Let me continue to share with you the typically wrong thinking styles of most quant traders.(3)

Let me continue to share with you the typically wrong thinking styles of most quant traders.

  1. Optimizing parameters is necessary.

Are you still optimizing the parameters of your strategies?
After you finish reading this post, you may change your mind and stop doing so.

From our experience, optimizing parameters is no more than wasting CPU time.
Please remember these words:
If there were parameters appropriate to the current market, there would have been parameters appropriate to all the market movements.
Yes, all work or none work.

Why? I can’t prove it by using formulas.
Because this rule is completely from our practical experience. We have tested a lot of times and concluded this rule.
And this rule is really marvelous. It’s subject to the restriction of generalization.

You know, if we optimize a lot of parameters, the trained models must have over-fitting problems.
So, if we want to make a model subject to the restriction of generalization, we only can use up to 5 parameters.
Too few parameters make the profit factor very low. Most of them will not be greater than 1.

You may ask, how about the probability to win a trade if the model is trained by optimizing up to 5 parameters? The probability is quite low, only 53% - 55%. But it’s still a good model. Because the generalization is quite good, no over-fitting problem.

Then, now that the probability is so low, why do we need to optimize the parameters?
That’s why I said optimizing parameters is no more than wasting CPU time.

You may ask that if we don’t need to optimize parameters, how to get a model with a combination of parameters to fit all the market movements?
This article is focusing on the experience that optimizing parameters is not necessary, how to get a model without the requirements to optimize parameters is another topic, and actually, it is a really difficult goal to achieve. Let’s discuss it in the future.

Next time, let’s discuss another wrong thinking style: B-book brokers are not good.


Yes, I think so too.
Because the parameters optimized only fit the historical data.
Regarding whether the parameters can fit the data in the future, who knows?
No guarantee.

Yeah, that’s why the improved_martingale goes quite well even if the signals are randomly generated.
The most important thing is the parameters are just discrete points in an infinite range.
Having good backtesting results is no more than winning the lottery.
It may happen in the past, that doesn’t mean it will happen twice in the future.

1 Like