Adding filters whenever a strategy loses money makes 89% of backtests “better”—but it’s a trap.
A lot of people have an instinctive reaction when tuning a strategy: if it loses money, add restrictions, like “don’t short when BTC is up on the daily chart” or “pause opening positions when the broader market surges.”
I systematically tested 13 signals against various BTC and whole-market filters, running a total of 28,000 configurations.
At first glance, the results look very tempting: after adding filters, total returns improved in 89% of configurations.
But when you break down the numbers, the truth is glaring:
- The average number of trades fell by about 25% (to 75% of the original)
- Average profit and loss per trade barely changed
Put simply, the filters didn’t help you pick better trades. If a strategy has negative expected value per trade, doing 25% fewer trades naturally means losing 25% less—but fundamentally, it’s still losing.
I also specifically tested “stop shorting during sharp BTC rallies and stop going long during sharp BTC drops” over 1/3/5/30-minute periods: none of the 8 versions passed. They blocked only 0.5%–2.3% of signals, essentially just fitting noise.
In another test, I used 17 BTC and market-state indicators to predict “whether it’s better to go long or short next.” In 68 tests, not one widened the long-short profit-and-loss gap enough to cover costs.
Now, when I judge whether a filter is actually any good, I look at just one number: did the filter significantly improve the “average net return per trade”?
If it only reduces total losses by “trading less,” that’s not optimization—it’s covering up the fact that the signal doesn’t work.
What’s the most complex filter you’ve added, and did you later check whether average profit and loss per trade actually improved?
#合约交易 #量化交易 $BTC
Not investment advice.
A lot of people have an instinctive reaction when tuning a strategy: if it loses money, add restrictions, like “don’t short when BTC is up on the daily chart” or “pause opening positions when the broader market surges.”
I systematically tested 13 signals against various BTC and whole-market filters, running a total of 28,000 configurations.
At first glance, the results look very tempting: after adding filters, total returns improved in 89% of configurations.
But when you break down the numbers, the truth is glaring:
- The average number of trades fell by about 25% (to 75% of the original)
- Average profit and loss per trade barely changed
Put simply, the filters didn’t help you pick better trades. If a strategy has negative expected value per trade, doing 25% fewer trades naturally means losing 25% less—but fundamentally, it’s still losing.
I also specifically tested “stop shorting during sharp BTC rallies and stop going long during sharp BTC drops” over 1/3/5/30-minute periods: none of the 8 versions passed. They blocked only 0.5%–2.3% of signals, essentially just fitting noise.
In another test, I used 17 BTC and market-state indicators to predict “whether it’s better to go long or short next.” In 68 tests, not one widened the long-short profit-and-loss gap enough to cover costs.
Now, when I judge whether a filter is actually any good, I look at just one number: did the filter significantly improve the “average net return per trade”?
If it only reduces total losses by “trading less,” that’s not optimization—it’s covering up the fact that the signal doesn’t work.
What’s the most complex filter you’ve added, and did you later check whether average profit and loss per trade actually improved?
#合约交易 #量化交易 $BTC
Not investment advice.