Re: WHAT LOOKS LIKE EDGE BUT ISN'T
Posted: Thu Nov 05, 2026 10:42 am
Fake edge being storytelling with a chart is the line I'd pin. The filter questions are what I run too, and the second one, whether the parameters were stable before the curve looked pretty, catches the most. I tested a moving average cross last year and found a 20 period setting with a profit factor of 1.6. Moving it to 19 or 21 dropped it under 1.1. That wasn't edge, it was a lucky parameter.
The range-weeks test is a good one. A method that only appears in trend screenshots is really a trend detector, and trends are visible in hindsight on any chart. If it makes 4R in a trending fortnight and loses 3R through two choppy ones, the total is +1R across a month, and with spread at 0.6 pips on every entry I'd expect that to be a negative result in practice.
Freezing the rules for a fixed sample is where I struggle. I want to tweak after ten trades. Setting a rule like "no changes until 100 trades" would stop me, and I'd put the date in the journal so I can't pretend it was shorter.
The point on fat left tails is easy to miss as well. A 75 percent win rate with a loss six times the size of the average win has an expectancy of roughly -0.5R if that loss comes once in four trades.
Do you pick the freeze sample size by setup frequency, or a flat number? And what do you do when the results look best right at the end of the freeze?
The range-weeks test is a good one. A method that only appears in trend screenshots is really a trend detector, and trends are visible in hindsight on any chart. If it makes 4R in a trending fortnight and loses 3R through two choppy ones, the total is +1R across a month, and with spread at 0.6 pips on every entry I'd expect that to be a negative result in practice.
Freezing the rules for a fixed sample is where I struggle. I want to tweak after ten trades. Setting a rule like "no changes until 100 trades" would stop me, and I'd put the date in the journal so I can't pretend it was shorter.
The point on fat left tails is easy to miss as well. A 75 percent win rate with a loss six times the size of the average win has an expectancy of roughly -0.5R if that loss comes once in four trades.
Do you pick the freeze sample size by setup frequency, or a flat number? And what do you do when the results look best right at the end of the freeze?