Almost every commercial EA description now includes the phrase "no martingale." It has turned into a badge, roughly the way "no repaint" did for indicators. The phrase is easy to write and impossible to verify from the outside, so a buyer either trusts it or does not.
When we started building our own USDJPY system, that bothered us — because the choice is not free. Refusing loss-recovery sequences has a visible cost: recoveries take longer, and the equity curve looks less impressive in testing. A decision with real costs should rest on measurements rather than on reputation, so before ruling anything out, we tested the alternatives properly.
This article describes how those tests were set up, what the numbers showed, and what changed in our system as a result.
Why martingale refuses to die
A martingale overlay increases position size after a loss, so that the next win recovers the accumulated deficit. Averaging-down and grid systems are close relatives: they add positions against an adverse move and wait for a partial retracement to close the whole cluster in profit.
These ideas persist for an understandable reason — over most test windows, they work. The win rate per completed sequence is high, and the balance curve looks smooth because losses stay unrealized while a sequence is open. The catastrophic outcome, the losing streak the account cannot absorb, is rare enough that a multi-year backtest may simply not contain one.
None of this requires bad intent from a developer. Someone can honestly test a grid system over several years, see a clean upward curve, and conclude that it is safe. The problem sits deeper than honesty: a backtest rewards systems that defer risk into rare tail events, and such systems look excellent right up until the tail arrives.
We wanted to know whether there was anything real underneath the smooth curves. That is a testable question.
Turning a slogan into a hypothesis
Instead of adopting a policy, we defined a study. We prepared 14 variants — 8 martingale progressions, 4 averaging-down schemes, and 2 grid configurations — and applied each of them to 12 years of USDJPY historical data. (The 12-year window belongs to this particular study; it is shorter than the two decades we use for full-system backtests.)
The pass condition was declared before anything ran. A variant would only be adopted if it improved risk-adjusted results, and if we could explain where the improvement came from in terms of market behavior. A larger ending balance was, by itself, not a pass.
That second condition carries more weight than it may appear to. Ending balance is the most seductive number in backtesting, and it is exactly the number that loss-progression systems inflate most easily. Acceptance criteria that can be satisfied by scaled-up exposure alone will, sooner or later, be satisfied by scaled-up exposure.
None of the 14 variants met the criteria. Zero were adopted.
What the doubling sequence actually did
The most instructive case was the classic double-down martingale: lose, double the lot, repeat until a win closes the sequence.
On paper it looked like a clear improvement. Total profit over the test came out four times higher than the same baseline without the overlay — and if we had stopped reading at that line, this article would have a very different conclusion.
The path told another story. On the historical route through the data, the account was down 18% of equity within the first year of operation, and producing the fourfold profit required a maximum lot 32 times the base size. The profit multiplication tracked the lot multiplication. Nothing about the trades themselves had improved; the same signals were simply carried through losing stretches with progressively heavier exposure.
The diagram below illustrates the mechanics of a doubling sequence. It is an example of the method, not a chart of our system.
Six consecutive losses are not exotic. Any strategy that trades for years will meet them. At that point the trader is carrying 32 times the intended lot, the open deficit equals the sum of five escalating losses, and the next decision gets made under exactly the kind of pressure a backtest never feels.
More profit is not the same as an edge
The four-times-profit result deserves a closer look, because the reasoning generalizes to any sizing overlay you might be tempted to test.
A trading edge lives in expectancy per unit of exposure: on average, does a unit of risk return more than it costs? A position-sizing rule operates after that question is already settled. It decides how much exposure to take, not whether the exposure is worthwhile. Doubling lots after losses cannot turn a neutral signal into a profitable one — it can only redistribute when profits and losses are realized, and scale everything up along the way.
That is what "profit ×4, max lot ×32" means in plain terms. The overlay added leverage, not skill. A similar expectation could be reached by simply trading larger fixed lots, but then the added risk would be visible in every single trade, and nobody would mistake it for a strategy improvement. Martingale produces the same enlargement while hiding it in the sequencing, which is why it keeps passing casual inspection.
This gives you a fast diagnostic for any exciting backtest: when profit grows, check whether exposure grew with it. If the two scale together, the sizing rule is doing the "improvement," and it will do the same to the losses when the bad stretch arrives.
Why the study was built as a direct test
The study had a second purpose beyond settling our own design question. Our product documentation states that betting systems were tested and rejected, and we wanted that sentence to rest on direct measurement rather than on theory or reputation.
For that reason, the variants were not tested as standalone strategies. Each one was applied as a sizing overlay on top of fixed signals and evaluated across deliberately stressed test cells, so that any difference in outcome could be attributed to the sizing layer alone. Every cell came back the same way: zero edge attributable to the sizing layer.
Probability theory has been saying this about betting progressions for a long time — a sizing rule cannot change the sign of per-trade expectancy — and we did not expect a different theoretical result. We ran the test anyway, because "the theory says so" and "we measured it on our data, and here is the result" are different levels of claim. A commercial product should rest on the second one.
Why backtests flatter these systems
The distortion that makes loss-progression systems look good in testing is structural rather than accidental, and it is worth understanding even if you never trade one.
A balance curve is drawn from closed trades. While a martingale sequence or a grid cluster is open, its growing deficit exists only in equity, not in balance — so the chart most people screenshot does not show the periods of greatest danger at all. A system can spend days carrying a deep open drawdown and still print an unbroken staircase of closed profits. Judged on the balance curve alone, its risk is invisible by construction.
The second distortion is the test window itself. The event that ends a loss-progression account — the streak longer than the progression can absorb — is rare by design, and a window of several years may simply not contain it. Our own 12-year study is not immune to that logic, which is exactly why the double-down result mattered so much: even inside a window where the system survived and quadrupled its paper profit, the path already showed an 18% equity decline and a 32× lot. We did not need to wait for the fatal streak to see what the profit was made of.
So when you review any EA that recovers losses by adding size, look at equity rather than balance, and treat a clean multi-year curve as the beginning of the investigation, not the end of it.
What we changed in our own system
Rejecting loss-progression systems is not just a matter of leaving a feature out. It shaped several decisions that have to hold together.
The first consequence is absolute: nothing in our system increases lot size after a loss. No recovery mode, no "smart" doubling, no compensation multiplier. The temptation to add one returns every time a losing period shows up in testing, which is exactly why the rule has to be categorical rather than situational.
The second consequence is harder to accept. Our system stops itself permanently if equity falls 18% from its peak, and no feature tries to win the loss back. Across our long-term testing, every recovery mechanism that relied on increased position size raised the probability of ruin instead of reducing it. If capital survives, the operator can make a calm decision about restarting; code under drawdown pressure should not be making that decision with enlarged lots.
These choices have a visible cost. Recoveries after losing periods are slower, and the equity curve never gets the cosmetic smoothing that loss-progression produces. We consider that an honest price. The figures in this article come from backtests on historical data — they describe the past, not the future.
If you want to check a sizing overlay yourself
The study design above translates into something any MetaTrader 5 user can reproduce, on your own EA or on one you are evaluating, in about an evening.
Separate the signal from the sizing, and run the same signal with fixed lots first so a baseline exists. Then enable the overlay and compare three things rather than one: expectancy per unit of exposure, the maximum lot reached anywhere on the path, and the worst equity point — not just the closed-balance curve, which hides open drawdowns by construction. Finish with the qualitative question: what market behavior would have to hold for the extra profit to be real? If the only honest answer is "losing streaks must stay short," what you measured was leverage.
Closing
"No martingale" appears in our product description too. The difference we can offer is that, in our case, the sentence is a measurement: 14 loss-progression variants went into a pre-registered study, and none came out. The concepts discussed in this article are implemented in Bushido, a commercial USDJPY portfolio EA developed by Origin-U Project — which, as a direct result of this work, contains no mechanism for trading its way out of a loss. Origin-U Project is an independent quantitative foreign exchange research project based in Japan.
