This is the simulation for the portfolio using the 4 systems I described before for the USDJPY (0.25% risk per trade): image The CAGR is 16.8% and the maximum drawdown is 19.7%. The annualized Sharpe ratio is 1.18. The results are inferior to ...

Machine Learning with algoTraderJo

The hour does not matter here because I am not combining them in an ensemble (making decisions based on all system inputs), I am simply building a portfolio (trading them together as independent systems). You cannot test the parameters found on ...

Machine Learning with algoTraderJo

You can also build a portfolio using systems for only one symbol since the mining space contains several candidates that can be combined successfully. Here are 4 USDJPY candidates using linear regression that give decent results (there are many ...

Machine Learning with algoTraderJo

In case anyone wants to perform the data-mining assessment here is the information about the mining process: image The genetic optimization settings for the ast.config (note that F4 uses gaul (open source genetic library) for optimizations): ...

Machine Learning with algoTraderJo

Yes indeed! Linear regression already shows what you can do with ML using simpler models. More is not always better. This begs the question of when we need higher complexity and when we don't. When do we gain by introducing complexity and when is it ...

Machine Learning with algoTraderJo

But there is still sooooo much to explore... What if we used NN? What if we used SVM? What if we used KNN? What if we used random forests ? What if we created a large portfolio using all of these ???

Machine Learning with algoTraderJo

With the linear regression systems I have posted for USDJPY, GBPUSD and EURUSD we can already create a very decent machine learning portfolio. The portfolio below uses a 0.4% risk per trade. This trades 3 major pairs for about 25 years using linear ...

Machine Learning with algoTraderJo

Awesome KaBo . Having others reproduce is a very important part of any research process.

Machine Learning with algoTraderJo

The GBP/USD is complicated. Finding a profitable ML algorithm for this pair is always a challenge. Using linear regression we can find something "decent" with A=12 GMT+1/+2, B=120, C=9, D=10 using a stop loss of 50% of the ATR(20) using our classic ...

Machine Learning with algoTraderJo

You are spot on. We are evaluating for data-mining bias which has nothing to do with curve fitting bias (what you are talking about). What we see with mining bias is whether the system is the product of real causal relationships within the ...

Machine Learning with algoTraderJo

This is a classic problem with testing, you simply cannot circumvent the mining bias issue using historical data. There is no such thing as using a historical "testing set" only once. What happens if you develop something and then it bombs out on ...

Machine Learning with algoTraderJo

To start the week with something interesting I want to show you that the machine learning approach using MLE/MSE values can also be applied successfully with Neural Networks. The backtest I am posting here shows you the results of using a Neural ...

Machine Learning with algoTraderJo

Awesome post

Machine Learning with algoTraderJo

argmax k { Sharpe(algo(A_k, B_k, C_k, D_k)) } when doing mining on the real data is not necessarily greater than argmax k { Sharpe(algo(A_k, B_k, C_k, D_k)) } on the average random data set. When mining bias is significant you will find that the ...

Machine Learning with algoTraderJo

It can be done in many ways. A simple way is to choose a statistic to track, like the Sharpe ratio. In this example you build a distribution of systems according to their Sharpe ratio as expected from random data sets and you compare with the ...

Machine Learning with algoTraderJo

You don't use a fixed system, you apply the same mining process to each random data set. If you performed a genetic optimization to find A,B,C,D then you repeat the exact same process on random data N times, to build a distribution of expected ...

Machine Learning with algoTraderJo

Nice contribution

Machine Learning with algoTraderJo

The mining bias problem is not difficult to solve. We have a strategy X that is the result of performing a mining process A on a dataset B. Our null hypothesis is that the relationships found by the system are spurious, meaning that the results of ...

Machine Learning with algoTraderJo

The GMT offsets and the DST are extremely important because market opening times and daily volatility cycles also change with DST. You need to correct your data to the proper timezone. Also use a higher spread as pipmeup said.

Machine Learning with algoTraderJo

I always trade using GMT +1 non-DST/+2 DST timestamps so the data posted by Pipmeup was automatically corrected within the framework I use to match this setup (GMT +1/+2).

Machine Learning with algoTraderJo