We can also test adding more inputs to the random forest model to see if we can get even better predictions by looking at more data sources. For example what happens if we attempt to add 8 bar returns for the USDJPY to our previous simulations? This ...

Machine Learning with algoTraderJo

All those are valid suggestions and links, however before we go into exploring better random forests algorithms there are many things you can do to improve the results of the random forest model. When working in ML (even more in FX) I have always ...

Machine Learning with algoTraderJo

Those are actually very good questions. In sample/out of sample testing is useless IMHO because you would simply find some way of weighting the systems that would work on your "out of sample" period with equal effectiveness. There is no real out of ...

Machine Learning with algoTraderJo

I use the shark library for training using linear regression. You can check the library's linear regression page for more information on how the algorithm works. Shark is open source so you can also take a look at the library's source code if you ...

Machine Learning with algoTraderJo

There are many ways in which we can attempt to get better trading results. So far we have discussed portfolios, ensembles, changing predictions targets, changing models and changing our training parameters (how many examples to use for training on ...

Machine Learning with algoTraderJo

Theil-Sen is a more robust linear regression estimator so it obtains better results for some sets of data but I am unsure if it will give you better predictions in the case of this exact linear regression algorithm. I have not used Theil-Sen before, ...

Machine Learning with algoTraderJo

We are only using daily bars at the moment. Lower timeframes have to be treated differently because of daily volatilty cycles caused by market opening|closing times. When we move to lower timeframes on the thread you will see how this is done. ...

Machine Learning with algoTraderJo

We can also try to make predictions using a random forest model instead of a linear regression model. I use a target that attempts to predict the highest and lowest return points within the next 10 bars, using the return of the past 2 bars as input, ...

Machine Learning with algoTraderJo

The programming framework I use allows to develop/test/trade trading strategies on Linux and it is also possible to compile it for use in ARM devices. I have tested doing this on a Raspberry Pi, where it works perfectly well. It also uses a python ...

Python Trader code and skills sharing

We can also play with the prediction frontier to see what horizon prediction brings the most favorable results. Changing the target to predict the lowest and highest returns within a 2, 5, 10 and 20 day period gets us the balance curves shown below. ...

Machine Learning with algoTraderJo

To end the week I want to leave with a new regression based strategy. In analogy to what we did with classifiers we can also attempt to make regression models that try to predict targets that are more strongly related with our ability to profit from ...

Machine Learning with algoTraderJo

I get the same results for 2007-2014 when executing tick-for-tick simulations using Dukascopy tick data and using candles. Since we're on the daily TF and the stops are so wide this does not make a very large difference. The difference we get must ...

Machine Learning with algoTraderJo

Talking about reproducing results, I talked to Daniel at Asirikuy last week and he agreed to include all the machine learning functions I send him on the main Asirikuy programming framework releases. If any of you are Asirikuy members and have any ...

Machine Learning with algoTraderJo

When doing regressions we can also try other models to see if we can improve the results obtained using the linear regression approach. The first thing that comes to mind is to use models that are more complex, that can find relationships that may ...

Machine Learning with algoTraderJo

Your rules and definitions are ok. Note that I do all calculations (including the ATR calculation) on daily data that has been corrected to GMT DST +2, non-DST +1. If you can please post some balance graphs so that I can take a quick look and see ...

Machine Learning with algoTraderJo

I would like to strongly emphasize the importance of the daily rate structure. Daily candles can be constructed in a wide variety of ways (weekly starting and ending times, GMT shifts, DST or no DST, etc). The way in which the candles are structured ...

Machine Learning with algoTraderJo

These suggestions do not work (already tried them extensively) due to these reasons: There is no constant relationship between trade returns. Systems generally have periods of large consecutive profits/loses and then they can have periods where they ...

Machine Learning with algoTraderJo

Yes, the ensemble models are also using a stoploss of 60% of the ATR20. Sure, Neural Network are a topic we will start covering in the future. There are lots of things I want to go through before going into NN models.

Machine Learning with algoTraderJo

Let us make this thread a little bit more interactive Post some ideas you would want me to try to improve the machine learning results I have posted so far (predictions on the daily timeframe on the EURUSD), I will try them out and will post ...

Machine Learning with algoTraderJo

When using ensembles an interesting thing you can do is to take signals that are favored by a given % of your models. Playing with this threshold can help you adjust trading frequency and avoid not trading whenever you have low disagreement (for ...

Machine Learning with algoTraderJo