Why the winning rate of trading systems is not relevant

Usually when you read some articles about some students, banks or quants researching or operating a new trading system you will find an explanation that their system works because they predict more than 50% of all cases correctly. Some authors even write “more than 51%” to underline that it is really, but really more than in half of all cases.

In the first moment you think “Yeah, that’s logical, when they predict the direction of the price development in 51% of all cases correctly and fail in 49%: they make money”.

I know that most of you see the mistake of this thought. For the others a short explanation: Even if you fail in 99% of the cases you can make money if you lose 99 out of 100 times one $/€ and the 1 out of 100 times you win 100 $/€.

What counts in the end is the gain, not the winning rate. No surprise. I know.

But what does that mean for machine learning algorithms, or more specific: what does that mean for neural networks. Does it in front of this background make sense to use a classification network to predict the direction of a stock or a currency pair. Yes and no.

Why ‘no’? That’s what I described above.

So, why ‘yes’? What counts is the relation between the winning rate, the average loss and the average gain. The average loss is again limited by the stop loss, the average gain by the take profit. And for a given stop loss and a given take profit the winning rate might in fact be a benchmark.

And how to find the best relation? I recommend the good old iterations through ranges. No back propagation, no mind blowing technology, simple number crunching.

If you think that an evolutionary network might an option or if you just want to contact me: do not hesitate to write an email or leave a comment.

Kind regards