The Steadiest Win Rate All Day Is the One That Breaks First Under Leverage

My backtest table has a column I'd never pulled up on its own before this week: hour_performance, 12,004 rows, every symbol-family combo split into 24 buckets by UTC hour. I went back to the same four symbols I wrote about a few days ago -- the ones that clear my pipeline's top-candidate filter, all range_bottom_reclaim -- and asked a different question than last time. Not "what's the worst drawdown." Just: does this thing trade the same way at 3am as it does at 3pm?

MEGAUSDT does. Across all 24 hours, its win rate never drops below 91.9% and never exceeds 100%. That's an 8.1 percentage-point band, held for 1,714 trades, with every single hour carrying a real sample -- the thinnest hour still has 51 trades, the busiest 95. There's no hour of the day where this thing has a bad time. It's the steadiest symbol in the set by a wide margin.

It's also the one I already found has the worst breakeven leverage of the four: roughly 7x, from converting its worst max_mae into a liquidation ceiling. The symbol with zero bad hours is the symbol that liquidates first.

The other three don't hold still

*Breakeven leverage figures from the worst-case max_mae conversion in my previous post. †BILLUSDT trades so rarely outside one hour (50 of its 285 total trades sit in hour 23 alone) that most of its 24 buckets have under 15 trades -- its 40pp win-rate swing is thin-sample noise, not a real pattern, and I'm flagging that rather than pretending it fits the story.

ARIAUSDT swings 60.7 percentage points between its worst hour (39.3% WR at UTC 3, on 84 trades) and its best (100% at UTC 6, on 65 trades). LIGHTUSDT swings even more -- 75 points, from a 25.0% win rate at UTC 7 on a real 72-trade sample up to 100% at UTC 23 on 74 trades. Both of these have solid sample sizes in every single bucket, same as MEGAUSDT. The swing isn't noise. It's a real symbol that trades well in some hours and badly in others. And both of them tolerate more leverage than MEGAUSDT does.

Why I didn't expect this

My first guess, before I ran the numbers, was that hour-to-hour stability and drawdown severity would move together -- a system that behaves the same way all day "should" have tighter, more predictable losses too. That's not what the data says. MEGAUSDT's stability is about win rate: it wins in roughly the same proportion no matter when you run it. Its breakeven leverage is about the size of the single worst loss it ever took, which happened once, at max_mae of -14.24%, and the hour-performance table doesn't even tell you when that one trade happened -- just that average win rate near that hour (UTC 7, -14.24% max_mae for MEGAUSDT sits inside hour 7's row) still ran 98.7% WR on 77 trades. One ugly tail inside an hour that otherwise looks completely normal. Consistency in the frequency of wins says nothing about the size of the one trade that goes badly wrong.

Zooming out past these four: across the full 12,004-row hour_performance table, UTC hour 2 averages the lowest expectancy of any hour (0.124) and UTC hour 7 the highest (0.211) -- a 71% spread purely by hour of day, pooled across every symbol and family I've backtested. I'm not claiming hour 2 is cursed; pooled averages mix hundreds of different symbols with different listing dates and liquidity regimes, and that's a population-level number, not a rule for any one pair. But it's consistent with what the four-symbol breakdown shows at a smaller scale: time of day moves the numbers more than I assumed before I looked.

What I actually do with this

I stopped treating "stable win rate across hours" as a proxy for "safe to leverage." They're two separate questions and this table answers one of them, not both. Before I'd size a position on any of these four, I still run its breakeven number through the safe leverage calculator with a real buffer under the breakeven line -- same as I wrote last time -- and I now also check whether the win rate holds up around the specific hour I'd actually be trading it, using the win rate calculator against that bucket's sample rather than the symbol's 24-hour average. A system that's 95.6% on average can still be the 39.3%-WR version of itself for one specific hour window, and the average won't warn you. I also keep the max drawdown calculator open next to it, because the hour where win rate looks fine is exactly the hour where a single bad max_mae tail -- the kind that doesn't show up until it's already liquidated you -- does its damage. Steadiness in how often you win and survivability of the one time you lose big are not the same property, and this table is the first time I've had numbers that show the gap between them directly.