LEARN · EN · easyquanttrading.com
We forged 3,655 strategies in 30 days. 90% were eliminated.
Most platforms show you what survived. Almost none show you what died, because the ratio is embarrassing. We publish it anyway, because it is the only honest way to read any strategy list, ours included. Every number below is read directly from our own production database over a 30-day window. Nothing is rounded up, sampled or cherry-picked.
- 9,180 automated forging runs emitted 55,684 raw candidates in 30 days.
- 3,655 unique strategies survived de-duplication. 362 are still live: a 9.9% survival rate.
- 1,655 strategies were scored by the Deflated Sharpe Ratio. Exactly 1 passed.
- The simplest engine survived 23x more often than the most complex one.
- 45.9% of the strategies that survived de-duplication are near-duplicates of each other.
- Only 46.2% could be reproduced on a real MetaTrader 5 terminal.
The funnel, unfiltered
Everything starts as a candidate. Most of it dies. This is the first table most platforms never publish, because the top row is embarrassing and the bottom row is the only one that gets marketed.
| Stage | Count |
|---|---|
| Automated forging runs | 9,180 |
| Raw candidates emitted | 55,684 |
| New strategies stored after de-duplication | 5,564 |
| Duplicate candidates skipped | 572 |
| Unique strategies on the shelf | 3,655 |
| Still live after every gate | 362 |
| Eliminated | 3,293 |
30-day window, 2026-08-21 to 2026-09-19. The shelf count is lower than the stored count because strategies that are authoritatively unprofitable are deleted rather than hidden.
Why 90% were eliminated
This is the part that matters. A platform that only tells you a strategy failed has told you nothing. The reason is the actual product: it tells you whether your search is finding signal or memorising noise.
| Recorded reason for removal | Strategies |
|---|---|
| Out-of-sample collapse (severe or high) | 1,078 |
| Degenerate signal: fired on 95-100% of bars | 421 |
| In-sample profit, out-of-sample loss (textbook overfitting) | 165 |
| Failed the Deflated Sharpe Ratio test | 128 |
| Weak or stale: decayed off the shelf | 50 |
| Empty shell: zero trades generated | 22 |
| Dataset retired from sale (crypto delisting) | 6 |
Reasons are recorded for 2,931 of the 3,293 removals. A single strategy can carry more than one reason, so rows overlap and do not sum to the total.
The single most important number: PSR versus DSR
This is the finding we would most want a competitor to have to answer for. We score strategies with two related statistics. The Probabilistic Sharpe Ratio (PSR) asks how likely the observed Sharpe is greater than zero. The Deflated Sharpe Ratio (DSR) asks the same question after correcting for how many strategies we tried before this one looked good.
The same population of strategies scores a median PSR of 0.956 and a median DSR of 0.042. Nothing about the strategies changed between those two columns. The only thing that changed is that we admitted how much we searched.
If you are shown a strategy list with a Sharpe ratio and nothing else, you are effectively being shown the PSR column while the DSR column is withheld. That gap is not a rounding detail. It is the difference between a research finding and a coincidence.
| Measure | Value | What it tells you |
|---|---|---|
| PSR, median | 0.956 | Looks excellent: the Sharpe looks reliably positive |
| DSR, median | 0.042 | Collapses once you correct for the number of trials |
| DSR pass rate | 1 of 1,655 (0.1%) | Almost nothing survives a multiple-testing correction |
| Judged trustworthy | 10 of 3,655 | The strictest honesty verdict in the pipeline |
Computed on the live population of forged strategies. DSR is shown, never hidden; failing it is recorded, not deleted.
The engine paradox: complexity made things worse
We run eight different strategy engines, from fixed indicator templates to differential evolution and multi-objective genetic algorithms. The intuitive expectation is that the more sophisticated optimisers should find better strategies. Measured on out-of-sample survival, the opposite is true.
Template, the least clever engine in the building, survived about 23 times more often than DE. A plausible reading is that more expressive search finds it easier to fit noise, so it produces candidates that look better in-sample and generalise worse. This is exactly the failure mode that a search-heavy platform has an incentive not to publish.
| Engine | Forged | Still live | Survival rate |
|---|---|---|---|
| Template | 310 | 121 | 39.0% |
| Tree-GP | 982 | 137 | 14.0% |
| Linear-GP | 395 | 26 | 6.6% |
| NSGA-III | 836 | 49 | 5.9% |
| NSGA-II | 109 | 4 | 3.7% |
| MOEA/D | 314 | 11 | 3.5% |
| PSO | 416 | 9 | 2.2% |
| DE | 293 | 5 | 1.7% |
Live means the strategy is still on the shelf after every gate. Rates are computed on forged counts, not on candidates emitted.
Nearly half of what survives is the same strategy twice
A large search does not produce a large number of independent strategies. It produces one attractive region of strategy space, discovered repeatedly in slightly different clothing. If you count those as separate bets, you will badly overestimate how diversified you are, and every overfitting correction you then apply will be too optimistic.
We cluster the live shelf by return correlation and label rather than delete. Of the strategies we have checked so far, 45.9% are near-duplicates of another strategy we already hold.
| Redundancy status | Strategies |
|---|---|
| Distinct: passed the correlation check | 1,100 |
| Redundant: highly correlated with another | 934 |
| Not yet checked | 1,621 |
Nothing is deleted. Redundant strategies stay available and are labelled, because the redundancy rate is itself information about the search.
Verified on a real MetaTrader 5 terminal
A backtest that we computed ourselves is not independent evidence. Where possible we replay strategies on a real MetaTrader 5 terminal and compare against our own numbers. Only 46.2% of attempts reproduce cleanly. The rest are reported as they are, including the ones that diverge.
| MT5 verification result | Strategies |
|---|---|
| Verified: reproduced cleanly | 449 |
| Unsupported: symbol or timeframe not available | 397 |
| Weak: reproduced with material differences | 70 |
| Failed | 31 |
| Divergent | 14 |
| Mismatch | 11 |
Unsupported is not a failure of the strategy; it means the terminal could not run that symbol and timeframe. We report it separately rather than counting it as a pass.
What the survivors actually look like
This is the number most marketing pages lead with, and it is the one we trust least on its own. These are the 362 strategies that passed every gate, measured over the window in which they were selected. Because the selection and the measurement share a window, these figures are optimistic by construction and are not a forecast.
| Metric (live strategies, n=362) | Median | 25th percentile | 75th percentile |
|---|---|---|---|
| Total return | 77.9% | 56.4% | 103.0% |
| Sharpe ratio | 1.34 | 1.01 | 1.92 |
| Maximum drawdown | 11.4% | 8.7% | 17.7% |
Backtested, not live-traded. Historical results do not guarantee future performance.
What we do about it
- Baseline technical screening runs before a strategy is listed at all, so obviously broken candidates never reach you.
- Each additional validation is displayed separately instead of being collapsed into one score. You can require the checks you care about.
- Failures are labelled, not deleted silently. If a strategy fails an honesty gate, you can see that it failed and why.
- Untested is never painted as passed. A grey stage stays grey; it is never shown as green.
- Deflated Sharpe Ratio and Probability of Backtest Overfitting are shown as reference and never hidden.
- Strategies can be marked untrustworthy by design. That is a feature, not an error state.
What this data is not
This is one platform's output over one 30-day window, on metals and FX instruments. It is not an industry-wide study, and the absolute rates depend on our search budgets, our engines and our gates. If you ran a smaller search with stricter priors, your survival rate would be higher, and that would not mean your strategies are better.
The transferable finding is the shape, not the level: the gap between uncorrected and corrected significance, the penalty for search complexity, and the redundancy rate. Those three patterns will appear in any sufficiently large strategy search, including one you run yourself.
None of this is investment advice. Historical results do not guarantee future performance.
FAQ
- How many trading strategies survive backtesting?
- In our own production data, 362 of 3,655 unique forged strategies were still live after every gate, a 9.9% survival rate over a 30-day window. The exact rate depends heavily on search budget: the larger and more expressive the search, the lower the survival rate tends to be.
- What is the Deflated Sharpe Ratio and why does it matter?
- The Deflated Sharpe Ratio is a Sharpe ratio corrected for the number of strategies you tried before this one looked good. In our data the median PSR was 0.956 while the median DSR was 0.042 on the same strategies. Without the correction you would conclude almost everything works.
- Does a more advanced algorithm produce better strategies?
- Not in our measurements. The simplest engine we run survived at 39.0%, while differential evolution survived at 1.7%, roughly a 23x difference. More expressive search appears to find it easier to fit noise.
- Why publish data that makes your own product look bad?
- Because the alternative is to publish only the survivors, which would make our strategy lists unreadable and would hide exactly the information you need to judge them. A high failure rate is not a defect in a search process. Concealing it would be.
- Why were so many strategies eliminated for a constant signal?
- 421 strategies fired on 95-100% of bars, which means the condition was effectively always true and carried no timing information. These pass a naive fitness screen because they trade a lot, so they have to be caught by an explicit signal-health gate.
- Can I reproduce these numbers?
- The underlying counts are read directly from our production database and the figures shown here are computed from it. We state the window so the numbers can be re-derived and checked against the live transparency report as it moves.
More guides
Not investment advice. Historical results do not guarantee future performance. EasyQuant is a research factory — you execute on accounts you control.