I pointed a gaming GPU at the market and told it to find me one clean edge. Not a portfolio, not a system — one honest setup that made money without a trick under it. The card ran for roughly five hours: 386 symbols, 220,000 configurations per symbol, about 85 million backtests in total on a single NVIDIA RTX 3070. Around 8,370 configs cleared the first screen. Fifty made the final shortlist.
All fifty were the same trap wearing different tickers. Let me show you.
The leaderboard looked incredible
On paper the shortlist is a dream. Median win rate 95.2%. Median profit factor 3.04, topping out at 3.47. Every one flagged "good for deep test." If I posted that screenshot with no context, half of crypto Twitter would ask to buy the bot.
Then I looked at the one column nobody screenshots.
Every winner averaged down
| Metric across the top 50 | Value |
|---|---|
| Median win rate | 95.2% |
| Median profit factor | 3.04 |
| Share of trades that averaged down (median) | 54.5% |
| Configs with dca-rate under 40% | 3 of 50 |
| Worst-case excursion on the best pair | −14.2% |
| Unique symbols in the whole top 50 | 4 |
Read the third row. More than half of every "winning" strategy's trades only closed green because the bot added to a losing position and waited for a bounce. Forty-seven of the fifty leaned on averaging down for at least 40% of their trades; the greediest one did it 72% of the time. That is not an entry with an edge. That is a martingale with good manners.
Here is why the win rate looks so pretty: when you buy a dip and then buy the deeper dip and then buy the deeper dip after that, your average entry keeps sliding down toward the price. A small reclaim bounce drags the whole stack back to breakeven, and the trade books as a win. Do that on 1,600 trades and you get a 95% win rate. You also get a position sitting 14% underwater at its worst, every time, waiting on the one dip that doesn't reclaim.
Four tickers, and the recipe fell apart on three of them
The most damning part isn't the averaging-down. It's that the entire fifty-name shortlist came from just four symbols, and thirty of the fifty slots were a single coin — MEGA — which happened to be in a long grind higher during the test window. The exact same strategy family, applied to the other pairs it surfaced, did this:
| Symbol | Win rate | Profit factor | Avg down rate |
|---|---|---|---|
| MEGA | 93–98% | ~3.14 | ~55% |
| BILL | 44–97% | ~1.78 | ~55% |
| ARIA | 48–89% | ~1.71 | ~42% |
| LIGHT | 33–37% | ~1.87 | ~63% |
Same logic, same code, same averaging-down engine. On MEGA it reads 97%. On LIGHT it reads 33% and averages down even harder to get there. A real edge holds its shape when you move it to a new chart. This one shattered. That spread — 97% here, 33% there — is the fingerprint of a curve fit, a setup that memorised one coin's recent behaviour rather than learning anything about markets.
Why the raw win rate is the wrong number
Profit factor and win rate are both headcounts of a sort. They tell you how the closed trades landed and nothing about how much rope you gave the open ones. The number that actually decides survival is maximum adverse excursion — how deep the trade went against you before it came back. Mine sat at −14.2% on the best candidates, on 1,600 trades apiece. That is fine right up until leverage enters the picture. At 10x, a −14% excursion is a −140% account move. The bot never "lost" on paper. It just lived one bad candle away from a margin call, 1,600 times.
This is the same lesson my whole trading year keeps teaching me. Big win rate is not profit. My five live paper systems ran 81 trades at a 15% net win rate and finished at −$0.567 — a system that loses 85% of the time landed near flat because the winners were large and the losers were capped. The averaging-down configs are the exact mirror image: they win almost every time and stay dangerous because the rare loss is uncapped. Both stories point at the same truth. Count the size, not the frequency.
What I actually did with the shortlist
I killed it. Eighty-five million backtests, five hours of GPU, and the honest output was zero deployable edges — just fifty flattering pictures of the same martingale. I didn't paper-trade a single one of them, because I already know how that movie ends. The account looks great until the one dip that doesn't reclaim, and then it's over in a single trade.
The screen wasn't wasted, though. It taught me what to grep for before I ever trust a green backtest again.
What I check now, before believing any backtest
- Read the averaging-down rate first, win rate last. If most trades needed a second or third entry to survive, the win rate is measuring my willingness to add risk, not the quality of the setup.
- Put the max adverse excursion through real leverage. A −14% excursion at 10x is a blown account. I run the worst drawdown through risk of ruin and drawdown recovery before anything else.
- Move the strategy to three other charts. If a 97% setup becomes 33% on the next symbol, it learned one coin, not the market.
- Trust MAE and expectancy over profit factor. Profit factor loved every one of these. The excursion column is what told the truth.
→ How deep can your averaging-down survive? · → Risk of ruin · → Profit factor calculator
FAQ
Why does a 95% win rate backtest still blow up?
Because the win rate is bought with averaging down. Every dip, the bot adds and waits for a bounce, so most trades close a few ticks green. The cost is a deep unrealized drawdown that stays invisible until one dip fails to reclaim — and that single trade is big enough to erase hundreds of small wins.
What is the "averaging-down rate" and why does it matter?
It's the share of trades that needed at least one add to survive. In my top 50 the median was ~55% and the max was 72%. When more than half your trades only work because you doubled down, the win rate measures risk appetite, not edge. The DCA survival calculator shows how far that can go before it breaks.
How many configurations did the screen actually test?
386 symbols × 220,000 configs ≈ 85 million backtests, on one RTX 3070 in about five hours. Only ~8,370 configs cleared the first screen and just 50 reached the deep-test shortlist.
If the win rate is fake, what number should I trust?
Maximum adverse excursion and expectancy. MAE says how deep the trade went underwater — mine sat near −14%. Expectancy says the average result per trade after everything nets out. A strategy is real only if it survives its own worst drawdown and still has positive expectancy.
How do I check my own strategy for this trap?
Count how many of your winning trades went meaningfully underwater first, and by how much. Then run that worst drawdown through risk of ruin. If recovery needs a return you've never actually produced, the win rate is lying.