One of the things you will keep hearing when you start getting into algorithmic trading is how the best strategies are supposed to work on all symbols, all timeframes and even possibly all markets. People will tell you that a strategy developed on a single instrument will have a higher probability to fail because a system that works on two or more instruments will inevitably be fit to tackle more market conditions. On today’s article I want to challenge this idea, showing you clear evidence that the above is not true. Through this post you will read why the above has simply not been true historically and I will also discuss why this is most likely the case. Is there any advantage to using strategies that work on several symbols? Keep reading to find out!

–

–

The idea to have strategies that work on several symbols has been alive practically since the start of systematic trading. The idea is that if you have a strategy that works on two different instruments, for example corn and oil, you will have a strategy that is more robust since it can survive to market conditions on both markets. If the corn market changed to resemble the oil market your strategy would survive, this means that chances of success should indeed be higher. In Forex this usually means to have strategies that will work across several different currency pairs. In line with the previous example this should add robustness to a trading strategy since it can simply tackle many more market conditions.

To test whether the above is true I performed a simple experiment. I used a price action based mining to generate strategies (600 in each case) that worked either on the EUR/USD alone, on the GBP/USD alone or on both the EUR/USD and the GBP/USD. Only strategies with an R² > 0.9 and a trading frequency above 10 trades/year were selected in all cases. Strategies were created using an in-sample period from 1987 to 2001 and a pseudo out of sample period from 2001 to 2016. The pseudo out of sample was used to test how many strategies behaved profitably outside of mining conditions. If the assumption that being profitable on more symbols implies more robustness is true then we should see an unequivocally higher number of profitable systems in the pseudo out of sample in the EUR/USD+GBP/USD case when compared with either the EUR/USD or the GBP/USD single cases.

–

–

The results – summarized above – show clearly how this simply isn’t the case. The systems generated for the EUR/USD+GBP/USD are more likely to fail than both the EUR/USD and the GBP/USD generated systems. The difference is also not small with an almost 20% difference with the EUR/USD systems and a more than 10% difference with the GBP/USD systems. This means that systems generated on the combination of the two pairs are far less robust than the systems generated to trade on only one of the two pairs. The intersection of strategies that work on both the EUR/USD and the GBP/USD in the in-sample period is indeed less likely to succeed than the separate groups of systems that work on the single symbols as a whole.

There is also a substantial difference when it comes to the average profitability of the generated strategies. In the case of the EUR/USD systems the profitability is almost 400% that of the EUR/USD+GBP/USD systems while in the case of the GBP/USD systems the profitability is almost 50% higher. Not only does generating systems for the two symbols imply a reduction in the probability of pseudo out of sample success but it also implies a reduction in the average profitability for the systems. **This means that it is not better to generate strategies that work on multiple symbols, it is in fact heavily counter-productive**.

–

–

Of course the above results could be a rare exception for the EUR/USD+GBP/USD reason why I rerun the above experiment using 6 different pair combinations. In all cases the results I obtained were exactly the same, a reduction in the probability of success and the average profitability of the systems during the pseudo out-of-sample period compared with the results ran for single symbols. This means that the above are not a rare example of what happens but actually a good indication of the average case we might expect. Repeating the experiment with more systems instead of 600 reveals similar results in terms of the differences between the single and multiple symbol results.

But why is this the case? Well, multiple symbol systems would only be expected to work better if there was indeed some sort of cross-exchange between market conditions between symbols (if the GU at some point behaved like the EU and vice versa). However it seems that this is simply not the case, at least when looking at FX currency pairs. In this case there is a penalty when we look at conditions that exist on both the EU and the GU because we are in effect under-fitting the strategies. We are not allowing them to adjust to the typical behavior of each pair because we only want behavior that is common to the two pairs. Since the pairs might have noise in common – much more than any other type of behavior – then we might fit less to some underlying important characteristics of the pairs and more to irrelevant data. Since we’re not picking up the “essence” of the pairs when doing multi symbol trading, the probability of failure increases.

In currency trading each pair is generated by the commercial transactions that are unique to each one of them. Since countries and the companies and people that work within them are so unique, it is not surprising to see that market behavior does not “migrate” across symbols. The special characteristics of the EUR/USD do not become those of the USD/JPY or the GBP/USD just in the same way as Japan does not suddenly turn into Britain and the United States does not suddenly turn into Australia. *With the above evidence it does seem clear to me that single symbol system development in the FX market is definitely the way to go*. If you would like to learn more about trading system creation and how you too can generate strategies and portfolios to trade the currency markets please consider joining Asirikuy.com, a website filled with educational videos, trading systems, development and a sound, honest and transparent approach towards automated trading.

Hi Daniel,

Interesting results. Wonder if you have also tested with other R-squared values (both higher and lower)? I do recall some authors pointing out that the same systems only need to show similar profitability on different symbols, though similarity had not been quantified.

Best Regards,

Ed

Hi Ed,

Thanks for writing. I have indeed tried lower and higher R² values and for the 2 pair cases the results are the same. For the 6 different pairs I have tried the results of the multiple pair systems are always worse than for the single pair systems, regardless of a higher or lower R². However once the R² drops substantially (to around 0.6) both single system and multiple pair systems show similar results, as results become basically random (% of profitable systems in the p-OS drops below 50% and mean profitability becomes negative). Of course do let me know if you have any other questions or comments,

Best Regards,

Daniel