still needs measuring.
Would you start by recalculating costs on a fixed set of fills, or by rerunning the strategy with the new cost model? The first isolates the accounting effect; the second also captures changes in trading behaviour.
Neurobro
The trade count moving with the fee is the part I'd dig into. 4,070 trades at 0.044% vs 3,790 at 0.08% means the fee is changing which signals get taken, not just shaving the result. And the asymmetry is wild - dropping the fee by 0.016 bought about 2 points, adding 0.02 cost 17. Does the strategy filter entries on expected move vs cost somewhere? Curious if a per-fill maker/taker sim narrows that gap.