Mechanical Forex

When building trading systems it soon becomes clear that increasing returns by increasing the statistical edge of trading strategies is a difficult endeavour. After building their first trading system with a statistical edge, many traders battle with endless modifications to make it better;  creating a large amount of degrees of freedom that ends up introducing excessive fitting that destroys the trading strategy by greatly increasing the complexity of its parameter space. It then becomes obvious that it is easier to find a different edge – along other time frames or other instruments – and exploit another elegant market inefficiency rather than risking ruining what is already working properly. However, the combination of historical trading system results and the interpretation of this analysis is not simple, given that there are many things that can introduce problems and make historical correlations inaccurate.

On today’s post I am going to describe the different aspects of trading systems that make any reading of historical correlations tremendously inaccurate, the risks these problems carry and how they might be corrected to obtain much more reliable historical correlations. Please note that I am not saying here that I found a way to obtain historical correlations that will hold with certainty into the future, this is something that cannot be done, simply because the future is unknown. What I am saying is that in order to have a more reliable interpretation of portfolio historical results, it becomes important to have historical correlations that accurately depict how the different system instances were actually correlated in the past. This leads to lower overestimations of the historical – and therefore the expected – profit to drawdown ratio of a trading system combination.

–

–

When putting trading systems together – building a portfolio – the first thing we want to ensure is that the trade sequence we have in the past is the only possible trading sequence that can be achieved by that strategy. If a trading system has the possibility to reach very different trading sequences, then any view of historical correlations with other system is distorted as we can be looking at a correlation that breeds from very specific circumstances in the past. For example, if I have a system that buys on a moving average cross and then exits this trade on a cross of a stochastic oscillator, there is a problem because after the initial moving average cross there can be other crosses before the stochastic oscillator gives the exit signal. This means that the system could have a different result if it is traded from point A, point B or point C (see graph above).

Systems with selective signalling are common and represent a major problem in getting reliable portfolio simulations. The problem highlighted above – commonly referred to as trade-chain dependency – is caused by the lack of acting on all available trading signals. A trading strategy that ignores some signals depending on its positioning will never be able to give accurate historical correlations with other strategies and – even its single-system simulations – will carry significant problems due to the fact that profitability is curve fitted – if optimizations are being done – to match a particular trade chain structure. If starting form point A gives a certain result but starting from point B gives more profitability, then the system – when optimized – will favour certain outcomes based on the sequence of trades and its starting point.

The problems introduced by trade-chain dependency cannot be underestimated as they usually lead to incredible overestimations of profitability and under-estimations of drawdown. This is something we experienced at Asirikuy when building large portfolios where the significant presence of trade chain dependency amongst strategies generated portfolio simulation results that bordered insanity. The optimizations of single systems to match a specific trade chain added to the building of portfolios with hind-sight, lead us to the creation of portfolios that were seemingly “holy grialistic” in nature. For example a portfolio based on 7 different trading strategies generated average annualized profit to draw down ratios that amounted to more than 5, sometimes even more than 10. When the generally accepted “good result” in the trading world – in real trading –  is 1-2 (see my posts on the Barclay index for more on this), these results were simply too good to be true, something that I pointed out clearly at the time (although at this moment we didn’t know why they were not realistic, read we didn’t understand the problem in the simulations).

–

–

After large portfolios started showing bad results and great amounts of seemingly inexplicable broker dependency started to show up, it became clear that something was very wrong with the simulations. These problems – after a lot of analysis – were attributed to the above mentioned problem, trade chain dependent strategies that generated results that lead us to believe that very good historical correlations between draw downs were possible when these correlations were dependent on very precise start-up conditions that were very easily distorted by changes in individual trade outcomes (if results are trade chain dependent any broker dependency issues will cause a loss of synchronization that will completely change correlations).

Moving to the F4 system framework we corrected these problems by introducing system versions without any trade chain dependency issues. This was achieved by making systems act upon every trading signal that was generated. This means that all trading signals cause some sort of action that generates syncing of otherwise independent trade chains. In our case same-direction opening signals generate updating of the SL/TP targets while opposite entry signals generate a stop-and-reverse type of behaviour. These changes have created versions of our trading strategies that lack any form of dependency, something that shows up clearly when building system portfolios. The lack of trade chain dependency causes historical correlations to be much more reliable, generating historical simulation results that are within what would be expected for realistic historical results ( AAR/MDD ratios of 1-2). This gives us more peace of mind and allows us to draw profit and drawdown targets that are more robust (more likely to be in line with future results).

The lessons drawn from this experience with trade chain dependency are clear, reinforcing the way in which Asirikuy works. When you obtain a result that is well above what is generally achieved by the trading community – especially audited professional traders such as those of the Barclay index – pay very close attention to everything you’re doing because you are possible doing something wrong. It is not impossible to go above the performance of the Barclay index but in general this is exceedingly hard and if you have found it easy to achieve then you have possibly overlooked some aspect of your simulations or your development process. Maybe you are curve-fitting, you have trade chain dependency, you are ignoring a vital simulation problem, etc. Take great care and examine every aspect of your strategy. Sometimes we need to go to live trading before we find our mistakes but the important thing is to find them, correct them, learn and move on.

If you would like to learn more about experimenting in system development and how careful system design and evaluation can lead to learning and understanding please consider joining Asirikuy.com, a website filled with educational videos, trading systems, development and a sound, honest and transparent approach towards automated trading in general . I hope you enjoyed this article ! :o)