Combining strategies: overlap, correlation, and diversification

Combining strategies means deciding how they share one portfolio

A multi-strategy portfolio runs more than one return process under a common capital and risk budget. The aim is often diversification: a loss in one strategy may be offset or avoided by another that responds differently.

Different names do not guarantee different risk. A momentum strategy and a breakout strategy can both buy the same futures during the same trend. A VWAP reversion strategy and an opening-range fade can both become short the same instrument during a strong upward repricing. Their logic is different, but the portfolio experiences the combined position and combined loss.

Combining strategies therefore requires more than a correlation matrix. The researcher must inspect positions, instruments, timing, return dependence, drawdown overlap, capacity, and competition for capital.

Diversification has several layers

A low full-sample correlation can coexist with severe tail co-loss. The period when diversification is most needed deserves its own analysis.

Why combining can help

If strategies respond to genuinely different mechanisms and states, their idiosyncratic outcomes may partially offset. A slow trend sleeve, a market-neutral relative-value sleeve, and an intraday strategy can have different clocks, holdings, and failure modes.

Diversification cannot turn a weak or untradeable strategy into a strong one. Adding a negatively correlated stream with negative expected net return may smooth history while destroying capital. Each component should first have a clear role and defensible standalone evidence.

The combination also creates behavior that standalone backtests cannot show: orders compete for cash, one strategy may block another under exposure limits, positions can net, and portfolio halts can change later trades.

Put components on a comparable basis

Before allocating, align:

• date range, timezone, and session treatment;
• return definition and starting capital;
• fees, spread, impact, and financing;
• execution fidelity;
• mark-to-market and missing-data policy;
• risk measure and evaluation frequency; and
• treatment of inactive periods.

A daily-return strategy and an intraday trade list cannot simply be placed in adjacent spreadsheet columns and correlated without resolving their common clock and equity accounting.

An illustrative two-strategy portfolio

Illustrative example. Suppose Strategy A and Strategy B each have a standalone risk target of 10 units. On a given day:

• A requests a +6-unit long position in an index future.
• B requests a −4-unit short position in the same future.
• the portfolio has a maximum gross open risk of 12 units and an instrument limit of 8.

At the account level, the requests can net to +2 units if the execution and policy permit. But gross demand was 10 units, and both strategies still have distinct ownership of their trades and P&L. If B’s short is rejected or delayed, the portfolio is not hedged as the standalone return series implied.

Now assume both strategies request +6. Their standalone backtests each fit a 10-unit budget, but together they breach the instrument limit. Priority, clipping, or rejection changes which return stream actually occurs.

This is why a portfolio simulation must evaluate shared capital and admission, not merely add historical curves.

Designing the combination study

Start with transparent baselines:

1. Each strategy alone under the common economics.
2. Equal capital across strategies.
3. Equal standalone risk under explicit caps.
4. The proposed allocation and admission policy.

Add strategies one at a time. For each addition, measure:

• change in net return and drawdown;
• marginal and total volatility;
• position and instrument overlap;
• ordinary and downside correlation;
• simultaneous drawdown duration;
• strategy contribution and concentration;
• rejected or clipped opportunities;
• netted versus synchronized turnover; and
• combined capacity and impact.

Use rolling and stressed windows descriptively, but do not optimize a new set of weights for every historical window and call the result stable.

Correlation is a summary, not the decision

Correlation measures linear co-movement under a chosen sample. It can be noisy when histories are short, misleading when return distributions are nonlinear, and unstable across states.

Add drawdown-overlap tables, conditional correlations during portfolio losses, and holdings-based diagnostics. Look at the actual dates of the worst combined outcomes. The objective is to understand shared causes, not to invent a new story for each loss.

Important failure modes

• Different backtest clocks or pricing rules make streams incomparable.
• Full-sample correlation hides common tail losses.
• One sleeve dominates portfolio risk despite equal capital.
• Standalone simulations each assume capital that cannot be used twice.
• Netting reduces ordinary turnover but synchronized exits raise impact.
• Optimized weights exploit covariance noise.
• A portfolio halt or risk limit changes the opportunity set in ways a summed curve cannot represent.

Real-world case

Historical case. Published US HML value and Mom momentum research series provide a reproducible illustration of differently behaving factors. For each month from January 2000 through December 2019, an illustrative blend allocates half to the published HML return and half to the published Mom return, then compounds the blended monthly result.

Over that selected interval, Mom compounded to 1.054, HML to 1.561, and the 50/50 blend to 1.478 before any implementation costs. Their maximum drawdowns were 57.85%, 38.98%, and 33.85%, respectively. The calculation uses retained monthly momentum and five-factor archives with data through May 2026. The blend did not produce the highest ending value; it produced a less severe historical drawdown than either component in this particular window.

This is not evidence that a fixed 50/50 factor portfolio is optimal. The inputs are long-short research factors, the blend assumes free monthly rebalancing, and the window is selected. It demonstrates the question a combination study should ask: did the components lose at different times strongly enough to change the portfolio path after realistic ownership and costs?

Try it in Arizmic

Further reading

• Kenneth French Data Library — Defines and hosts the monthly momentum archive and monthly five-factor archive retained for the selected combination chart. They are reproducible long-short research factors, not executable strategy sleeves with shared capital and costs.
• Markowitz, “Portfolio Selection” (1952) — Establishes how covariance enters portfolio risk and why combining imperfectly related return streams can change the whole portfolio’s variance.
• Engle, “Dynamic Conditional Correlation” (2002) — Provides an estimable model for correlation that changes over time, reinforcing that one full-sample matrix is not a permanent portfolio property.
• Longin and Solnik, “Extreme Correlation of International Equity Markets” (2001) — Uses extreme-value methods to show asymmetric dependence in international equities, with correlation rising in bear markets rather than merely whenever volatility is high.
• DeMiguel, Garlappi, and Uppal, “Optimal Versus Naive Diversification” (2009) — Compares 14 allocation models with a simple 1/N benchmark out of sample, demonstrating why estimated optimal weights need a transparent baseline.