Two commenters told me my trend bot was overtrading. They were right, and it didn't save the strategy
retained text
Follow-up to a backtest I posted last week: trend bot, about 565 trades a week, paying roughly 11% of the account in fees every week. Two people in the comments said the same thing, that the 3 hour average was flipping inside the noise and I should make the signal persist for a few closes before acting on it. So I ran it.
Setup, so you can tell me where it's wrong. 15 minute bars, 12 coins, 6 weeks to 14 Sep, paper. Entry is a close above the N-bar average with the average rising, exit is a close back below. The change is that the close now has to happen k times in a row before anything fires. k=1 is the original rule, which is how I check the baseline still reproduces the trade counts I published (it does, within 4 trades in one week out of six). Costs 0.1% a side plus 0.05% spread.
Per week, averaged over the six:
rule trades/wk fees/wk fees % acct return/wk 6wk total return/wk at zero cost
3h avg, act on first close 565 $1,129 11.3% -15.63% -64.6% -4.34%
3h avg, hold 2 closes 343 $685 6.9% -8.10% -40.8% -1.24%
3h avg, hold 3 closes 255 $510 5.1% -5.13% -28.1% -0.02%
3h avg, hold 4 closes 214 $427 4.3% -4.61% -25.7% -0.34%
6h avg, first close 362 $723 7.2% -8.73% -43.5% -1.50%
12h avg, first close 225 $450 4.5% -3.98% -23.3% +0.51%
6h avg and hold 2 closes 230 $459 4.6% -4.41% -25.1% +0.18%
BTC over the same six weeks: +2.7%, -3.4%, +23.4%, +0.2%, +3.3%, -4.3%.
They were right about the turnover. Holding for 2 closes cuts trades 39%, 3 closes cuts 55%, and fees go from 11.3% of the account a week to about 5%. Halving the cost roughly thirds the loss, which is a bigger effect than I expected going in.
Something I didn't expect at all: hold-3 on the 3 hour average and a plain 12 hour average end up within 30 trades a week and about a point of return of each other. So waiting for confirmation and just using a longer lookback seem to buy more or less the same thing, at least here.
The last column is what killed it. That's return with the fees added back, so what each rule would have done trading free. It's approximate, fees compound with everything else so it flatters the losing weeks a bit, though nowhere near enough to matter: the best of them makes about half a percent a week, over six weeks where BTC did 23% in one of them.
So the filter worked and there was nothing under it.
Which would you test first, a volatility gate so it only fires when range is expanding, or a minimum hold time? Or is 15 minute bars across 12 coins just the wrong place to be looking for trend at all.document fetched 2026-09-26 00:43 · canonical https://reddit.com/r/algotrading/comments/1wpw54x/two_commenters_told_me_my_trend_bot_was