The Edge Ratio: Does a Strategy's Raw Edge Predict Out-of-Sample Survival?

Every backtest report ends in a single number: net profit. It feels final — the strategy made money, or it did not — but net profit is not one thing. It is the tangled outcome of at least two separate skills: whether your entry signal caught a genuine directional move, and whether your exit and money management then managed to bank it. Optimise net profit directly, as most strategy builders do by default, and you cannot tell which of the two you actually improved. You might have found an entry with real predictive power and wrapped it in a stop-loss that quietly bleeds it dry. Or — more dangerously — you might have found an entry with no edge at all, and disguised that fact behind an exit so finely fitted to the historical data that it managed to monetise pure noise. Both strategies can produce an identical-looking equity curve in the backtest report. Only one of them has any right to expect that curve to continue.

Is the entry earning its keep, or is the exit doing all the work?

That question predates most of what passes for strategy analytics today. In Way of the Turtle, Curtis Faith — one of the original Turtle traders — described a way to measure an entry signal in isolation, stripped of exits, stops and position sizing entirely. He called it the Edge Ratio, or E-ratio, and it is one of the first things I look at in StrategyQuant X, before I let an optimiser anywhere near a strategy’s exit logic.

What the Edge Ratio actually measures

The method is simple enough to compute by hand, though obviously not at the scale a generator produces strategies. Take every entry signal a strategy fires — long or short, it makes no difference — and, for a fixed number of bars after each one, track two numbers:

Maximum Favourable Excursion (MFE) — the best unrealised profit the trade reaches at any point during that window.
Maximum Adverse Excursion (MAE) — the worst unrealised loss the trade reaches at any point during the same window.

Neither number cares what your actual exit rule does with the trade. A position that gets stopped out for a small loss in the backtest can still show a large MFE, if price wandered nicely in its favour before wandering back again — and that is exactly the point. You are measuring what the market offered the signal, not what your exit chose to take from it.

Raw price moves are not comparable across instruments or volatility regimes, so each excursion is normalised by the Average True Range (ATR) at the time of entry — turning a raw price move into a volatility-adjusted figure you can compare between EURUSD and the S&P 500, or between a quiet year and a violent one, on equal terms. The E-ratio itself is then just a ratio of averages, taken across however many entries the strategy produced:

E-ratio = average(MFE ÷ ATR) ÷ average(MAE ÷ ATR)

An E-ratio above 1 means that, on average, price travelled further in the trade’s favour than against it during the window you measured — the entry has a genuine directional lean. An E-ratio at or around 1 means the signal is, for these purposes, indistinguishable from a coin flip: price wanders as far against the position as for it, and any profit your backtest shows is being manufactured entirely by the exit, the stop placement, or luck.

The choice of how many bars to measure over is not incidental. Too short a window and you are really just measuring the first tick or two of noise around the entry; too long and you start picking up the influence of whatever the broader trend was doing regardless of the signal. A sensible horizon is one that roughly matches how long the strategy intends to stay in a trade — which is also why the E-ratio works as well for a mean-reversion scalper as it does for a position trader, provided you measure each on its own natural timescale rather than borrowing someone else’s.

Why bother separating the two

The reason this is worth the extra step is that exits are extremely good at hiding a weak entry. Give a curve-fitting optimiser enough freedom over stops, targets and trailing logic, and it will usually find a way to extract a positive net profit from almost any entry condition on in-sample data — including one with no real edge behind it at all. The exit becomes an expensive, fragile patch over a signal that was never predictive to begin with, and patches like that tend to fail exactly when you need them, which is out of sample.

The E-ratio does not care about any of that, because it is computed before the exit logic ever touches the trade. It grades the entry on its own merits — what the market actually did after the signal fired, not what your rules did with the opportunity it created. A strategy with a strong E-ratio and mediocre net profit tells you something you can work with: the entry has real information in it, and the exit is the part that needs attention. A strategy with a weak E-ratio and excellent net profit tells you something to be suspicious of: the results are being carried entirely by exit engineering, which is precisely the kind of result that tends not to survive contact with new data.

I see this constantly when reviewing batches of machine-generated strategies. Two candidates arrive with near-identical net profit curves: one built on a genuine signal wrapped in a mediocre exit, the other built on nothing wrapped in an exit clever enough to hide it. Net profit cannot tell those two apart — by construction, it grades the finished product, not the ingredients. The E-ratio can tell them apart, because it never looks at the exit at all.

Does raw edge survive into out-of-sample?

That last claim — that a strategy with a strong E-ratio should be more likely to hold up beyond the data it was built on — is a hypothesis, not a certainty, and I wanted to know whether it actually held before I started relying on it. So I set up a direct test: does a strategy’s in-sample Edge Ratio actually correlate with how it performs on true out-of-sample data, or is it just a plausible-sounding statistic that happens to feel rigorous?

Setup. I worked with strategies built and evaluated on two markets: EURUSD on the 60-minute chart, and the E-mini S&P 500 futures contract (ES), also on the 60-minute chart. For each strategy I recorded the in-sample Edge Ratio alongside its performance over a true out-of-sample period — data untouched by either the generation or the selection process — and measured the Pearson correlation between the two.

Result. On both markets the correlation was positive: strategies with a higher in-sample E-ratio tended to go on to perform better out of sample than strategies with a lower one. On EURUSD it was a real but fairly modest relationship. On the S&P 500 contract it was considerably more pronounced — a fairly strong positive correlation between in-sample edge and what the strategy actually went on to do on data it had never seen.

My own reading of that gap is that it is plausible rather than surprising, not that it is explained. A liquid index future like the E-mini S&P 500 is traded by a different mix of participants, under different microstructure, than a spot forex pair like EURUSD, and it would not be shocking if a raw directional-edge measure travelled more cleanly through one than the other. But that is interpretation on my part, not something the correlation itself proves — the study measured a relationship on two specific instruments, not the mechanism behind it.

I want to be careful about what this does and does not show. It is one study, on two instruments and one timeframe each, over one particular slice of history — not a universal law of markets. A correlation, even a fairly strong one, is not proof that a high E-ratio causes good out-of-sample results, and it is no guarantee that the relationship holds with the same strength on every symbol, timeframe or asset class you might apply it to. The honest reading is narrower than a sweeping claim, and more useful because of it: on the two markets I tested, raw entry edge measured in-sample carried real information about what happened afterwards, out of sample. That is enough to justify using it as evidence. It is not enough to justify trusting it blindly.

How I use it in practice

None of this makes the E-ratio a replacement for true out-of-sample testing, walk-forward analysis or Monte Carlo robustness checks. It is a filter I apply earlier, before those slower and more expensive checks, to decide which strategies out of a large batch deserve the time.

Treat it as a ranking tool, not a pass or fail gate. I do not discard every strategy with an E-ratio under some arbitrary threshold. I use it to rank a batch of generated strategies against each other and prioritise the ones with real entry edge for the deeper testing that follows.
Read it alongside net profit, not instead of it. A high E-ratio paired with disappointing net profit is often the more interesting strategy, not the worse one — it usually means the entry is sound and the exit logic is what needs reworking, which is a far easier problem than inventing a new entry from nothing.
Match the measurement window to how you actually trade. The number of bars you track MFE and MAE over should sit in the same neighbourhood as your intended holding period. An E-ratio measured a few bars after entry tells you very little about a strategy meant to hold for weeks.
Recompute it across market regimes if you can. An entry that shows a strong E-ratio only in a trending sample and collapses in a choppy one is telling you something about when it works, not just whether it works — treat that as extra information, not noise to average away.
Do not let it replace true out-of-sample validation. The correlation I found was real, but it was a correlation, not a certainty. The E-ratio earns a strategy a closer look. It does not earn it a live account.

Used this way, it pays for itself early in the pipeline — a fast, exit-agnostic sanity check that tells you, before you have spent any optimisation time on stops and targets, whether there is something in the entry worth building around at all.

The point

Net profit tells you what happened. The Edge Ratio tells you which half of the strategy is responsible for it. That distinction matters more than it sounds, because the two halves fail for different reasons: a weak entry fails because the market never gave it anything to work with; a weak exit fails because the strategy was handed a real opportunity and mismanaged it.

My own testing suggests this is not just a tidy theoretical idea. On EURUSD and the S&P 500, in-sample edge measured this way carried real signal about what a strategy went on to do out of sample — modestly on one, more convincingly on the other. That is not a promise, and it is not a substitute for the validation steps that come after it. But as one more evidence-based filter for deciding which strategies out of a generated batch deserve your capital and your attention, it has earned a permanent place in how I evaluate a system before I trust it with either.

If you would like a second opinion on whether a strategy’s edge is real before you commit capital to it, get in touch.