Edge Finder starts from a different question than the rest of the Developer Assistant. The other tools help you validate a strategy you already have in mind. Edge Finder comes earlier: on this instrument, what kind of approach is worth looking for at all?
You give it a symbol and one or more timeframes. It runs hundreds of combinations — different families of approaches, each with its own parameter grid, on both sides — and gives you two levels of answer. First: which families show signs of life on that instrument and which are noise. Second: within the promising families, which specific probes hold up best, with the statistical metrics, the resulting equity curve, and EasyLanguage code to start from.
Edge Finder does not produce strategies. It produces research hypotheses measured under an anti-overfitting protocol, to tell you where your development time is worth spending.
An important note. QANTHOS is a tool for analysis, validation and discovery, not a financial advisor. Edge Finder does not prescribe strategies to trade: it measures the historical behaviour of generic approaches on an instrument and reports the statistical evidence. The same banner appears in the tab, above the results: research hypotheses, not strategies. Validate with full development and WFA. Every judgement and every operational decision remain yours.
1. The problem it addresses
Anyone who develops systematic strategies knows the situation. You pick an instrument, try a trend-following idea, it doesn’t work, you try another, then a mean reversion, then a pattern. After a few weeks something works — and at that point it is impossible to know whether you found an edge or simply drew the lucky result out of dozens of attempts.
This is the multiple testing problem, and it is why a raw Sharpe Ratio says almost nothing when it arrives at the end of a long exploration. The more combinations you try, the more likely one of them looks brilliant by pure chance. The backtest doesn’t know how many times you tried before; you do, but you rarely account for it.
Edge Finder inverts the approach: instead of hiding the number of attempts, it counts them and uses that count to correct the verdict. Every combination tried goes into a trial registry, and that registry feeds the statistical deflation of the final metrics. A result that survives this correction has said something; one that doesn’t was noise, and you know it immediately rather than after six months of development.
2. How it works: the protocol
The sequence is fixed and cannot be relaxed from the interface — that is the point: a protocol you can loosen protects you from nothing.
2.1 Chronological split with embargo
The data is divided in time order: the first portion is In-Sample (IS), the second Out-Of-Sample (OOS). The default is 0.70, meaning 70% of the data in IS, adjustable from the IS/OOS Split field in the parameter bar.
Between the two portions a few calendar days are discarded as an embargo. Without it, a trade opened across the boundary would live half in IS and half in OOS, and information would leak from one into the other. The embargo band removes the problem at the root: no bar and no trade can straddle it.
2.2 Selection on IS, judgement on OOS
Each probe runs across a parameter grid. The “best” combination in the grid is chosen looking at the in-sample only. The verdict on that combination is then formed looking at the out-of-sample only.
The separation is deliberate and worth understanding: if the winner were chosen by looking at the OOS, the OOS would stop being out-of-sample and become a second in-sample in disguise. It is the most common — and hardest to spot — mistake in this kind of research.
2.3 Minimum sample gate
A combination takes part in the verdict only if it produces enough trades on both sides of the split. Below that threshold it is not rejected: it is excluded, which is a different thing. Likewise, if too few daily observations are available, the Sharpe is not estimated at all and the metric comes back as not calculable.
This is a recurring choice throughout QANTHOS: better to state “I don’t know” than to return an authoritative-looking number computed on a sample that cannot support it.
3. The families of approaches
Probes are grouped into families, and the first results table reasons at that level. It answers the big question — what kind of behaviour does this instrument show? — before going into detail.
| Family | What it looks for |
|---|---|
| Trend Following | Directional persistence: a move that has started tends to continue. |
| Mean Reverting | Return towards a reference value after a departure from it. |
| Swing / Pattern | Recurring price configurations over a small number of bars. |
| Volatility | Behaviour tied to range expansion or compression. |
| Bias / Seasonality | Regularities tied to the calendar or the clock, over fixed time windows. |
The family verdict is the best among the verdicts of its probes, each judged on its own parameter grid. The choice is not obvious and deserves a note: pooling different probes and opposite sides into a single distribution would mask a real but isolated edge, drowning it in the median of everything else.
4. The four verdicts
There are four levels, in decreasing order of strength. The wording is deliberately cautious: none of these verdicts means “winning strategy”.
| Verdict | What it says |
|---|---|
| PROMISING | Solid OOS evidence across every criterion: clearly positive deflated DSR, Profit Factor above threshold, entry quality confirmed by the E-ratio. |
| INTERESTING | Evidence is there but below the top level. Worth investigating, not worth excitement. |
| WEAK | Positive DSR, but at least one criterion fails — typically the Profit Factor or the E-ratio. |
| NOISE | No evidence. Includes the most treacherous case: high DSR but a non-positive grid median, meaning an isolated peak in a sea of negative results. |
Two mechanisms adjust the verdict after it is first assigned.
The instability cap. The OOS is split into segments and the number of segments where the Sharpe stays positive is counted. If coherence is insufficient, the verdict drops by a single step: PROMISING to INTERESTING, or INTERESTING to WEAK. Never two steps at once, and never below WEAK. The reasoning stated in the code is that a strong but time-concentrated edge remains a candidate worth investigating, not something to discard. If fewer than two segments are estimable, coherence cannot be measured and the verdict is left untouched — it is only annotated.
Costs never reject a candidate. This is an explicit methodological principle. DSR, Profit Factor and Sharpe are computed on gross P&L. The reason: Edge Finder looks for hypotheses, and a hypothesis that doesn’t cover costs today may become tradable once you build development filters on top of it. Rejecting it now would mean throwing away raw material because it isn’t yet a finished product.
5. The metrics
| Metric | What it is | How to read it |
|---|---|---|
| DSR OOS (gross) | Deflated Sharpe Ratio: the out-of-sample Sharpe corrected for the number of attempts made, for skewness, and for the tails of the return distribution. | The headline metric. It answers “how likely is it that this Sharpe is not an artefact of how many trials were run”. |
| DSR excluding top-3 | The same DSR recomputed after removing the three trades with the highest gross P&L. | The comparison against the headline is everything. A DSR that collapses without its three best trades is not an edge — it is three lucky hits. |
| Median OOS SR | The median out-of-sample Sharpe across the entire parameter grid for that probe and that side. | Separates a plateau from a peak. A non-positive median is a veto: the verdict becomes NOISE even with a brilliant best. |
| IQR | Interquartile range of the same distribution. | Measures how much the result depends on the parameters chosen. A wide IQR means behaviour changes sharply as parameters move. |
| E-ratio | Ratio of mean favourable excursion to mean adverse excursion in the bars following entry, both normalised by the ATR computed at the moment of the signal. | Measures the quality of the entry, separately from exit management. A value around 1 says the entry has no predictive value — and no stop or target can fix that. |
| Headroom | Gross average trade divided by the cost of one round turn. | Informational, never a criterion. Below 1 it means development filters will have to more than double the average trade for the hypothesis to become tradable. |
| N trials | Total number of combinations evaluated in the run. | This is the number that drives the deflation. The higher it is, the harsher the correction applied to every DSR. |
For those who want the detail. The Sharpe feeding the DSR follows the convention required by the original formulation: computed on daily returns reconstructed from the equity, not on per-trade P&L, and not annualised — the observation count is in days. The annualised value shown elsewhere is the same Sharpe multiplied by the square root of 252, and exists purely for readability. The Wilder ATR used to normalise the E-ratio is the one computed at the signal, so it depends only on data already available at that moment.
6. From verdict to individual combination
The tables step down in scale: from the family, to the probe with its side, down to the single combination of parameters and timeframe. The last level lists in-sample and out-of-sample trades, the gate outcome, the selected exit horizon and the metrics for that specific row.
From here the detail window opens with the equity curve. A vertical line marks the boundary between in-sample and out-of-sample, with the embargo band highlighted: everything to the right of that line is material the selection never saw.
One element of the panel deserves explanation. Exit N* is the exit horizon in bars chosen by the system as the one that maximises the E-ratio — and it is chosen looking at the in-sample exclusively. The same discipline as in 2.2 applies: if N* were chosen on the OOS, the OOS result would stop being independent.
7. EasyLanguage export
From a combination’s detail view you can generate the corresponding EasyLanguage / PowerLanguage code, ready to paste into MultiCharts or TradeStation.
It is worth being precise about what is being exported. The code reproduces the probe’s entry and exit logic with that combination’s parameters: it is a starting point for development, not a finished strategy. It contains no context filters, no risk management, no money management, and its metrics were computed gross of costs.
The natural path is: export, take the code to the platform, build on top of it what is needed, and at that point — and only then — put the result through the normal validation process.
8. Flow validation
Above the results tables sits a collapsible panel called Flow validation, containing probes whose outcome is known in advance: always long, always short, a random probe with a fixed seed, and an oracle that deliberately peeks at the next bar.
They are not looking for an edge. They verify that the pipeline is working: the oracle, which cheats, must come out dominant on the shortest horizon. If it didn’t, the problem would be in the calculation chain, not in the markets.
These probes are excluded from the verdicts and do not enter the trial count that drives the deflation — otherwise they would distort the correction applied to everything else.
9. Common mistakes
Reading the best and ignoring the median. The most frequent and the most expensive mistake. A high DSR with a negative grid median describes an isolated peak: move one parameter by a notch and the edge disappears. The two columns must always be read together.
Mistaking a hypothesis for a strategy. A PROMISING verdict says that on that instrument this kind of approach deserves development. It does not say the probe as it stands is tradable — it has no filters, no risk management, and its metrics are gross.
Treating headroom as a gate. By construction it isn’t one. Low headroom describes how much work the filters will need to do, not a rejection.
Forgetting that N matters. Adding timeframes and probes makes the exploration broader but also stricter: N grows, deflation grows, and the DSR of every combination comes down. This is not a flaw, it is the correct price of a wide search.
Changing the split until the numbers work. If a combination only becomes interesting at a particular IS/OOS Split value, the useful information is not the new result: it is that the combination is fragile.
10. Workflow
- Enter the symbol and select the timeframes to explore.
- Check the IS/OOS Split, any round-turn commission override, and the worker count.
- Start the analysis and wait for it to finish; the number of trials run is reported at the end.
- Read the family table: which kinds of approach show signs of life on this instrument.
- Step down to the probe table within the promising families, always comparing DSR against the median.
- Open the detail view of interesting combinations and examine the equity across the IS/OOS boundary.
- If the hypothesis convinces, export the code and continue development on the platform.
11. Further reading
- Surface Explorer — once the hypothesis has become a strategy with parameters of its own, check whether those parameters sit on a plateau or on a ridge.
- Bias Explorer — goes deeper into the calendar and clock regularities that appear here as the Bias family.
- Portfolio Rules — where strategies developed from these hypotheses get selected and put into a portfolio.