100% Win Rate, Negative Expectancy: What Ten Years of a Martingale Grid Actually Looks Like

A martingale grid is the most seductive shape in retail trading. Equity curves go up and to the right. The win rate sits near 100%. Drawdowns look tiny right up until the one that isn't. We took a real one, replayed it over ten years of daily bars on six instruments, and then started removing its protections one at a time to find out which part was actually making the money.

The answer, in short: all of the profit came from being allowed to lose without limit.

The setup

The system is a mean-reversion grid built on a 250-day EMA. Deviation is the percentage distance of price from that EMA. When deviation crosses below −1.4% it goes long; when it crosses above +1.4% it goes short. From there it adds to the position every time price moves 0.8 × ATR(19) against the entry, and the lot size multiplies by 2.7 every seven adds. It closes the whole basket at +$31 of profit. The maximum loss parameter is set to $99,999. Starting size is 0.1 lot.

One structural note that matters later: because entries only fire from a flat book, an existing position is never reversed by an opposite signal. It can only leave through the profit target or the maximum loss.

The headline numbers

Ten years of daily bars, the script exactly as written (2.7× sizing, no effective stop):

Two things stand out. First, how wildly the trade count varies: 79 on EUR/USD, 1 on the VIX. That is not noise. Deviation sits outside the ±1.4% band on 70% of EUR/USD bars but 95% of VIX bars, and the system can only re-arm itself while price is inside the bands. The fewer bars it spends inside, the fewer entries it can take.

Second, look at the peak sizes. Peak exposure on gold was 6.24 lots — roughly $940,000 of notional, starting from 0.1 lot. That is the martingale doing exactly what it was designed to do.

Finding 1: the stop that isn't a stop

A maximum loss of $99,999 is not a risk limit. It is the absence of one, written as a number. On Bitcoin it was breached exactly once in ten years, and that single exit was for about −$100,000 — enough to turn 74 winning trades into a net loss of −$91,930.

This is the whole shape of the strategy in one row of a table. Seventy-four small, reliable wins and one catastrophic loss, and the catastrophic loss is bigger.

Finding 2: a 100% win rate tells you nothing

On gold with no effective stop, the worst closed trade out of 67 was +$33. Not one position was ever booked at a loss. The best was +$5,114.

That is not skill. It is structural: a grid that is never forced to close will keep adding until a small bounce is enough to reach the target, so every exit is by definition a winner. The losing positions are still there — they are simply open, sitting at −$45,667 at the worst moment rather than being booked.

If you are evaluating a system and the win rate is at or near 100%, the useful question is not how it wins. It is what happens to the positions it has not closed yet.

Finding 3: switching the martingale off made it worse

The obvious fix seems to be to remove the lot multiplier and keep everything else. We ran it: every add stays at 0.1 lot.

Gold collapses. The VIX never closes a single trade in ten years. The mechanism is simple: with flat sizing, each additional 0.1 lot barely moves the average entry, so the basket needs a much larger reversal to reach +$31. Without the escalating lot sizes it cannot get there, positions stay open, and eventually they hit the maximum loss instead. The martingale's final, enormous add is what drags the average entry close enough to the current price for a small bounce to close the trade.

So the martingale is not a risk overlay bolted onto a profitable system. It is the system. It is the mechanism that produces exits, and the exits are where all of the profit lives.

Finding 4: cap the loss and the edge disappears

The last test is the one that matters for anyone actually considering running this. We kept the 2.7× sizing and simply capped the loss at realistic levels.

Gold flips from +$27,786 to −$15,999 the moment the loss is capped at $5,000. Every instrument that produced meaningful profit goes negative once the loss is realistically capped. The stop-exit count on gold climbs 0 → 5 → 9 → 15 as the cap tightens, which is the mechanism laid bare: with a tighter cap, more baskets are cut off before they can recover.

An uncapped grid can always wait for a rebound. That waiting is not free — it is the entire source of return.

What to look at instead

For any system that adds to positions, the useful metrics are all about the open position rather than the closed ones:

• Worst mark-to-market against net profit. Gold made $27,786 while sitting on $45,667 of open loss at its worst. That ratio — about 1.6× — is the real risk/reward, and it is far worse than the win rate suggests.
• Peak exposure in money, not lots. "6.24 lots" sounds modest. "$940,000 of notional from a 0.1-lot start" does not.
• How often the system is armed. If the trigger condition is only true 5% of the time, your sample of trades is far smaller than the calendar suggests.
• What happens when you cap the loss. If the answer is "it stops working," the returns were payment for tail risk, not edge.

What this backtest does not tell you

These figures are gross of spreads, commissions and swaps, which matters enormously for a system whose average winner is a few hundred dollars and which trades 0.1-lot clips. They are measured on daily closes, so intrabar execution would differ. Contract sizes for index CFDs, single stocks and crypto are broker-specific and were approximated with standard conventions — so treat the dollar figures for those instruments as indicative rather than exact. Finally, ten years is still one sample: the Bitcoin result rests on a single stop-out.

You can run all of these variations yourself on the chart page, which replays the same logic on real daily bars and lets you flip the lot multiplier and the loss cap to watch the numbers move.

How these numbers were produced

Every figure above comes from replaying this grid through the simulation module behind this site's strategy pages, over ten years of daily bars taken from the same candle feed that powers /chart. The instruments tested include gold (XAUUSD), EUR/USD, Bitcoin and VIX. Unless a test says otherwise, lots scale 2.7x every seven adds, the basket closes at +$31, and the loss cap is the variable being changed.

Replay the same parameters on a different data vendor and the numbers will move — a grid this sensitive to its worst bar is supposed to. What survives the change of feed is the shape, not the decimals: the return was payment for unlimited tail risk, not edge. That is the part worth taking away, and it is the part you should test on your own data before risking anything.