The Mysterious Case of DerivAI_EA

We've buried a lot of trading robots in this series. The Cemetery Census killed 29 out of 30. The Two Survivors found exactly two worth keeping, out of thirteen genuinely different ideas tested with full rigor.

This one's different. This one didn't die cleanly, and it didn't survive cleanly either. It's still an open case — but the investigation itself is the interesting part.

The find

Sorting through an old folder of trading robots nobody had touched in months, one name stood out: DerivAI_EA. A 772-line multi-timeframe system — higher-timeframe trend bias, execution-timeframe entries, candlestick pattern detection, support/resistance zones, and something calling itself an "AI signal score" from 1 to 10. The kind of thing that reads like a kitchen-sink strategy someone built by throwing every trading idea they'd ever heard of into one file.

A quick smoke test on gold (XAUUSD): $10,000 → $56,556.

We've learned not to trust a number like that. The last time something this size showed up — a different EA, +1,292% — the follow-up check timed out and the number was never confirmed. Filed under "probably a mirage," never proven either way.

This one, we didn't let get away.

Clue one: it survives the harder test

The first smoke test uses a fast approximation — the tester fills in the gaps between price bars instead of simulating every real tick. It's quick, but it's not the real thing.

We reran it with full tick-by-tick simulation. If the number was an artifact of the approximation, it should have collapsed.

$10,000 → $52,464. Lower, but in the same neighborhood. A real result doesn't evaporate under more scrutiny — it just gets more precise. This one held.

Then we nudged its settings — moving averages a few bars faster, a few bars slower, the stop-loss wider. Every single variation still landed strongly positive, clustered together. That's the signature of a real, stable pattern, not a lucky specific number. This was the most robust neighborhood we'd seen in fifteen separate investigations.

Everything was pointing toward: we might have actually found something.

Clue two: the plot twist

Then we asked the obvious next question — a trend-following, pattern-reading, multi-timeframe system shouldn't only work on one market. So we ran it on Bitcoin. On the major currency pairs. On synthetic indices.

BTCUSD: flat, nothing. EURUSD: −25.7%. GBPUSD: −55.3%. USDJPY: −63.7%.

Every non-gold market either did nothing or lost badly. Whatever this thing was doing, it wasn't a transferable trading skill. It was gold-specific, full stop.

That's strange on its own. But then we went back further in time — testing 2022 to 2024, a period the earlier smoke tests hadn't touched.

$10,000 → $8,983. A loss.

The same unmodified EA, same market, different years: dead losing in one stretch, wildly profitable in another. Not a decline. Not a gentle fade. A flip.

The reveal

Here's where the mystery actually resolves.

We pulled the full trade log for the 2022–2026 window. 18,479 closed trades over three years — roughly 17 trades a day, every single day, on gold, on a 15-minute chart.

Then we added up the wins and the losses separately.

Gross winning trades: $113,212. Gross losing trades: $113,190.

Those two numbers are, for practical purposes, identical. The raw edge — before spread, before commission, before any real-world cost — was a coin flip. Fifty-fifty, functionally.

That's the whole mystery, solved in one sentence: this isn't a strategy that works in trends and fails in chop. It's a strategy with no real edge at all, that happened to get carried by one exceptional, real, well-documented gold rally big enough for its winning trades to outrun its own trading costs. In 2022–2024, gold didn't move far enough per trade to pay that toll 17 times a day. In 2024–2026, it did — spectacularly. Same coin, same flips, different-sized prizes on the winning side.

It's the same lesson we keep re-learning in this series in a new disguise every time: trade too often, and the house always eventually finds you — unless the market happens to hand you an unusually large winning streak to hide behind.

Attempting a fix

Once you know the actual disease, you can test actual medicine — not just throw filters at it and hope.

We tried making the EA pickier about which setups it takes (MinScoreToTrade, raised from 6 to 8 out of 10). That alone flipped the losing 2022–2024 period from −10.2% to +3.5%, while the winning period dropped from spectacular to still-strong (+145%). A real, targeted improvement — not a lucky number, since it held up across a proper unseen-data test too: +56% on 2025, +90% on a locked-away slice of 2026 evaluated exactly once.

We also tried slowing the whole thing down — the same fix that turned our own BTC strategy from a loser into a winner months ago. Here, it made everything worse in every period tested. This particular EA, unlike ours, genuinely needs to trade fast to work at all. Some fixes don't transfer.

Where the case stands

Unsolved, technically. MinScoreToTrade=8 is a real, evidence-backed improvement — but the older 2020–2022 history is still a net loser under it, and "the last two years were kind to gold traders" is not the same as "we found a repeatable edge." It needs a longer track record before it earns a place next to our two actually-validated strategies.

What we can say for certain: the mystery wasn't in the market. It was in the trade count. Seventeen trades a day is not a trading strategy — it's a tollbooth, and for most of its life, this one was paying more toll than it collected.

We're leaving this case open. If gold keeps trending the way it has, we may have more to report.

---

Method note: MetaTrader 5 Strategy Tester with full tick-level simulation, cross-symbol testing across seven markets, and a full trade-log audit (18,479 trades) to identify the underlying raw edge before costs. Results are historical simulation, not investment advice, and this specific system is not currently live on any account.