Can Currency Correlation Cut Drawdown? Our Surprising Failure!

A beginner-friendly summary of the verification: “Can Currency Correlation Cut Drawdown? Our Surprising Failure!”.

What’s the idea?

We’re always looking for ways to make our algorithmic FX trading systems (EAs) more robust, and a big part of that is managing drawdowns. A drawdown (DD) is simply the peak-to-trough decline in your trading account balance – basically, how much your equity dropped from its highest point before recovering. Lower drawdowns mean a smoother ride and less stress! My hypothesis for this experiment was that a significant portion of our system’s drawdown, especially in our robust5 setup (which focuses on all-yen cross pairs like EURJPY, GBPJPY, etc.), comes from what I call “correlation drawdown.” This happens when multiple currency pairs move in the same unfavorable direction simultaneously. Think of it like putting all your eggs in one basket – if all your trades are highly correlated, a single market event can hit them all at once, leading to concentrated risk and a bigger DD. So, the bright idea was this: what if we could dynamically adjust our leverage based on the average correlation between these currency pairs? If the average correlation was high, we’d reduce our leverage. This would hopefully shrink the concentrated risk and, in turn, reduce our drawdown. Then, when correlations were lower, we’d have room to increase leverage again (since DD is often the main constraint on how much leverage we can use). To test this, I took our core FX system (version 1.4.1, which already incorporates volume-targeting, a stock market filter, and higher timeframe analysis) and added a “correlation overlay.” This overlay calculated an adjustment factor (corr_lev) based on the current pair correlations, clipping it between 0.6 and 1.4 to keep the leverage adjustments within reasonable bounds. Crucially, this was designed to be applied without any “hindsight” – meaning the system wouldn’t know future correlations when making decisions.

How I tested it

First, I needed to understand the typical correlation environment for our robust5 yen cross pairs. I found that the average pair correlation had a median of 0.63, ranging from 0.42 to 0.80. As expected, yen crosses tend to be quite highly correlated, meaning they often move together. One of the most interesting initial findings was about how this new correlation-based leverage adjustment interacted with our existing stock market filter (which uses the US500 index to gauge overall market risk). I analyzed the relationship between the “correlation leverage” and the “stock leverage” and found that their interaction was essentially +0.00. In simpler terms, they operated on completely different axes; they were non-redundant. This was exciting because it meant our new correlation overlay wasn’t just doing the same job as an existing component. For context, a previous study (Research 98, which explored “breadth” filters) did find redundancy with the stock filter, meaning it was essentially duplicating efforts. So, for this experiment, having a truly separate mechanism felt like a promising start!

What happened?

Unfortunately, despite the promising initial diagnosis, the results were not what we hoped for. Across the entire testing period, the system’s drawdown actually worsened, going from -6.1% to -7.4%. This is a significant step in the wrong direction! Consequently, our return/drawdown (r/DD) ratio, which measures how much profit you get per unit of drawdown (higher is better), also declined from 0.29 to 0.24. And to add insult to injury, the monthly profit remained largely unchanged. I also broke down the results into sub-periods. The correlation adjustment only showed a positive impact on drawdown in 2 out of 4 windows. In fact, for the periods 2018-2021 and 2021-2024, the drawdown actually got worse. Even the Profit Factor (PF), which is Gross Profit divided by Gross Loss (and should be >1 for a profitable system), only saw a slight improvement from 3.65 to 3.23 – a minimal change that doesn’t outweigh the increased DD.

What I learned

Here’s the crucial insight into why our clever idea stumbled: Correlation, by itself, doesn’t distinguish direction. Think about it this way: when all the yen cross pairs are rising together, their correlation is high. For a trend-following system like ours, this is often a fantastic opportunity – a strong, synchronized move that we want to ride! However, our “high correlation -> reduce leverage” rule would kick in and cut back our exposure during these favorable, simultaneous upward movements. In other words, reducing leverage based only on the magnitude of correlation ends up penalizing both the “bad” simultaneous downward moves and the “good” simultaneous upward moves. The result? Drawdown doesn’t necessarily decrease, and our overall returns get chipped away. What truly causes drawdowns isn’t just “the magnitude of correlation,” but specifically “simultaneous adverse movements downwards.” And here’s the kicker: this “downward risk-off” scenario is already captured and managed by our existing stock market filter (US500). That filter is designed to identify periods of general market risk aversion, which often correlates with currency pairs moving unfavorably. It understands direction. So, while our correlation magnitude adjustment was indeed operating on a “separate axis” from the stock filter (as we found with the +0.00 interaction), it ultimately proved to be unhelpful. It was a separate axis, but an unproductive one. The big takeaway from this experiment, much like with our previous breadth study (Research 98), is that our existing stock market filter is already doing a superb job because it understands direction. Simply using the magnitude of correlation isn’t effective for reducing drawdown; what’s useful is a direction-aware signal (like identifying “risk-off” periods), and that’s already built into our system. Therefore, no changes will be made to version 1.4.1 of our system based on this research. While the experiment didn’t yield the desired results, it provided valuable confirmation that our current approach to managing risk, especially with the stock market filter, is robust and effective. Every “failed” experiment is just another step in learning and refining our trading strategies!