Have you tested whether backtesting on 10-20 years of data is really meaningless? This is an opinion you have. Consider that it might be very wrong. Look at the 1H data distribution of returns from 10 years ago and from this year. Do you see them ...

Machine Learning with algoTraderJo

Talking about this type of system, the exit we decide to use in trades and the ML algo is one of the salient points in the successful implementation of ML strategies of this kind. Since we try to predict each trades' result, the exit will set to ...

Machine Learning with algoTraderJo

For those of you with access to F4 I will email Daniel the functions to create a system as described on this post, he will probably debug and polish them further and should add them to the next F4 update.

Machine Learning with algoTraderJo

It seems that cloud mining for machine learning has started giving some fruit : url . Those of you interested in doing DMB analysis might find that post interesting. Look at the GBP/USD case, tons and tons of bias. If those systems had been traded ...

Machine Learning with algoTraderJo

It is very difficult and it takes a lot of energy and effort to do DMB analysis, then perform demo/live tests. Sorry but if you want information you will need to put this energy in yourself. I won't do the hard work for you. As I said, I'm not ...

Machine Learning with algoTraderJo

Jo is back with some more machine learning fun. I will be posting for a few days before leaving again. I am terribly busy with my own trading and coding, so there is not much time for this FF thread. If you're familiarized with this thread you will ...

Machine Learning with algoTraderJo

The probability to get those results out of random chance is still too large as your testing period is still way too small. Calculate the confidence with which you can say that your method is performing according to your backtesting results Vs. luck ...

Machine Learning with algoTraderJo

You must really be kidding. Think about what you're asking. You're asking me to show you how I risked my own money and achieved a profitable result so that you can make a clearer decision to trade the same family of methods I use. I get absolutely ...

Machine Learning with algoTraderJo

That's what I'm talking about! Those are solid ideas that would allow you to realistically test the mining bias of ensembles. It will still take a lot of time but it is now in the realm of possibility.

Machine Learning with algoTraderJo

As I said, a valid idea but such an idea does not lead to the desired ensemble results as I posted here, due to the bias introduced in the selection of these "top systems". Often the best ensembles come from mixing signals that perform rather badly ...

Machine Learning with algoTraderJo

Sometimes we get true enhancing ensembles with very nice performance: SL=70% TP=40% A=23 B1=300 C1=6 D1=60 B2=125 C2=18 D2=80 image Performance for ensembles can be jaw-dropping! This is the type of performance that we would want if we expect ...

Machine Learning with algoTraderJo

This will not work. It is very important here to understand that ensembles are not the same as portfolios. Having two very good systems does not mean that combining systems one and two in an ensemble is a good idea because an ensemble modifies trade ...

Machine Learning with algoTraderJo

Effectively true. The search space becomes humongous as the number of potential systems increases exponentially with the number of parameters. We went from looking at a generation exercise with a potentially manageable number of systems (10-100K) to ...

Machine Learning with algoTraderJo

Look at this result on the USDJPY: SL=80% TP=340% A=13 B1=75 C1=11 D1=190 B1=50 C1=20 D1=200 image Often ensemble results will appear to be combinations of very similar strategies that create a result that is not much better than the original one. ...

Machine Learning with algoTraderJo

Chapeau! Did you ever try the new raspberry pi 2? They are supposed to be 6x as fast as the original raspberries for the same cost! Not that this would make it cost effective, but at least it would lower the difference.

Machine Learning with algoTraderJo

So ensembles, let's move onto them a bit... Suppose we want to use a system analogous to the one I have exposed so far where multiple positions can be taken at a time with a given SL|TP goal but we want to use a two algo machine learning ensemble to ...

Machine Learning with algoTraderJo

A member has actually modified the framework to compile on ARM (made it run on a Raspberry Pi!). I have never tried this myself but I know people have built F4 on different architectures.

Machine Learning with algoTraderJo

I have a cluster with enhanced PowerEdge R320 rack servers connected by an Infiband network structure. The number of servers I can use changes (as I work with other people who do their own types of experiments) but I generally can access at least ...

Machine Learning with algoTraderJo

The USDJPY can also yield higher linearity systems using this multiple position machine learning. Here is an example: SL=90% TP=390% A=16 B=275 C=19 D=120 image I think we have seen enough examples of how high linearity and good sharpe ratios are ...

Machine Learning with algoTraderJo

Daniel has allowed me to show a graph from a data-mining bias cloud experiment (using USDJPY hourly data from 1986-2015 in this case). In this experiment a machine learning algo is being evaluated to see if the generated systems (systems with R2>0.9 ...

Machine Learning with algoTraderJo