Most traders looking for an edge look for a better strategy. The measurable win in this article does not come from a better strategy. It comes from putting ordinary strategies next to each other so that their bad days land on different dates.
One strategy's profit and loss over time, treated as a series in its own right. Two return streams are uncorrelated when knowing what one did on a given day tells you nothing about what the other did.
What is the free lunch in diversification?
Combining return streams that do not move together reduces the portfolio's volatility while its expected return stays at the average of the parts. Nothing is traded away to get that reduction, which is why Harry Markowitz's 1952 result is usually called the only free lunch in finance.
For an equal-weight portfolio of N return streams, each with volatility σ and the same pairwise correlation ρ between them, portfolio volatility is:
σ_portfolio = σ · √( ρ + (1 − ρ) / N )
Portfolio volatility against the number of equal-weight return streams, each with 10% volatility. Computed from the portfolio-variance formula above — not from backtest data.
With σ = 10% per stream, that formula gives:
| Streams | ρ = 0 | ρ = 0.1 | ρ = 0.2 | ρ = 0.4 | ρ = 0.6 |
|---|---|---|---|---|---|
| 1 | 10.0% | 10.0% | 10.0% | 10.0% | 10.0% |
| 2 | 7.1% | 7.4% | 7.7% | 8.4% | 8.9% |
| 4 | 5.0% | 5.7% | 6.3% | 7.2% | 8.1% |
| 15 | 2.6% | 4.0% | 5.1% | 6.6% | 7.9% |
| 20 | 2.2% | 3.8% | 4.9% | 6.6% | 7.9% |
Read the bottom curve. The second stream removes almost 30% of the risk. The fourth has cut it in half. By fifteen it is down about 75%. Ray Dalio built Bridgewater on this curve and called it the Holy Grail of Investing: fifteen to twenty good, uncorrelated return streams cut risk by roughly 80% and raise return-to-risk about fivefold.
Why is correlation the price of the free lunch?
The benefit collapses as correlation rises. At ρ = 0.6 the curve falls from 10% to about 8% and then flattens — the twentieth stream does nothing the third did not already do. Diversification is not "own more strategies", it is "own strategies that disagree".
This is the part that gets skipped. A book of six iron condors with different entry times is not a diversified book — it is one trade in six instalments, and it will take one bad afternoon in six pieces.
So the practical question is not how many strategies do I run. It is what is the actual correlation between them, measured on daily profit and loss, rather than assumed from the fact that they carry different names.
What does a genuinely uncorrelated 0DTE book look like?
It looks like thirteen unremarkable strategies whose combined drawdown is smaller than almost any of them alone. The book below is one saved $100,000 portfolio, one contract per strategy, measured over 250 trading days from 15 September 2025 to 11 September 2026.
| Measure | Value |
|---|---|
| Total P&L | $137,378 |
| CAGR | 139.7% |
| Max drawdown | −1.68% (−$3,684, on 31 July 2026) |
| MAR (CAGR ÷ max drawdown) | 83.13 |
| Sharpe / Sortino | 8.34 / 15.16 |
| Peak margin | $59,342 |
| Win rate | 63.6% over 2,609 trades |
| Winning / losing days | 174 / 76 |
| Best / worst day | +$2,911 / −$3,193 |
The thirteen streams come from three families: multiple-entry iron condors that sell premium (MEIC), trend-following credit spreads gated on EMA crossovers (METF), and reverse iron condors that are long gamma (RIC).
| # | Family | Entry | Detail | CAGR | Max DD | Win rate | MAR | Risk share |
|---|---|---|---|---|---|---|---|---|
| 1 | MEIC | 11:43 | both sides | 16.1% | −2.92% | 55.4% | 5.5 | 23.3% |
| 2 | MEIC | 14:20 | calls only | 22.1% | −1.71% | 63.3% | 13.0 | 16.3% |
| 3 | MEIC | 15:20 | both sides | 10.0% | −1.03% | 59.1% | 9.7 | 5.9% |
| 4 | MEIC | 14:34 | no stop | 5.4% | −1.27% | 88.6% | 4.3 | 3.6% |
| 5 | METF | 11:15 | calls, EMA 5/40 | 19.4% | −2.94% | 73.1% | 6.6 | 16.8% |
| 6 | METF | 15:34 | both, EMA 5/20 | 19.0% | −2.01% | 60.9% | 9.4 | 15.3% |
| 7 | METF | 10:54 | puts, EMA 20/40 | 11.3% | −2.36% | 64.2% | 4.8 | 5.9% |
| 8 | METF | 15:20 | calls, EMA 5/20 | 10.4% | −0.62% | 74.2% | 16.9 | 3.4% |
| 9 | METF | 10:50 | puts, EMA 20/40, no stop | 3.7% | −3.16% | 97.8% | 1.2 | 2.1% |
| 10 | METF | 13:03 | puts, EMA 5/20, no stop | 6.3% | −0.04% | 99.3% | 63.3 | −0.2% |
| 11 | METF | 14:10 | puts, EMA 5/20 | 2.5% | −0.45% | 66.2% | 5.5 | −0.9% |
| 12 | RIC | 09:33 | long gamma, VIX gate | 5.9% | −1.95% | 47.3% | 3.0 | 5.6% |
| 13 | RIC | 11:00 | long gamma, 60-delta legs | 6.1% | −1.56% | 24.8% | 3.9 | 2.9% |
Two things stand out before any correlation is calculated.
There is no best strategy in the list. The one with the best MAR (63.3) earns 6.3% a year. The one that earns the most (22.1%) has a bigger drawdown than nine of the others. Win rates run from 24.8% to 99.3%. Ranked on any single column, this list would be pruned down to something far worse.
Two streams carry negative risk share. Risk share is a stream's contribution to total portfolio variance. A negative number means that stream's daily P&L runs against the rest of the book often enough that adding it reduces the total. Those two are not the best performers. They are the ones that disagree.
How do you read a correlation matrix of 78 pairs?
Look at the distribution, not just the average. Thirteen strategies produce 78 unique pairs; this book's pairs average 0.009, but the useful fact is that 55 of them sit inside ±0.1 and only two exceed +0.2. An average near zero built from large positive and large negative pairs would be a very different portfolio.
Correlation here is always Pearson correlation of daily P&L, computed over the days on which both streams actually traded. Rows and columns below are numbered as in the table above.
Pairwise correlation of daily P&L between all thirteen strategies. CashFlow Engine backtest data, 15 Sep 2025 – 11 Sep 2026.
| Measure | Value |
|---|---|
| Pairs | 78 |
| Average correlation | 0.009 |
| Pairs below zero | 38 |
| Pairs within ±0.1 | 55 |
| Pairs above +0.2 | 2 |
| Range | −0.298 to +0.343 |
Almost half of the pairs are negative, which means half the book is, on any given day, partially hedging the other half.
The one concentration worth naming is the highest pair, +0.343: an iron condor entered at 15:20 against a trend spread entered at 15:20. Same hour, same direction of risk. Nobody planned it — it is what happens when two selections are made independently on the same market. It stays in the book, and it is the first thing to look at if the book ever needs trimming.
Where do the hedges in a 0DTE portfolio actually sit?
In pairs nobody designed. The strongest negative pair in this book is −0.298: an iron condor selling calls at 14:20 against a trend spread selling puts at 14:10 — opposite sides of the market, ten minutes apart. Both strategies were selected on their own merits; the opposition between them was discovered by measurement afterwards.
The next strongest negatives follow the same logic: a put-side trend spread against a long gamma position at −0.267, a morning trend spread against that same long gamma position at −0.189. Premium sellers and gamma buyers make money on opposite kinds of days.
This is the practical shape of portfolio construction. You do not engineer the hedges. You measure which combinations happen to hedge, and then decide to keep them.
Get the weekly 0DTE research email
Research notes and product updates, straight to your inbox. Unsubscribe anytime.
Is risk contribution the same as drawdown contribution?
No, and conflating them is expensive. Of the two long-gamma streams in this book, neither has a negative risk share — they contribute 5.6% and 2.9% of portfolio variance, so by that measure they add risk. Adding them still cut the worst drawdown of the year almost in half.
Contribution to total portfolio variance per strategy. CashFlow Engine backtest data, 15 Sep 2025 – 11 Sep 2026.
Variance and drawdown are different questions. Variance asks how much the daily P&L moves around on an average day. Drawdown asks what happens over the worst consecutive stretch. A stream that loses a little on most days and wins a lot on exactly the days the rest of the book is bleeding will raise variance and lower drawdown at the same time.
If you optimise a book on variance alone, you will throw those streams out. Measure both.
What happens at each step of building the portfolio?
Each step below is the same measurement re-run on a nested subset of the same thirteen strategies over the same 250 trading days, so every line is measured rather than argued. Return rises at every step; drawdown does not.
CAGR above the line, maximum drawdown below it, at each nested build step. CashFlow Engine backtest data, 15 Sep 2025 – 11 Sep 2026.
| Step | Streams | CAGR | Max DD | MAR | Avg ρ | Peak margin |
|---|---|---|---|---|---|---|
| 1. One iron condor alone | 1 | 16.1% | −2.92% | 5.5 | — | — |
| 2. The four iron condors | 4 | 54.2% | −2.26% | 23.9 | 0.119 | $19,600 |
| 3. Plus the seven trend spreads | 11 | 127.6% | −3.01% | 42.5 | 0.026 | $58,350 |
| 4. Plus the two long-gamma streams | 13 | 139.7% | −1.68% | 83.1 | 0.009 | $59,342 |
Step 2 is the ordinary version of diversification. Four premium sellers instead of one: more than three times the return, and the drawdown goes down, from 2.92% to 2.26%, because their average correlation is 0.119 rather than 1.0. They are related, but they are not the same trade.
Step 3 is where most books stop. Adding a different family — trend-following instead of premium selling — pushes the average correlation down to 0.026 and the return to 127.6%. But the drawdown goes up, to 3.01%. More streams did not mean less pain. The ratio improved; the worst stretch did not.
Step 4 is the one that is hard to predict from theory.
Can two losing strategies reduce a portfolio's drawdown?
Yes. The last two streams added to this book are reverse iron condors with win rates of 47.3% and 24.8% — one loses three trades out of four. Adding them took the maximum drawdown from 3.01% to 1.68% and raised the return from 127.6% to 139.7%, roughly doubling MAR.
Cumulative P&L of the whole book against two of its families. The long-gamma line is almost flat — its value is when it rises, not how far. CashFlow Engine backtest data, 15 Sep 2025 – 11 Sep 2026.
Both are long gamma: they pay when the market moves and bleed when it does not. Both are profitable over the year, which is why the return rose. But the reason the drawdown fell is that they make their money on completely different days — the sharp, directional sessions that hurt a book full of premium sellers. They are an airbag. Most of the time you carry it for nothing.
What it cost: peak margin went from $58,350 to $59,342 — about $1,000 more buying power on a $100,000 account, roughly 1%. That is the whole price of cutting the worst stretch of the year in half.
The general lesson is uncomfortable and worth stating plainly: the strategies that improve a portfolio most are often the ones that look worst on their own. Any selection process that ranks candidates individually and keeps the top of the list will systematically discard them.
What does diversification not remove?
The tail. A 1.68% maximum drawdown over 250 trading days is a pleasant number and an incomplete picture: this book still loses on 76 of 250 days, needs about $59,000 of peak margin, and has an arithmetic worst case near $11,902 — around 12% of the account — if every defined-risk spread reached maximum loss in the same session.
- The book still loses often. 76 losing days against 174 winning days. Roughly one session in three is red.
- The tail is not diversified away. Correlation between strategies has nothing to say about that $11,902 figure. It is set by position sizing and by how many spreads are open at once.
- Correlation is not a constant. Every figure here is an average over one year. In a genuine liquidity event, correlations between risk-taking strategies tend toward one exactly when you need them not to. A measured 0.009 describes the past window; it is not a promise about the next crisis.
- Margin is real. Thirteen streams need about $59,000 of peak buying power on a $100,000 account. Diversification is free in expected return, not in capital.
Diversification reshapes the ordinary bad week. It does not remove the extraordinary bad day.
Mandatory pit stop: Options trading involves significant risks and is not suitable for every investor. Past results are no guarantee of future performance.
Where is the selection bias in these numbers?
In the selection itself. The thirteen strategies were chosen by looking at the same 250 trading days that are then reported as the result. That is selecting on the data you report, and every number in this article is inflated by an unknown amount because of it.
A longer window does not fix this. Running the same book over 24 months would still include the window the selection was made on, so the measurement would be less out-of-sample than it looks. Genuine out-of-sample means either the period before the selection window, measured on its own, or forward data the selection could not have seen.
Ten of the twenty-three candidates were measured and left switched off: three ratio spreads (one call against two puts), three further iron condors, one trend spread, two long-gamma positions and one long strangle. The ratio spreads were excluded on structure rather than on results — undefined risk on one side — so that the traded book is defined risk all the way through. The other seven were measured and did not earn a slot. They remain in the file and can be re-measured at any time.
Treat what you have read as a research book, not a track record. The only way to settle it is forward, which is why this exact selection is being traded live and reported against this baseline — including in the weeks when the backtest looked better than reality.
How can I measure this on my own book?
Four steps, none of which requires taking anything above on faith: backtest every candidate individually on one shared window, measure the pairwise correlation matrix on daily P&L, build in nested steps and re-measure each one, then test removals as well as additions.
- Step 1
Backtest candidates individually
Run every strategy on its own over the same window, with the same contract count and the same fill assumptions. Different windows make the correlation matrix meaningless.
- Step 2
Measure the matrix, do not assume it
Compute pairwise correlation on daily P&L over the days both streams traded. Two strategies can share a family, an entry time and a ticker and still be weakly correlated — and two that look unrelated can turn out to be the same bet.
- Step 3
Build in nested steps
One stream, then the family, then a second family, then the hedges — measuring CAGR, maximum drawdown, MAR, average correlation and peak margin at every step. A step that improves the ratio while raising the drawdown is telling you something.
- Step 4
Test the removals
Switch the least flattering streams off and look at what happens to the drawdown. That is how the airbag effect in this book was found.
Our own figures come from second-level backtest data, one contract per strategy, with no market-event filters applied. The same four steps work with any backtester that exports a daily P&L series.
Frequently Asked Questions
- How many uncorrelated strategies do I need?
- Most of the benefit arrives early: the second stream removes almost 30% of the volatility, the fourth cuts it in half, and the curve is flattening well before twenty. Dalio's framing of fifteen to twenty assumes genuinely uncorrelated streams. Four that disagree beat twelve that agree.
- Is low correlation the same as low risk?
- No. Correlation describes how streams move relative to each other. It says nothing about how large any one of them can lose. A book of thirteen weakly correlated strategies can still have a maximum potential loss of 12% of the account in a single session, as this one does.
- Why did the drawdown go up when strategies were added?
- Adding a stream adds its own bad days. If those land near the existing book's bad days, the deepest stretch can get deeper even while the average correlation falls. That is exactly what step 3 of the build shows.
- Does a negative correlation mean one strategy hedges the other?
- Partially, and only on average over the measured window. A −0.3 correlation means the two streams have often moved in opposite directions, not that one will cover the other's loss on any particular day.
- Can I just pick the strategies with the best individual numbers?
- That is the failure mode this article is about. Ranking candidates individually and keeping the top of the list discards precisely the low-win-rate, low-return streams that improve the portfolio. Select on what a candidate does to the book, not on what it does alone.
- Are these results live?
- No. Every figure is hypothetical backtested performance over a single 250-day window, and the strategies were selected on that same window. The live version of this book is reported separately.
Terms & Definitions
- Return stream
- One strategy's profit and loss treated as a time series in its own right.
- Correlation
- A number from −1 to +1 describing how two series move together; here always Pearson correlation of daily P&L.
- Maximum drawdown
- The deepest peak-to-trough decline in account equity over the measured window.
- MAR
- Annualised return divided by maximum drawdown — a return-per-unit-of-pain ratio.
- Risk contribution
- A single strategy's share of total portfolio variance; it can be negative.
- Long gamma
- A position that gains as the underlying moves and decays while it stays still.
Related reading
Educational disclosure: This article describes portfolio mathematics and software analytics. It is not financial advice, not a recommendation, and not a trading signal. All performance figures are hypothetical backtested results over a single historical window, and the strategies shown were selected on the same window that is reported. Backtested results do not reflect actual trading and may not be repeatable. Options trading involves substantial risk, including loss of the entire investment.