You're mistaken about what the DMB analysis tries to do. It does not try to evaluate whether your system sees something on the real data that is not present in random data. It seeks to find out whether your process can generate a similarly ...

Machine Learning with algoTraderJo

The train/cv/test process has nothing to do with DMB. If your process is successful it does not mean that an equal process wouldn't be successful on random data. The system creation process tells you nothing about DMB, this is why you need to ...

Machine Learning with algoTraderJo

You cannot perform a DMB analysis "just at the end". The DMB must include the entire creation process, otherwise you could just mine and modify as you please. The idea of the DMB is to show you how probable it is for what you have created to have ...

Machine Learning with algoTraderJo

It's a pity I don't have more time to go into this in detail. Just a few more words. These methods always fail because they involve two things: one is reuse of data and the second is bias in the selection of the train/validate/test sets. I'm not ...

Machine Learning with algoTraderJo

Besides how good or bad the method you use is, the point is also that for any system generation process, no matter the method used, you still need to repeat the method on random data sets and perform the DMB analysis. You simply cannot skip this ...

Machine Learning with algoTraderJo

But would you go live with something with a bad test score and a good validation score? Of course not. You would tweak things until you had both a good validation score and test score. This tweaking increases DMB a gazillion times because of the ...

Machine Learning with algoTraderJo

No, that doesn't work to replace the DMB analysis. That type of test tries to remove another source of bias (not DMB) but curve fitting bias but it does not even work for that purpose. The train/cross-validate/test analysis does not work to do this ...

Machine Learning with algoTraderJo

I want to add that the current "Munay" system in the F4 framework contains some nice trailing stop functions that give you a lot of room to explore. IMHO low DMB systems for the GBP/USD can be found using this system if you use the right set of ...

Machine Learning with algoTraderJo

At Asirikuy they do have GPU system mining. With that software you can do 1 billion of system tests a day... Mind boggling. The software is limited to price action based strategies but still, have never seen that sort of mining power elsewhere. I ...

Machine Learning with algoTraderJo

Definitely, test and test and test!

Machine Learning with algoTraderJo

IMHO retraining very often (on every decision) and testing this for long periods of time is not statistical overkill, it's what's necessary. Keep on testing, you'll start seeing this yourself, especially when you face live trading.

Machine Learning with algoTraderJo

1. Never, I always go live with strategies as soon as I validate them. I have come to believe that strategies have "expiration dates" after you find them, if you take your time to paper trade them you're wasting precious time you could be reaping ...

Machine Learning with algoTraderJo

You're partly wrong. I will explain. Imagine there is a trading instrument that went up from 2005 to 2010 and then down from 2010 to 2015. You create a system for 2 months in 2015 when the instrument was going down. Even if you have 10,000 trades, ...

Machine Learning with algoTraderJo

The GBP instruments are difficult. Even if you find something profitable the DMB is going to be high most of the time. I just have about 5-10 GBP/USD systems because of this. What you say all matches my experience, the TP tends to be bad (results ...

Machine Learning with algoTraderJo

In the systems I have posted retraining is done after every trade. This means that patterns are constantly being redefined in the 25 year backtests. For me it is very different to have a trading system with a 10 year backtest with 1000 trades than a ...

Machine Learning with algoTraderJo

I have never searched for one so I cannot tell you if there is. There might or might not be one, you would have to search for it.

Machine Learning with algoTraderJo

There are other "light" machine learning method besides linear regression... Any ideas?

Machine Learning with algoTraderJo

I advice you read babelproofre's posts on the matter. Read his blog as well. He has already pointed to several sources where you can read more about this type of DMB analysis. I have been using this analysis for the past several years and so have ...

Machine Learning with algoTraderJo

A test on the USD/JPY shows useful potential systems as well. See this one as an example: SL = 90% 1D/ATR20 A= 1 (GMT +1/+2) B= 215 C= 5 D= 100 image I like the lack of additional parameters of this type of trailing. We do not make our DMB search ...

Machine Learning with algoTraderJo

Highly stable, strong compounding results are common with this trailing stop modification. This is another example: SL = 110% 1D/ATR20 A= 6 (GMT +1/+2) B= 205 C= 6 D= 120 image See the SL value of 110%. How does it compare with values for similar ...

Machine Learning with algoTraderJo