Market Anomalies EA review: a 35 cent edge on a nine month window

2026-08-27, based on our dated catalog run of v1.8 on USDJPY. The entry is unlisted while the gold-only decision stands and the re-run on the new data world is pending, status on the audit page.

Market Anomalies EA looks unremarkable in the best possible sense. No grid, no martingale, no hidden basket. Our forensics count a median size escalation of 1.0 after losses, volume raised after 30.9% of losses versus 29.3% after wins, which is the signature of a system that does not care whether the last trade won. At most 3 positions are open at once on a single symbol, and long and short overlapped in exactly 1 moment across the whole run. The reconstructed floating drawdown is 6.38% against a closed trade drawdown of 6.35%, so what the balance curve shows is what the account lived through. After the grid and martingale families in this catalog, that is a pleasant thing to measure.

The run netted +$190.23 on a $1,000 account. Then you divide by the number of trades, and the article begins.

Thirty five cents a trade

544 trades produced a gross result of +$233.70. Commission took -$41.43 and swap took -$2.04, which leaves the net of +$190.23 and puts costs at 18.6% of gross. Per trade that is an expected payoff of $0.35, on an average position size of 0.148 lots.

Our sensitivity section asks the obvious follow up question. How much extra cost per trade does this system survive. The answer is $0.35, the break even shock, and it is the same number as the edge because that is what the edge is. Add half a dollar per trade, which on 0.148 lots is a modest amount of spread widening or slippage, and the same 544 trades come out at -$81.77 with a profit factor of 0.89. At one dollar per trade the result is -$353.77.

Read this the right way. A thin edge per trade is not a defect by itself. Many good systems live on small margins and make it up in frequency. What the number does say is that the distance between this backtest and a live account is measured in cents, so every assumption about spread, commission and fill quality carries the whole result. The risk this audit points at is not a structure that hands the account back in one session. It is an edge that is too thin to survive a broker who is slightly worse than ours.

The same fragility shows in the leave best out test. Remove the 10 most profitable trades from the observed set and the run falls from +$190.23 to +$15.10. Remove 20 and it turns negative at -$64.43. Twenty trades out of 544 carry the entire result.

Seven days out of two hundred and three

That concentration is not a statistical curiosity, it is the shape of the run. 80% of the profit was made in 7 trading days, which is 3.5% of the 203 trading days in the sample. The three best days contributed +$29.50, +$28.00 and +$27.12, against a total net of $190.23. The three worst cost -$18.67, -$18.47 and -$17.88. On the other 196 days the system essentially traded to a draw.

Alongside that sits the direction. 85.3% of the profit came from the long side on 79% long trades, and over the traded span USDJPY moved from 152.94 to 158.51, up 3.64%. Our audit files that as a long beta suspicion rather than a verdict, and the reason for the caution is honest. A system that is long a rising pair four times out of five has not yet shown whether it reads the market or simply leans on it.

The window that is mostly not there

The report head prints a test period of 2020.01.01 to 2026.08.11 on M1. The history quality line prints 13% real ticks, with real ticks starting 2025.09.19. The first trade in the deal census sits on 2025-10-24 and the last on 2026-08-10.

So the six and a half year label describes about nine and a half months of actual trading. That is not the vendor's doing and it is not a flaw in the EA. It is the tick history of the broker feed this run used, and it means every metric in this article is a statement about one bull leg in one pair, not about six years of anything. We print the period the tester printed and then we print what is underneath it, because the report head alone would have let us annualise a nine month sample without comment.

One more line belongs here. The log carries 229 randomization prints, which means the EA deliberately varies its own behaviour between runs. A single backtest of a randomizing system understates the spread of possible outcomes by construction. The honest reading of +$190.23 is one draw from a distribution, not the result.

The rulebook kills the prop account, not the drawdown

This is the most interesting section of the audit, because the two ways of failing point in opposite directions. Measured against the IQ Capital Classic funded profile at full size, the account never dies of drawdown. The sweep records 0 deaths at every sizing step and a withdrawal rate of 21.85% per year at full size. On the FTMO Challenge the same trades pass phase one with probability 0.903 and both phases with 0.878.

Now the rules that are not about drawdown. The per position loss limit of the IQ Capital profile, 0.5% of the account, is breached 27 times at full size. A funded account is hard breached on the second violation. And the consistency rule, which caps the best day at 30% of the year's profit, is breached in both tested years, at 30.3% in 2025 and 30.1% in 2026. Those two numbers sit just over the line, and we say so rather than dress them up, but they sit over it in both years and the cause is structural. A system that makes 80% of its money in seven days will keep producing a best day near the cap.

Sizing down fixes one thing and costs another. At half size the position loss breaches fall to 0 and the modelled floor breach probability drops from 0.442 to 0.02, while the withdrawal rate falls from 21.85% to 11.04% per year. On the FTMO side the same move is brutal in the other direction, because the target stays fixed while the trading shrinks. Phase one pass probability goes from 0.903 at full size to 0.570 at half size, 0.178 at one third, and 0.0 at one sixth.

The trade off in one line. At full size this EA passes challenges and breaks position rules. At safe size it keeps the rules and stops passing challenges. There is no sizing in our sweep that does both.

What the payout model actually pays

On the observed path with a monthly sweep to the start balance, the run pays out in 6 of 11 months, totalling $193.61 at fixed lots, which is 23.68% per year on the $1,000 base. The median paid month is $42.96 and the best is $58.93. The number to look at is the other one. The longest dry streak is 3 months, and the sample is only eleven months long, so a quarter of the tested life paid nothing at all.

What we would need to change the verdict

The current dimensions read info on data quality and structure, ok on costs, and caution on concentration, regime and prop fit. None of that says the EA is bad. It says the measurement cannot yet separate the mechanics from the window.

The honest limits. One tester run of v1.8 on USDJPY M1 at 1:100 on a $1,000 deposit, an effective window of about nine and a half months on 13% real ticks, in sample by definition, and a randomizing EA measured once. The deal census is complete (544 trades, log coverage 1.0, 0 chronology violations, 0 duplicate deal ids) and the reconstructed end balance matches the report to the cent, so the arithmetic is solid. What the numbers support: the structure is clean, the edge is real but thin enough that costs decide it, the profit is concentrated in a handful of long days, and no confident verdict about live performance follows from this data. The vendor has since published v2.0, which we have not tested.

Related: the audit page for this EA · all audits · more articles