How a smarties analyzed time series data to build a quant algo trading system?
This was a discussion found on Linked In:
There are statistical tests that you can try to determine the relevance of a particular variable before training. There are also several machine intelligence models that adapt the structure and thus remove variables that are not relevant. (IEEE Trans. on Neural Networks, vol.1 no. 1 page 100).
I recommend that you do more reading on neural networks and their quirkiness. Try to define if you have too many input variables (high dimensionality), or irrelevant variables (statistical tests) or low information content (low correlation of input to output) and understand your problem. If you are a programmer I can suggest Tim Master’s books as a good reference in both understanding and in modifying the models.(available on Amazon).
As I had promised before, those who have sent me time series I would do my best to design trading system for them. But most importantly I would share part of my system design approach.
I have load the files for two systems from data sent by Cleber Gomeshttp://br.linkedin.com/in/clebergomes . The directory is called CGomes and it is in my linkedIn page www.linkedin.com/in/drtenorio .
Thanks Cleber for your challenge. Now you get to explain what the data is about.
The results look very promising. I did the first time series with the idea of pattern recognition and multi-agents and the second with double equity trackers. You should be able to gain something from it and I’d appreciate any kind of feedback.
I am know turning my guns to Louis’ series. I am going to do both a market trading system as well as a classifier/predictor for boolean variables. Let’s see how that goes.
Cleber is considering sending me the open equity of a trade system so that I can improve it with post-mortem analysis. If anyone has a trade system with intraday open equity calculations and the corresponding market bars, I would consider looking at it – I need the permission to post the data later on.
I know that a third person requested me to look at his data. I have misplaced your message. Could you please contact me again.
If after looking at the reports you have any suggestions please send them to me or post at the forum.
I got the results of your analysis on the time series I generated. First of all, thank you very much for the reports, they are very rich and I am sure I will be able to gain some insights from your methodology, and I may even be able to integrate them into my own trading systems. Definitely you have proven your point, generating profit from two random time series. Those series were generated by myself based on the well known ‘rand’ pseudo-random number generator in C++, and I can guarantee they are almost random in nature, with approximately as many up as down days. One series has a fixed step between time stamps, while the other is designed to simulate real financial data, with varying increments from day to day. Thanks again and congratulations!
The results for Louis’ data are available in my homepage www.linkedin.com/in/drtenorio .
I believe that you will get something from high dimensionality problems. Using meta approaches such as trackers, equity trading and Game Theory we can do this and more to your existing trading system.
You are correct in raising the possibility. But as I said in the report, this was just the beginning for us to continue to work on smoothing the equity curve.
What you suggest can be done and would further reduce drawdowns. I would also suggest:
– In a portfolio, thing about a stop trading switch (cash) for losing instruments and allow the re-allocation procedure to take place
– I have also shown how to combine in time several equity curves, so upon many consecutive losses you can switch strategies. As a matter of fact, the best strategy is to multi trade an instrument and switch to trade the equity curve generated.
– You can easily design a cash switch by counting how many consecutive losses you have and building a distribution and taking it from there. You can further improve it by making it a moving window distribution and dependent only on a probability therefore the system is parameterless again.