Compiled from the source rather than from the marketing pages. Where a figure appears below it was read out of the code that produces it.
Contents:
Edge Matrix takes a finished backtest, or a live account statement, and reports whether its statistical properties were consistent with a real edge. It does not generate strategies, does not optimise parameters, and does not place trades.
The distinction that shapes everything else: a strategy tester answers how a strategy performed on one sequence of trades. Edge Matrix answers how much of that performance could have been luck.
Accepted formats
| Source | Format | Endpoint |
|---|---|---|
| MT4 strategy tester | HTML report | /import/mt5 |
| MT5 strategy tester | HTML report | /import/mt5 |
| cTrader strategy tester | HTML report | /import/mt5 |
| cTrader History tab | XLSX workbook | /import/xlsx |
| MT4 / MT5 live or demo account | HTML statement | /import/mt5 |
| MQL5 signal page | CSV position export | /import/csv |
| Myfxbook | CSV export | /import/csv |
| Trading journals | CSV with a profit column | /import/csv |
One control, automatic detection. A single drop zone accepts every supported format. Files can be dragged onto it or chosen through a dialog, and routing is decided by the file itself rather than by which control was used.
Live and demo accounts, not only backtests. MT4 account statements and MT5 trade history are read directly. The dashboard reports absolute gain, the true maximum drawdown and the time-weighted return, the last of these because a deposit part way through an account's life otherwise flatters every simple return figure. The full statistical suite runs unchanged on a live record, and the integrity checks stay active, since a statement can be incomplete or unrepresentative just as a backtest can. Report wording adapts to the loaded account history rather than referring to a backtest.
Cash flows are separated from trades. Deposits, withdrawals and transfers are identified and excluded from the trade series. Without this, a deposit reads as a very large winning trade and every statistic downstream is wrong. On accounts funded over time, performance is measured against the capital actually at risk at the time of each trade rather than the full deposited total, which is how a broker or a track record service reports the same account.
The cTrader XLSX case, and why it asks for a balance. That export lists trades but carries no running balance, unlike the HTML report. Equity therefore has to be accumulated onto a starting figure that is not in the file. Every percentage scales with it: on one test file, assuming 10,000 when the account held 1,000 turned a 5.95 percent drawdown into 0.77 percent. The import asks for the figure, and if it is skipped the dashboard states on the result that the percentages rest on an assumed balance.
Truncation detection. Where the source carries header metadata, the reported trade count is compared against the parsed count, so a report that was cut short is flagged rather than analysed as if complete.
Import Data the parsed record, headline statistics, and the equity curve.
Projection resamples the strategy's own trade distribution forward over a chosen period and returns best, median and worst paths with confidence bands rather than a single line.
Optimal Period estimates how much historical data is worth re-optimising over, three separate ways: statistical (rolling variance stability), benchmark (asset class and timeframe), and spectral (memory in the return series), with a regime read alongside. Where the three agree the window is stable; where they disagree, that disagreement is the finding.
Monte Carlo see section 5.
Edge Matrix the 19 scored tests plus the plausibility gate, with a 0 to 100 score, per-test results and a written assessment for each.
Insights see section 8.
Compare up to ten strategies side by side on identical tests.
Report assembles the tabs already run into one document.
Nineteen scored tests, split 9 core and 10 advanced, plus one gate that is not averaged in. Each returns its own score and a written verdict rather than disappearing into a single number.
Core
using López de Prado's minimum track record length computed from the strategy's own expectancy and higher moments rather than a fixed rule of thumb
Advanced
with a Gini-style concentration
underwater series
Sharpe of 2.0 with lag-one autocorrelation of 0.3 carried more risk than a Sharpe of 1.5 near zero
VaR alone
The gate: Statistical Plausibility
Asks whether the combination of win rate, reward-to-risk and trade frequency is consistent with a durable edge at all. It computes the annualised Sharpe those figures imply and compares it against the highest verified track record on record. Results beyond that bound are flagged as more consistent with overfitting than with a durable edge.
It runs as a gate rather than being averaged in, because a curve that looks smooth on every other test would otherwise outvote its own detection. Every input, the formula and the sources are shown in the app and the report so the calculation can be checked rather than trusted.
It carries a diversification exemption. When a merged portfolio's Sharpe sits inside the mean-variance ceiling implied by its own correlation matrix, the elevated precision is honest aggregation across low-correlation sleeves and the flag is suppressed. Most implausibility screens would misfire on such a portfolio.
When the gate does engage, gated strategies are mapped into a capped band that preserves their ranking against each other rather than collapsing to one number, so a strategy that is only implausible sits above one that is implausible and weak everywhere else.
The test measures statistical plausibility. It is not a determination of intent.
Up to 50,000 simulations. Three resampling methods, and the difference between them is the point.
Permutation reshuffles the exact set of trades. Every trade is used once, so the final balance is identical on every run and only the path changes. Isolates sequence risk: was the smooth ride real, or helpful ordering?
Bootstrap draws a fresh sample with replacement. Some trades appear twice, some never. The final balance varies, which is the point. Answers how much rested on the particular trades that occurred. Assumes trades are independent.
Block bootstrap also draws with replacement but lifts runs of consecutive trades whole, so losses that bunched together can still bunch together. The honest one when a strategy shows volatility clustering, which most do.
Two methods can be compared in one run. Bootstrap against permutation separates sequencing luck from sampling luck. Bootstrap against block isolates what clustering cost, since the only difference is whether runs were kept together.
Block length is read from the data, not chosen. Selected automatically by the Politis and White procedure with the Patton, Politis and White correction, run on squared returns because drawdown is driven by clustering in the size of moves rather than dependence in the average trade. The chosen length is reported with the result. When a strategy shows no clustering the length falls toward one trade, and the result panel states that block bootstrap described the same thing as bootstrap rather than implying a difference that is not there.
Days are pooled from the calendar, not from trade dates. The pool of days available for resampling includes non-trading business days rather than only days on which a trade closed. Without this, a low-frequency strategy has its timeline compressed dramatically, because a record spanning ten years is replayed as if it occupied only the days it happened to trade.
Outputs drawdown at the 50th, 95th and 99th confidence levels, a robustness score and grade, risk of ruin, percentile rank, the share of profitable simulations, and a fan chart showing the original backtest against the percentile envelope.
The score is computed from the same method every time regardless of which is selected for display, so strategies stay comparable.
Compare up to ten strategies on identical tests, so ranking comes from the same measurements rather than from separate reports read in sequence.
Portfolio Mode merges up to twenty strategies and analyses them as one account. Requires at least two, since a single strategy is not a portfolio.
Correlation Pearson correlation of daily profit and loss across the union of close-days, with non-trading days counted as zero rather than skipped. This is the number that decides whether a portfolio diversified anything, and it is invisible when strategies are validated one at a time. Each pair can be opened for a scatter of daily results, an overlap percentage and a plain-language reading.
Correlation is computed on demand rather than on load, so users who never open it pay no computation.
Breaks the trade record down by weekday, session and month, and surfaces the combinations where the strategy lost consistently rather than occasionally.
Views Overview, Breakdown, Calendar. Configurable GMT offset and minimum sample size per bucket.
Reports the sample size with every finding, so a pattern resting on eight trades is visibly resting on eight trades.
Generated filter code. When a pattern is strong enough to act on, the filter condition is returned in the syntax of the platform the file came from, detected automatically: MQL4, MQL5, or C# for cTrader. It is a condition to drop into entry logic, not a finished function.
The caution that belongs with it: every filter added from in-sample data is another parameter fitted to the past, so the honest move is to re-validate afterwards and expect the score to move.
Eight self-contained HTML documents, each rendering without external assets:
| Report | Contents |
|---|---|
| Full validation | every tab that was run, assembled |
| Edge Matrix | the 19 tests, the gate, score and written assessments |
| Monte Carlo | simulation results and method explanation |
| Insights | overview, breakdown and calendar |
| Compare | multi-strategy ranking |
| Portfolio | merged account with correlation |
| Analysis | optimal period and projection, also available as PDF |
Reports explain their own method. The Monte Carlo section sets out which resampling method ran and why it suited the data, and shows the figures behind the automatic block length choice: the selected length, lag-one autocorrelation of trade results, the same measure on squared results, and the threshold each was judged against.
Reports adapt to the finding. A flagged strategy does not read as an endorsement anywhere in the document, including the abstract and the closing recommendation.
Reports report rather than praise. Figures are stated relative to a reference range without describing the strategy as exceeding or outperforming, because a report that argues for the strategy it is measuring is worth less to the person reading it.
Publishing. Any report can be published to a link that anyone can open with no account and no sign-in. The stored copy is frozen at the moment of sharing, so the figures a link shows today are the figures it shows in a year, and it keeps working after the session ends. Stored compressed and served compressed, so a burst of views costs one indexed lookup each and no computation.
Every published page states that the report was self-published, because a report chosen by the person selling the strategy is selected evidence, and a reader who has to work that out for themselves discounts everything else too.
Withdrawing. A link can be withdrawn at any time from a list in the app that also shows view counts. Withdrawing stops the page loading but leaves the reference resolvable, still reporting what was published. Content stays under the publisher's control; the record of it does not. A report shown to someone and later withdrawn cannot afterwards be denied.
Verification at /verify. Every report carries a reference. Entering it returns the strategy, date and headline figures recorded when that report was published, so a figure in a circulating copy that does not match was changed afterwards, and an unknown reference was never published.
Three properties keep the lookup from becoming a way to read other people's results. It answers only for reports whose publisher chose to make them checkable. Unknown, malformed, unlisted and withdrawn references return the same response, so nothing reveals which references exist. And the reference carries a random suffix, because the previous format was a plain timestamp with 86,400 values a day and could have been swept.
First validation free, no payment method requested. Not a demo: all tests, Monte Carlo at full iteration count, Insights, Compare, Portfolio Mode and every report, on a backtest of the account holder's choosing.
Credit packs, one-time purchases with nothing renewing. Packs are sold in several sizes, and the larger the pack the lower the cost per validation. Current prices are on the pricing page rather than here, so this document does not go out of date the moment they change.
Every pack unlocks the same application. There is no cheaper version and no feature reserved for a higher tier: the only difference between packs is how many strategies can be validated.
A credit is charged once per new backtest. Everything afterwards on that backtest is free and unlimited: every test, every Monte Carlo run, every Insight, every Compare, every Portfolio, every export, and loading the same file again. A backtest is identified by its trade data, so re-uploading the same result is recognised as the same strategy, while the same expert advisor over a different period produces different trades and counts as new.
The balance and what it has already bought are both visible. The tab bar shows the remaining balance for accounts using credits and hides it entirely for subscribers, who are never metered. Opening it lists every backtest the account has already paid to validate, with instrument, timeframe and when it was validated, each marked as free to load again. An import states whether a credit was spent or whether the file was already owned, rather than leaving a change in the balance unexplained.
Credits are never spent on a run that cannot complete. Portfolio Mode needs at least two validated strategies, and selecting more files than the balance covers is refused before anything is uploaded rather than after a credit has moved.
Subscriptions remain available for unlimited validation. A subscriber is never metered and the credit interface is hidden for them entirely.
Credits do not expire. A pack bought today can sit unused for a year and still be spent, which is the point of paying per strategy rather than per month: nothing is lost to a period in which you did not happen to be validating anything.
Hide EA name. A toggle replaces the strategy's real name with an anonymous serial everywhere: on screen, in every report, and in download filenames. Set before loading, the real name never appears at all. Intended for sharing screenshots, recording video, or submitting a report to a prop firm without revealing which strategy produced it.
The anonymous serial is generated at parse time and stays stable across every tab and report for one import.
The free browser tools upload nothing. Analysis runs entirely in the browser and the file never leaves the machine.
Free Monte Carlo simulator accepts MT4, MT5, cTrader and trading journal exports, runs 1,000-path simulation with drawdown confidence bands and a robustness score, entirely in the browser with no account.
The analysis, import, report and publishing endpoints. Analysis: /analyze/tests, /analyze/edge-matrix, /analyze/monte-carlo, /analyze/projection, /analyze/analysis, /analyze/insights, /analyze/compare, /analyze/portfolio-sessions, /analyze/portfolio-correlation.
Import: /import/mt5, /import/csv, /import/xlsx.
Reports: /report/html, /report/em, /report/insights, /report/compare, /report/performance, /report/portfolio, /report/analysis, /report/analysis/pdf.
Publishing: /report/share, /r/{token}, /report/shares, /report/share/{token}, /verify.
Account: /auth/me, /credits/balance, /credits/strategies.
relationship underlying the plausibility gate
Every test names the method it implements and the source it came from, so an implementation can be checked against the paper rather than taken on trust. The parameter choices and scoring weights are not published.
These belong in the documentation rather than in small print.
happened, or resamples of them. A 99th percentile drawdown is a statement about a historical distribution, not a floor.
advice.
checks built to catch specific failure modes, which is a weaker claim than it sounds and stays weaker than a live track record.
only the statistical fingerprint if one reached the results.
poor tick data is not rescued by analysis, and the appearance of rigour makes that worse rather than better.
says nothing about the reports that were not published.
← Back to guides