The Two Strategies That Survived Our Cemetery

Last week we published The Cemetery Census — we tested all 30 free trading robots that ship with MetaTrader 5. Twenty-nine died. One survived, and only barely.

That article was the funeral. This one is what happened next.

Once we had a proper testing pipeline built — real broker spreads, real overnight fees, out-of-sample validation, the works — we didn't stop at 30 free robots. We kept going. Trend-following systems. Mean-reversion systems. Institutional order-flow models (Wyckoff). Liquidity-sweep setups. A structure-trading engine we'd built ourselves. Harmonic pattern reversals. Session-timing effects.

In total: 13 genuinely different trading ideas, tested with full rigor. 11 died.

Two survived every gate we could throw at them. This is what they are, the real numbers behind them, and — because we test before we build, including on ourselves — an honest admission of what we don't know yet.

The testing bar (so the numbers below mean something)

Every idea we test has to clear four separate hurdles, in order:

1. Beat real transaction costs — not a theoretical backtest, the actual spread and overnight financing a real account pays
2. Survive a parameter neighborhood check — if nudging a setting by 10% destroys the result, it wasn't a real edge, it was a lucky number
3. Hold up on data it never saw — we tune on one period, then test once, unchanged, on a locked-away later period
4. Pass the platform's own tick-level Strategy Tester — the final, most realistic simulation before anything goes near an account

Most ideas — including ones with real theoretical pedigree — die at step 1 or step 3. That's not a flaw in the ideas. It's what an honest test looks like.

Survivor #1: BTC-USD, slow trend, 4-hour chart

The idea: a classic trend-following crossover — two exponential moving averages (27-period and 125-period), both directions, with a volatility-based stop and trailing exit. Nothing exotic. The "boring" family of strategy that keeps winning in our testing, the same family behind the lone Cemetery Census survivor.

What made it different from the versions that failed: timeframe. The exact same logic on a 1-hour chart came back roughly breakeven once real spread was charged. Moved to the 4-hour chart, the spread shrinks to a rounding error against the typical price swing — and the strategy clears every gate.

Validated results (Strategy Tester, 2021–2026, real spreads):

• $10,000 → $16,193 (+61.9%)
• Roughly 130 trades over 5.5 years — about 2 a week
• ~40% win rate. It loses more often than it wins. The wins are simply bigger.
• Max drawdown under 4%

Survivor #2: Gold (XAU-USD), slow trend, daily chart

The idea: the same crossover family, tuned for gold's own rhythm — 10-period and 50-period EMAs, daily chart.

Validated results (Strategy Tester, 2018–2026, real spreads):

• $10,000 → $23,820 (+138%)
• Only ~65 trades over 8 years — roughly one a month
• ~35% win rate
• Worst individual year: −1.2%. No year came close to a serious drawdown.

Gold's profile is the more interesting one: most years it does almost nothing, and then a real trend year (2023 in our data) carries the bulk of the return. It's less a monthly paycheck and more a slow-burning position that occasionally pays off big — which, if you understand why it works, is exactly the point.

What we buried to get here

For the record, and because we'd rather show the losses than hide them: a fast mean-reversion system (great win rate, killed entirely by spread costs), a structure/breakout trading engine (lost on 28 of 32 markets tested), Wyckoff spring/upthrust setups (couldn't be fairly tested without real volume data), a Donchian channel breakout (the actual 1980s "Turtle Traders" system — failed on every market we tried), an ADX-strength-confirmation entry, harmonic pattern reversals (caught a lookahead bug in our own code along the way — a good day, honestly), and a well-specified institutional liquidity-sweep model that turned out too selective to gather a large enough sample to trust either way.

Every one of those had a plausible story behind it. Plausible isn't the same as real.

The part we're not going to dress up

Here's where most trading content stops — right after the impressive numbers. We're doing the opposite.

Everything above is a backtest. A rigorously built one, tested against real costs and validated on data the tuning process never touched — but a backtest is still a simulation of the past, not proof of the future. As of today, both strategies are live on a demo account, trading real-time with real (if practice) money on the line. We have exactly one day of live data. One day tells you nothing.

What we're committing to publicly: we'll report back with the real, live results — good or bad — after a proper tracking period. If the numbers hold up anywhere near the backtest, that's a genuinely rare thing worth talking about. If they don't, we'll say that too, the same way we said MACD Sample turned $10,000 into $41.90.

That's the whole point of testing this way. The receipts don't stop being honest just because they're now about us.

Follow along — Field Report Nº 003 covers what actually happened.

---

Method note: MetaTrader 5 Strategy Tester and a custom Python validation pipeline (MetaTrader5 API), real historical broker spreads and swap rates, out-of-sample train/validate splits. Results are historical simulation; they are not investment advice, and past performance does not guarantee future results — a point this article exists specifically to take seriously.