Published
Last post I pulled the 50 rows my pipeline flags as top candidates out of 12,331 backtests and checked each one's breakeven leverage. This time I compared all 50 against each other, brute force, looking for the two with the closest expectancy per trade. The winner: a MEGAUSDT row and an ARIAUSDT row, both with exactly 1,600 trades, expectancy 0.28871864...% versus 0.28869824...% -- a difference of 0.00002 percentage points, closer than any other pair out of the 2,450 I checked. Net result over the full run: $461.95 versus $461.92. Two cents apart.
Everything else about these two rows is a different world. MEGAUSDT closes that result at a 95.4% win rate, averaging 9.3 minutes a trade. ARIAUSDT gets to basically the same two dollars at a 55.2% win rate, averaging 41.6 minutes a trade -- 4.5x longer to pay out the same money. My scoring pipeline ranked them 70 points apart (120.9 vs 50.8) mostly on profit factor and win rate. It never once asked how long the capital sat there to earn it.
The two rows, side by side
| Row | Trades | Net | Win rate | Avg hold | Return per hour held |
|---|---|---|---|---|---|
| MEGAUSDT | 1,600 | $461.95 | 95.4% | 9.3 min | 1.85%/hr |
| ARIAUSDT | 1,600 | $461.92 | 55.2% | 41.6 min | 0.42%/hr |
"Return per hour held" here is just expectancy per trade divided by average hold time, scaled to an hour. It's not a compounding forecast -- I'm not claiming you can chain either config back to back all day with no slippage, no funding, no capital limit. It's a measure of how efficiently each config used the time its position was actually open, nothing more. By that measure MEGAUSDT is 4.5x more efficient than ARIAUSDT for the exact same expectancy per trade and the exact same sample size. The expectancy match told me these two configs earn the same per trade. It took a second number -- hold time -- to find out one of them does it 4.5x faster.
It's not just these two rows
I ran the same per-hour math across all 50 top candidates. Average hold time ranges from 6.0 minutes to 41.6 minutes across the set -- a 6.9x spread. Return per hour held ranges from 0.40%/hr to 2.09%/hr -- a 5.2x spread. Mean hold across all 50 is 17.0 minutes; mean return per hour is 1.34%/hr. None of that spread shows up in the score my pipeline sorts by. The 9 BILLUSDT rows score 47-59, less than half of MEGAUSDT's 117-131, but their return-per-hour (1.16%-1.80%/hr) sits in the same band as MEGAUSDT's. Score ranks net result, win rate and profit factor. It has nothing to say about whether your capital is tied up for 9 minutes or 42 to get there.
What I actually check now
Expectancy per trade and net result answer "did this make money." They don't answer "how much of my time did it cost to make that money," and those turn out to be two different questions that can point in opposite directions on rows my own filter calls equally good. Before I size anything from this top-50 list, I now pull avg_hold next to expectancy, the same way I started pulling max_mae next to score after none of these 50 configs cleared 20x breakeven leverage on the drawdown side. A row that ties for best expectancy isn't tied for best use of capital. I run the hold-time number through the position size calculator to see what that config actually lets me deploy per hour of exposure, and check the slow end of the range against the time-to-liquidation calculator -- a 41-minute average hold with a fat right tail is a different funding and gap-risk conversation than a 9-minute one, even when the expectancy column says they're twins. Same two numbers. Same sample size. A clock neither score nor net result was ever going to show me.