What happens when you add a losing EA to your portfolio

The first article in this series showed our ten template strategies returning +272.59% at the same drawdown the average single strategy suffered, on an average pairwise correlation of 0.014. The obvious objection came immediately, and it is a good one:

"You picked ten strategies that worked, on the same history you measured them on. Of course they look good together."

This article answers that objection with a test the first one could not run: a universe of 322 strategies chosen before looking at whether they made money, and portfolios assembled at random out of it. Then, the practical version of the same worry — what happens when one or two of your EAs stop performing the way the backtest said they would.

The universe: 322 strategies, nothing filtered out

Thirteen strategy families — Supertrend, EMA crossover with ADX, RSI pullback, Donchian breakout, MACD, Bollinger reversion, Bollinger trend, Parabolic SAR, Stochastic, Williams %R, CCI, MFI, Ichimoku — expanded into 23 parameter variants and run across 14 markets on H1. Every one keeps its result, including the ones that lost half the account.

• 133 profitable, 189 unprofitable. Pick one of these at random and you are losing money 59% of the time. That is the honest base rate for "indicator strategy with a fixed ATR stop and no selection", and it is the number the rest of this article is measured against.
• Median winner: +21.4%. Median loser: -21.1%. Worst: -51.2%.
• Across all 51,681 pairs, mean correlation of daily returns is +0.018, median +0.008. So the orthogonality result from the first article survives without any cherry-picking — these strategies genuinely do not move together.
• The most correlated pair sits at +0.910 — the same market and the same family, differing only in the take-profit multiple. Adding that as your "second strategy" adds nothing.

Experiment 1: portfolios built at random get worse, not better

Draw N strategies from the universe at random, run them together at 1% risk each, 400 draws per value of N. No selection of any kind.

Read the last column first, because it is the only fair one. Return alone is not comparable across these rows — running 12 strategies at 1% each is simply more leverage than running 1, which is why the drawdown column climbs from 18.9% to 62.5%. The last column removes that: every portfolio is scaled until its max drawdown is exactly 20%, and then we compare returns.

Scaled that way, the answer is flat and negative — around −4% to −8%, no matter how many strategies you add. Meanwhile the probability of losing money rises from 57.2% to 79.8%, and median Sharpe drifts from -0.02 to -0.21.

If diversification were insurance, that column would rise. It does not.

The control: same experiment, filtered pool

Now change exactly one thing. Draw from the 65 strategies that were profitable with at least 60 trades, and repeat the identical procedure.

Same sampling, same equal-risk scaling, opposite outcome. Adding legs did not make the losing pool profitable, and it did not need to make the winning pool profitable — it multiplied what was already there.

Why: it is √N, and you can see it in the numbers

For roughly uncorrelated legs, combining N of them multiplies expected return by N and volatility by only √N — so the ratio of the two improves by √N. Nothing in that derivation cares about the sign of the expected return. A negative edge gets amplified exactly as efficiently as a positive one.

Take the single-leg Sharpe from the filtered pool (0.76) and multiply by √N:

Nearly exact at 2 and 3 legs, then drifting increasingly below the prediction. That gap is the price of the residual +0.018 correlation — real strategies are not perfectly independent, so the benefit decays as you stack legs. Which is the honest version of "how many strategies should I run": the first four buy you most of it.

Experiment 2: seeding a good portfolio with genuinely losing strategies

This is the question that prompted the article. Build a 6-leg portfolio from the winners, then replace k of those legs with strategies drawn from the 93 in the losing pool — real strategies whose own backtests lost a median of -21.1%.

Three things worth separating out:

• One bad leg out of six costs you more than half the return — from +162.7% to +73.7% at equal risk. That is a much bigger hit than "one sixth", because the bad leg is not merely contributing zero, it is actively subtracting.
• But it is still roughly twice as good as running one good strategy alone (+34.1%). One rotten leg does not undo the portfolio.
• Two bad legs and you are past the point. +22.7% is worse than just running a single screened strategy, and the probability of losing money over two years jumps from ~0 to 11.0%. At three, it is 71.0%.

Now look at the worst day column, which barely moves: -7.56% with no bad legs, -7.54% with three. This is the sharpest way to state what diversification does and does not do. The bad legs were diversified perfectly well in the variance sense — they did not make your worst day worse, because they still were not losing on the same days as everything else. They simply were not diversifiable in the mean sense. Correlation is about variance. Expectancy is not something correlation can fix.

Experiment 3: legs that stop working, rather than legs that were always bad

Experiment 2 is the pessimistic case: the leg has a genuinely negative edge. The more common real-world failure is softer — the strategy keeps trading, keeps taking the same risk, but the edge that showed up in the backtest is simply gone. So: take a 6-leg winning portfolio and subtract each chosen leg's mean daily return, leaving its volatility and its correlation structure untouched. The leg still trades, still swings the account, and now earns exactly nothing.

This is the result that justifies running a portfolio in the first place:

• One dead leg out of six: +104.5% at equal risk — still three times a single strategy.
• Two dead: +66.9%, still twice.
• Three dead — half the portfolio: +40.7%, still ahead of +34.1%.
• Four dead: +19.8%, and still profitable in 98.8% of draws.
• All six dead: -3.2% — noise, as it must be. There is no free lunch hiding in the correlation structure.

Compare that to the concentrated alternative. If you run one strategy and it is the one that stops working, you are at zero — there is no "four out of six" for a portfolio of one. Experiments 2 and 3 bracket the realistic range: a degraded leg costs you somewhere between the Experiment 3 row (edge gone) and the Experiment 2 row (edge inverted).

Experiment 4: pick with last year's data, live through next year

Everything above shares one flaw — the same one this article accused the first article of. Experiments 2 and 3 damage a leg artificially: I chose how badly, knowing the whole two years. And the filtered pool in the control was filtered on the window it was then measured over.

So here is the clean version, and it is simple enough to state in one line:

Why select on correlation rather than on profit? Because of the finding above: profit did not carry forward at all. Of the 152 strategies that were profitable in year one, only 42.1% were still profitable in year two (34.5–50.1% at 95% confidence — an interval that contains 50%), and in-sample Sharpe predicted out-of-sample Sharpe at -0.034, p = 0.58. Ranking by backtest profit is ranking by noise. Correlation is the one measurement that held.

The candidate pool is those same 152 in-sample winners, and all three selection rules draw from it, so the only thing that varies is how the legs are chosen:

• Lowest correlation — greedily pick the N legs with the lowest average pairwise correlation, using year-one data only
• Highest correlation — the same procedure, run backwards
• Random — ignore correlation entirely

Each is repeated 300 times; every round first draws 40 random candidates and then selects within them, so what comes out is a distribution rather than one lucky portfolio.

The results

Read the two correlation columns first, because they are the reason the rest of the table is meaningful. What was selected in year one stayed selected in year two: the low group was picked at -0.089 and came out at -0.055; the high group was picked at +0.356 and came out at +0.319. The selection rule works. The question is what it buys.

It buys risk reduction, consistently and increasingly:

Comparing within each round — same candidate subset, only the selection rule differs — the high-correlation portfolios drew down more in every case, and the gap widened with size: 5.7 percentage points at 3 legs, 11.5 at 5, 18.1 at 8. Not one of those intervals touches zero. Worst single day tells the same story: at 8 legs, -6.80% for low correlation against -11.34% for high.

What it does not reliably buy is return. At 8 legs the low-correlation portfolios did finish +16.4 points ahead — but at 3 legs they finished -15.5 points behind, and at 5 the gap was -0.1 with an interval straddling zero.

The 3-leg reversal is worth understanding rather than explaining away. Three maximally correlated strategies are, in practice, three bets on the same market. Out-of-sample that year, that concentrated bet happened to be right — which is exactly what concentration does: it widens the distribution in both directions. It cost 31.19% drawdown against 24.44%, and by 8 legs, where concentration can no longer diversify anything, the same rule finished -27.4% against -13.0%.

The curves

One round of the 5-leg test, drawn out. This is the round whose return was the median of the 300, not the best one:

The high-correlation portfolio was five GER40 strategies — MACD at two different targets, SAR, Williams %R, and an RSI pullback. Five charts, five EAs, five sets of spread, one bet. The low-correlation portfolio spread across US30, NAS100, XAUUSD and BTCUSD.

And the blue line deserves a moment: that is the single strategy with the highest in-sample Sharpe in the entire pool — 2.62, the one any reasonable person would have picked. It returned -2.82% out of sample. It is also, by a wide margin, the smoothest line on the chart, at 11.57% drawdown against 31.76% and 38.31%.

Which is the honest summary of the whole experiment. Selecting on correlation gave us control over risk and no control over return — and that is not a disappointing result, it is the result you should expect once you know that correlation persists and profit does not.

What did survive: risk, not return

If a backtest cannot predict next year's return, the obvious question is what — if anything — it can predict. So we ran the same split on every number a backtest produces: measure it in year one, measure it again in year two, and see how well the first predicts the second. Across all 291 strategies (and all 42,195 pairs, for the correlation row):

Two clean groups, and the line between them is not subtle. Everything about risk carried forward. Nothing about profit did. How badly a strategy can hurt you on its worst day was the single most predictable thing after correlation; how much it made was indistinguishable from noise.

The correlation row is worth one more look, because it is what the rest of this series is built on. Mean pairwise correlation was +0.0191 in year one and +0.0190 in year two — the same number to three decimals. And the 1,145 pairs that correlated above +0.3 in year one still averaged +0.446 in year two. Redundant strategies stayed redundant; orthogonal ones stayed orthogonal.

That flips the usual priority. Most people read a backtest for the return figure and glance at the drawdown. On this evidence the return figure is the part that will not carry forward, and the risk figures are the part that will.

The same procedure on a pool that does have an edge

Experiment 4 ran on a universe deliberately built without any screening, and that universe had no edge in either year — 47.1% of it was profitable in year one, 41.3% in year two. Selecting on correlation there controlled risk and could not manufacture a return, exactly as the mechanism predicts. The obvious next question is what the same procedure looks like when the pool does have an edge.

So we ran it on this site's 14 template strategies. Identical rule, and the selection still never sees the second year: compute the correlation matrix on 2024-08-03 – 2025-08-09, take the six templates with the lowest average pairwise correlation (capped at two per market so the result is not just one symbol wearing six hats), freeze them, and score only 2025-08-10 – 2026-08-16.

• Selected at an average pairwise correlation of -0.0394; measured out of sample at -0.0253. The orthogonality carried across, which is the +0.707 result above showing up in a concrete decision.
• Six legs, equal weight: +150.22% with a 7.33% max drawdown, Sharpe 3.82, worst day -3.04%, and 12 of 13 months positive.
• The same six run separately averaged +17.06% at 6.33% drawdown — so the combination multiplied the return roughly ninefold while the drawdown grew by about one percentage point.
• One of the six lost money out of sample (BTCUSD H4, -1.31%). That is the Experiment 2 scenario happening for real, and the portfolio absorbed it.

Turning this back on the first article

We now have the distribution the first article lacked. Drawing 10 screened strategies at random, 2,000 times, gives a median portfolio Sharpe of 2.09, a 90th percentile of 2.49, and a maximum of 3.33. The first article's ten-template portfolio scored 3.12 — above essentially the entire distribution.

That cuts two ways, and both need saying:

• The template screening is stricter than the one Experiment 4 tested — it also requires an out-of-sample profit factor above 1, which is a real hurdle rather than "made money overall". Ten templates beat almost every random draw of ten in-sample winners.
• Precisely because it sits that far out, it is the number most likely to be flattered by hindsight. Experiment 4 is the reason to take that seriously rather than as boilerplate. Treat 3.12 as an upper bound, not an expectation, and note that we have not yet shown any screen surviving a true forward test.

What to actually do with this

1. Do not expect a year of backtest to pick winners. That is the headline finding, and it is the opposite of what most retail workflows assume. 42.1% survival with a confidence interval spanning 50% means the ranking you spent a weekend producing may carry no information at all.
2. Diversification cannot rescue that. Adding legs multiplies the pool's expected edge by roughly √N in whichever direction it points. If you cannot establish that the sign is positive, more legs makes the outcome more certain, not safer — probability of loss went from 58.5% to 80.0% as legs went from 1 to 12 out of sample.
3. So spend your effort on the screen, not on the count. Things that plausibly help and that this test did not include: requiring out-of-sample profit factor above 1 inside the backtest, requiring the same logic to work across several markets, using far more than one year, and forward-testing on demo before funding. None of these are proven here — but "more history, harder hurdle, independent confirmation" is the only direction the evidence points.
4. Four to six legs captures most of the variance benefit. The √N table shows prediction and measurement separating after four, because real strategies retain some correlation. Beyond six you are adding charts, margin and VPS uptime for a shrinking statistical return.
5. Check the correlation before adding, not the backtest. A second strategy on the same market and family correlated +0.91 in this universe. It costs you a chart, margin, and attention, and buys nothing. This is the part of the analysis that held up out of sample: correlation persisted at +0.707 across the two years, performance at -0.034.
6. Judge legs individually, forever. The portfolio hides a dead leg well — that is the point of Experiment 3, and also its danger. A leg that has quietly stopped working will not show up in the portfolio equity curve until several of them have.

Where this test is weak

• Experiment 4 is one split and one out-of-sample year. 2025-08-04 to 2026-08-02 is a single draw from the space of possible futures. A different year could rank these strategies differently, and a 42.1% survival rate measured once has a confidence interval (34.5–50.1%) wide enough to contain both "slightly worse than a coin flip" and "exactly a coin flip". What it does not plausibly contain is "history reliably predicts the future", which is the belief the exercise was testing.
• One year of selection data is a weak screen by construction. A longer in-sample window, or a hurdle stricter than Sharpe > 0, might do better — this test cannot rule that in or out, it can only report that the common workflow failed.
• Two years, one path. 728 calendar days is not enough to rule out a regime that takes down uncorrelated strategies together. The first article found a 19-day stretch where 8 of 10 legs lost simultaneously despite 0.014 average correlation.
• The universe is one engine's opinion. All 322 strategies are single-position, bar-close, fixed ATR stop, RR take-profit. That construction is itself a bias — it disfavours scaling, grids, and intraday timing, and the 59% base failure rate is a property of that construction, not of trading in general.
• Experiment 3 uses the full-sample mean to zero out a leg's edge — it needs to know the future to remove exactly the right amount. It is a stress test, not a forecast. Its value is as a bound: it shows what survives when a leg contributes precisely nothing.
• Real spreads are included; swap and commission are not. Six legs pay six sets of both.
• Equal-risk scaling assumes returns scale linearly with position size. Margin, lot granularity, and compounding all break that at the top end.

Frequently asked questions

Does adding more EAs reduce my risk?

It reduces your volatility per unit of expected return, which is not the same thing. In this test, portfolios built from unscreened strategies had a higher probability of losing money as legs were added (57% at one leg, 80% at twelve). Diversification multiplies the edge you have; it does not supply one.

If backtests do not predict future returns, why backtest at all?

Because return is the one output that did not carry forward — the rest did. Correlation persisted at +0.707, worst single day at +0.610, max drawdown at +0.307, all with p < 0.0001. So a backtest is a risk instrument: it tells you how much this thing can hurt you and how it will move against your other positions, and both of those held up out of sample. It also rules strategies out, which is worth something. What it will not do is rank candidates by how much they are going to make.

I have 5 EAs and one has been losing for months. Should I remove it?

The measurement here says one bad leg out of six is survivable but expensive — it cost more than half the equal-risk return. Judge it on its own record, not on the portfolio's: if it has broken its own backtest expectations, the portfolio is not a reason to keep it. If it is merely flat while uncorrelated, Experiment 3 says it is costing you far less than it feels like.

How many strategies is the right number?

Four to six, on this evidence. The √N benefit is close to theory through four legs and decays after that because of residual correlation. Beyond six you are adding operational risk — charts, margin, VPS uptime — for a shrinking statistical return.

Can a strategy that loses money still help the portfolio?

Only if it is a genuine hedge, meaning reliably negative correlation, and this universe had very few of those. A merely uncorrelated losing strategy subtracts its loss and adds nothing back. Note the worst-day column in Experiment 2: the bad legs did not make the tail worse. They just cost money every day.

How do I know if my strategies are uncorrelated?

Backtest each one, export the equity curves, resample to daily, take the daily returns, and compute the correlation matrix. The rule of thumb that falls out of this universe: same market and same family is near-duplicate (0.91 at the top), different market and different logic is near-zero.