Stress the Costs and See Whether the Edge Survives

[Fairman]

Stress the Costs and See Whether the Edge Survives

A strategy has a gross expectancy of plus 0.15R per trade before costs. The costs in the test were 0.05R per trade: spread and commission. Test what happens if costs are higher.

Compute net expectancy at different cost levels. At 0.05R: 0.10R. At 0.10R, double the original: 0.05R. At 0.15R, triple: 0.00R. At 0.20R: minus 0.05R.

Interpret. The edge survives a doubling of costs, barely, and disappears at tripling. That is a thin margin of safety.

What could raise costs above the test? Wider spreads at the times you actually trade, slippage on stop orders, commissions that differ from the test, and swap on holds.

What could lower the edge? Out-of-sample decay, as seen with all strategies. If the true gross edge is 0.10R instead of 0.15, a cost of 0.10R wipes it out.

Put numbers on frequency. At 20 trades a month, net 0.10R yields 2R a month. If costs rise to 0.10R, net 0.05R yields 1R a month. At 1 percent risk, the monthly return falls from about 2 percent to 1 percent.

A strategy with a gross edge of 0.5R has more tolerance.

Always re-run backtests with costs at 150 percent of your estimate. If the result becomes negative, the strategy is fragile.

Your turn: compute the cost multiple at which your strategy's expectancy hits zero.

[FXS HongKong Scalper]

Doubling costs is a fair stress level, but for scalpers I would look at where the extra cost actually comes from. In our logs the average spread at a given hour is not the problem; the tail is. Most of the extra cost comes from a small share of trades taken in the first minute after data or at the daily rollover.

So besides a flat multiplier, try a second test: keep normal costs, then add 1 to 2 pips of slippage to the worst 10 percent of trades by time of day. With a 0.15R gross edge and a small stop, that alone can take expectancy close to zero. If it does, a time filter usually helps more than switching brokers.

[Amanhasnoname]

I ran this test on an M1 EURUSD system I was quite proud of, and it failed far earlier than I expected. The backtest assumed a fixed 0.2 pip spread plus commission, which came to about 0.06R on my small stop. When I replaced that with real spread data from my broker for the hours I actually trade, the cost was closer to 0.11R, almost double, and the net expectancy dropped to nearly nothing.

The second reply describes what I found too. The average spread wasn't the issue. It was a handful of trades taken right at the rollover and in the minute after data, where the spread was three or four times normal. Removing those two windows with a time filter recovered most of the edge without changing anything else in the strategy.

My own break-even multiple came out at roughly 1.8 times the original cost estimate. That's the number I now keep in my notes for each strategy, and anything below two gets traded at reduced size until I've seen at least three months of live fills.