Diversification usually means owning different assets. In systematic trading there is a more powerful version: running different strategy types simultaneously. Our backtest data contains a demonstration of why, including one result that surprises almost everyone who sees it.
Two systems, opposite personalities
The Breakout System buys momentum on JPY pairs: 57% win rate, larger winners, thrives in trending markets, suffers in chop. The Mean Reversion System fades extremes on range-bound pairs: 72% win rate, smaller steady gains, thrives in quiet markets, suffers in runaway trends. Different logic, different pairs, different best and worst environments.
What happens when you combine them
Markets alternate between trending and ranging regimes. Whichever regime is active, one of the two systems is in its element. The combined equity curve is smoother than either component: the combined backtest shows zero losing years across 2015 to 2024, with a max drawdown of -13.6% at 1% risk, and 73% of all months positive.
The counterintuitive result
Here is the finding that surprises people. In our backtest, take the standard 1% risk on both systems, then lower only the mean reversion risk to 0.5% while keeping breakout at 1%. Intuition says less risk, less drawdown. The data says the opposite: max drawdown gets worse, moving from -13.6% to over -15%.
Why? The worst breakout drawdown period (around December 2017) coincided with positive mean reversion months. The mean reversion profits were absorbing part of the breakout losses in real time. Cut the mean reversion allocation in half and you cut the shock absorber in half, exposing more of the raw breakout drawdown. The two systems are not just uncorrelated. In the moments that matter most, they are negatively correlated.
The practical takeaway
- A second strategy is not an optional extra. In the worst periods it is the risk management.
- Balanced risk between negatively correlated strategies often beats reducing one side.
- Evaluate any portfolio by its combined worst months, not each system's average months.
You can test every risk combination yourself in the interactive risk calculator, which runs on the full 6,000-trade dataset.
Correlation is a property of periods, not a constant
The headline correlation between our two systems across the full backtest is low, in the 0.06 to 0.18 range depending on the measurement window. But the more useful truth is that correlation is regime-dependent, and what matters is where the extremes land. Plot the breakout system's ten worst months and check what mean reversion did in each: positive or flat in the large majority of them. Then flip it: during mean reversion's worst stretches (sharp trend eruptions in its cross pairs), the breakout system was usually mid-run in its own trends. The systems do not merely fail to correlate. Their failure conditions are close to mutually exclusive by construction, because one is structurally long volatility expansion and the other is structurally short it. That is a much stronger property than a low average correlation, and it is the property that produces the drawdown result this article opened with.
Why not add a third system, or a tenth?
If two uncorrelated systems beat one, the greedy conclusion is that ten would be even better. In principle yes; in practice each addition must clear a rising bar. Every new system needs its own genuine edge (diversifying into a mediocre strategy dilutes returns without buying much smoothing), needs low correlation to the existing set specifically during their worst periods, and adds operational surface: more charts, more parameters, more things to monitor and re-verify. There is also a subtle statistical trap: the more candidate systems you evaluate for inclusion, the more likely one looks good by chance, which is the portfolio version of curve fitting. Our bar for a third engine is that it must survive the same multi-year development funnel as the first two, show negative correlation to the pair's combined worst months, and justify its complexity with a visible improvement to the portfolio's return-per-drawdown ratio. Candidates exist in the research pipeline. None has cleared the bar yet, and the bar staying high is a feature of the product, not a limitation.