In July 2026 we ran a deliberately clean-slate review of published trading strategies and expert advisors. The agents doing the work were given no prior JPTC findings, so nothing could be confirmed by assumption. Twelve discovery angles produced 136 raw candidates. Fourteen were examined in depth. Each survivor was then attacked by two independent reviewers, one looking at the evidence and one at whether it could actually be implemented. None of the fourteen cleared the bar. This page publishes what we found, including the parts that are inconvenient for anyone selling an EA, which includes us.
Published by JPTC Research Team under the JPTC editorial policy.
Read how we testThe reviewers were instructed to default to rejection. So zero survivors means nothing cleared a deliberately brutal bar, not that nothing works. That distinction matters and we are not going to blur it to make a headline. What gives the result weight is that several kills came from the reviewers own out-of-sample replications returning the wrong sign, not from opinion.
Read equity drawdown relative, never balance drawdown. Balance drawdown counts closed trades only, so it is structurally blind to an open basket of losing positions. An expert advisor that adds to a losing position can advertise a four to sixteen percent drawdown while its equity drawdown runs far higher. Any EA that can add to a loser is disqualified before you look at its returns.
The most-linked free EA repository we found, with 583 stars, has an author who states across all eleven of its backtest readme files that the pure strategies were not profitable long term and only became profitable with the grid technique. The grid engine adds positions at rising volume for up to fifty levels with the stop loss ignored. The bundled reports show equity drawdown relative between 56 and 94 percent, while the advertised figure is balance drawdown. All eleven ran with no out-of-sample period at all, on generated ticks.
One popular repository advertising a prop firm drawdown protection EA contains four files: a readme, an html page, two svg images and a download button pointing somewhere else. There is not one line of MQL in it. Its star count is higher than most repositories that contain a real EA.
We ran targeted code searches across MQL repositories for links to third-party verification services. The results were unrelated blogs and signal-seller catalogues. The count of open-source MQL EAs with a third-party-verified live track record, anywhere we could search, is zero.
Placing pending orders either side of a high-impact release is the most marketed event strategy in the retail world and we found no verifiable evidence for it. The mechanism works against itself: a major pair widens from under a pip to five or ten pips around the release, which typically triggers both legs on the spread expansion alone. Separately, most funded accounts breach immediately on any execution within two minutes of a selected release.
The flagship product in this category, at 2,800 dollars, markets a perceptron and claimed language-model integration. By the vendor own description the network gates a grid system, and the verified history it cites belongs to the grid mechanics of its predecessor. No ablation is offered showing the machine learning adds anything. Vendor-quoted drawdown reaches 56.93 percent on its high-risk settings. A long run of profitable months is the expected signature of a grid that has not blown up yet, not evidence of an edge.
A 2026 paper re-ran published language-model trading strategies across more than a hundred symbols over twenty-four years with survivorship and look-ahead controls, and the reported advantages largely evaporated. A separate audit of nineteen agentic trading studies found two of nineteen used a time-consistent split, one of nineteen modelled transaction costs, and none reached the highest reproducibility tier. The leakage is not fixable by filtering inputs, because the models reproduce in-cutoff prices from their own weights.
Most prop firms, FTMO included, publish no pass-rate data, and most figures that circulate online trace back to content sites citing each other with no sample size and no method. We followed one widely quoted claim, about traders adhering to specific daily and overall limits being far more likely to pass based on five thousand participants, to its named source article. That article contains no such statistic. Two sources do stand up. Topstep publishes its own numbers: in 2025, 16.8% of Trading Combines were completed and 33.3% of traders at the funded level received a payout. And FPFX Tech, a software vendor to prop firms, reported to Finance Magnates in 2024 that across 300,000 accounts at ten firms, 14% of traders passed and about 7% were ever paid. There is still no published comparison of algorithmic against manual trader outcomes at prop firms.
This is the finding we would most like traders to take away. We simulated a strategy with no edge at all: every trade risks 1% with a fixed stop and target, and the win rate is set so the expectancy is exactly zero before costs. On standard two-step challenge rules (a 10% and then a 5% profit target, a 5% daily and a 10% overall loss limit, no time limit) it passed both phases about one time in three before costs. With trading costs of 2% of the risk per trade it still passed about one time in four, and at 5% between one in eight and one in five, depending on the reward-to-risk ratio. Before costs and at low costs, the expected value of an attempt is positive purely from the asymmetry between the fee and the payout. Published exemplars of consistent traders often have exactly that shape: very high win rates at reward-to-risk ratios below one. Before paying another challenge fee, run your own win rate and reward-to-risk through that null and see whether your results are distinguishable from chance.
Every major firm we checked prohibits it with immediate termination and forfeiture of profit, detected at millisecond level. This is the abuse that led the platform vendor to revoke prop firm licences in February 2024, taking a large number of firms down with it. There is also a documented regulatory case where a broker zeroed an arbitrage trader account and was found to be within its rights.
We pointed the same machinery at our own work rather than at the market. Pre-registered gates, out-of-sample requirements, a coin-flip null control, and a synthetic-noise audit on the principle that if a pipeline finds an edge in noise then every result it has ever produced is void. We would rather publish a method that can fail than a number that cannot be checked.
An earlier version of this page said that prop firms publish no pass-rate data and that every circulating figure traces back to content sites. That was wrong. Topstep publishes its own figures and FPFX Tech published data across ten firms; both are now cited above and listed under sources. The same update added the numbers from our zero-edge challenge simulation.
No. It means none of these fourteen cleared a bar we set deliberately high, and that the reviewers were told to reject by default. Several were killed by replications that came back with the wrong sign, which is stronger evidence than an opinion, but zero survivors is a statement about the bar as much as about the strategies.
Because the alternative is competing on claims, and the claims in this market are not checkable. Everything above is either a public repository anyone can open, a published paper anyone can re-run, or arithmetic anyone can do. We would rather be judged on method.
The findings, the kill reasons and the sources are what we publish. The candidate list and our own strategy parameters are not published, for the same reason we do not publish setfiles.
Equity drawdown relative rather than balance drawdown. Whether it can add to a losing position. Whether there is an out-of-sample period at all. Whether any live record is verified by someone other than the seller. Most products fail on the first two.