My EA Made $87,000. It Still Almost Died Five Times!

In part 1, I showed the autopsy: two grid EAs across eight real cent accounts, about 25,000 trades and ~$87,000 in closed profit — and the single hidden flaw (deep recovery ladders) that killed one account and nearly killed several more. This is part 2: the rebuild. I ran the four historical disasters back through TrendLab Fusion in Strategy Tester to see whether the fixes actually hold. The testing didn’t just validate the design — it found three real bugs I had to fix. Here’s everything, including a benchmark that failed before it passed.

✅ What this is (and isn’t): Everything below is Strategy Tester and backtest data, except the eight real account histories from part 1. It is not live proof. The cent-account forward test is launching now, and I’ll publish those real results here as they come — wins and losses. Backtests can flatter; forward results are the real judge. Read this as “the rebuild survived the replays in testing,” not “this is proven live.”

TrendLab Fusion: what was kept, and what was fixed

The forensics in part 1 were clear: the engine made money almost everywhere up to basket depth 7, and died at depth 8+. So the rebuild keeps the proven DNA and bolts on the protection layer the originals never had.

Kept (the proven part): the 200-point grid, dynamic spacing, the M15 signal cadence, both trade directions, basket take-profit on the volume-weighted average. Added (the five fixes, each traceable to a real disaster):

• Hard depth cap at 6 — the depth-8+ death zone simply can’t happen anymore.
• H4 trend gate (EMA + ATR) on adds beyond depth 3 — the fix for “averaging into a freight train,” the exact mechanism of the GBPAUD and AUDNZD disasters.
• Hard basket stop, a fixed % of equity — any single basket’s worst case is now bounded.
• Global equity brake at 12% with a same-day cooldown.
• Evidence-based pairs and per-pair risk tiers — the dangerous pairs get tighter caps and stops automatically. AutoConfig only ever tightens risk, never loosens it.

The part I didn’t expect: the testing found three bugs

Here’s the honest, and honestly exciting, part. I didn’t write perfect code and watch it pass. The stress tests broke it — three times — and each break made the final version safer:

• Bug 1 — the midnight spread cascade. The very first replay stacked five ladder rungs in eight seconds just after midnight, when the rollover spread spiked. The fix: no new entries when spread is abnormally wide, plus a minimum time gap between rungs. The telling part: even with the bug creating an oversized ladder, the hard stop amputated it at −$258. The safety net caught a failure of the system itself.
• Bug 2 — midnight stop misfires. In the 6-month benchmark, two of three stops fired around midnight on baskets that weren’t really at the loss trigger — rollover spread and triple swap faked it. The fix evaluates the stop on mid-price. Proof it mattered: after the fix, those same baskets survived and closed at +$25 and +$23 instead of being stopped.
• Bug 3 — ladder re-escalation. A subtle interaction let lot sizes creep above the designed ceiling. The fix locks the ladder math so the ceiling holds forever.

This is why you test before you go live, not after. Every one of these would have cost real money on a real account. The stress test found them before the market could.

TrendLab Fusion disaster replays: the real numbers

The core question: run the four historical disasters — the ones that lost about $8,000 and destroyed accounts — back through Fusion. What happens?

Strategy Tester replays. Two replays used 100% real tick data (the GBPAUD blowup week and the AUDUSD ladder); two used simulated ticks where the real-tick archive didn’t reach back. This is exactly what part 1 promised: real replay results, not projections.

Read this honestly. Fusion did not turn these losses into big wins — that was never the goal, and anyone promising that is lying. What it did was exactly what it was designed to do: on two disasters the trend gate refused the doomed ladder entirely (small profits instead of catastrophe), and on the other two the hard stop amputated the loss at around −$200–$300 instead of −$2,400 to −$3,243. About $8,000 of destruction and two dead accounts became roughly −$355 with a worst drawdown of 2.58%. The disasters were capped, not erased.

The benchmark: an honest failure, then a pass

I also ran a 6-month benchmark, and I’m reporting it in full because the first run failed:

Same pair, same months, only the bug-fixes changed: an $860 swing from losing to winning. The full 5-pair portfolio over the same period returned about +$1,417 (~$236/month) with a 74.9% win rate across 1,388 trades and just one hard stop — max drawdown 2.02%.

The most honest number in this whole post

Here’s the comparison I refuse to soften. The original Trend Chief EA earned roughly $800/month in the same window — about three times Fusion’s rate — while carrying the exact ladder that eventually takes the account to zero. Golden Lab showed 18 months of profits and then went to $200 in the GBPAUD blowup.

Fusion makes about $236/month at fixed tiny lots, with a worst moment of −2%. It is slower. That’s the honest trade-off, and it’s the whole point: the extra profit the originals showed in good months was partly rented from the blowup that hadn’t happened yet. Fusion earns less per month and never faces that ending. I’ll take the slower, surviving version every time — because a strategy that dies gives it all back.

The “good trading days” myth (a bonus finding)

While parsing ~15,000 trades I tested a popular idea: that certain days or hours are “high-probability” and you should size up on them. Across the whole dataset, the win rate was flat at roughly 70–73% across every day and every hour. The occasional bigger-profit day was just where ladder resolutions happened to land — not an edge. So Fusion refuses to scale lots up, ever. Most EAs sell “smart lot boosting” as a feature; the data showed it’s the same disease as martingale, so I declined to build it. Defensive sizing only.

Honest caveats — because this is the whole brand

• The entry signal is a reconstruction (the originals were closed-source). The cadence matches; exact entries won’t. The forward test will validate real frequency.
• Two of the four replays ran on simulated ticks (the broker’s real-tick archive didn’t reach back that far); the other two were 100% real ticks.
• The benchmark used a fixed lot on a $10k test account. The cent launch is much smaller, so the same % stop means less dollar room — likely smaller, more frequent stops. The forward test will show this honestly.
• Everything except the eight real account histories is a backtest. Backtests are where strategies look their best. The real cent forward test starts now.

What happens next

TrendLab Fusion goes on a live cent account now — small real money, five evidence-based pairs, all the protections on. For the next 4–6 weeks I’ll log entries per day, every time the trend gate blocks a doomed add, every time the defensive brain reduces or pauses, and every hard-stop event. It only earns a bigger account if it stays positive with the stops behaving exactly to spec. If it doesn’t, you’ll read that here too.

That’s the deal this whole site runs on: the autopsy, the rebuild, and now the real-money test — published either way.

Important: All performance figures here are backtests / Strategy Tester results except the eight real account histories referenced from part 1. Backtest results routinely fail to survive live trading. Not financial advice. Grid and martingale-style strategies carry a very high risk of total account loss. Past and simulated results do not predict future outcomes. These are my own results and opinions. Never trade money you can’t afford to lose.