Machine Learning with algoTraderJo

Imagine a bag with 1000 marbles in it, where each marble is a single return. Bootstrap with replacement is taking one marble out, recording its value, and then putting it back in the bag, and repeating this another 999 times. Random permutation is taking all the marbles out, one at time without replacing them, until the bag is empty. Both of the above are single repetitions, so we must repeat the whole procedure again N times. This is much closer to a block bootstrap. Imagine a rod 1000 cms long, which we cut up into equal size lengths X ( e.g X = 5 cm ). These separate pieces are then treated exactly like the marbles above for the block bootstrap and block permutation respectively. The length X is typically chosen automatically, usually based upon the statistical significance of any auto correlation present in the time series.

Each of these procedures will produce different distributions to, obviously, test different null hypotheses.

In spite of being over my head, I can actually say yes I have reason to think so.

Non-linearity can be observed for different reasons, including changes in regimes -- transitions of an underlying hidden state. AFAIU, the work of Engel and Hamilton, Yuan (2013) and others supposes that there are fundamental states which correspond to and drive distinct market behaviors (i.e. distinct probability distributions for some feature set), and the transitions between these hidden states can be modeled from the observable data using a Markov switching process such as an HMM.

Some UW graduate lecture notes claim Markov regime-switching models have become "extremely popular" for modeling economic and financial time series including "business cycles, the term structure of interest rates, volatility in economic and financial variables, foreign exchange rate dynamics, inflation rate dynamics, etc."

There's a professed Ph.D. student in the Dukascopy community who posted an article last year describing an HMM he used to consistently outperform a random walk for a variety of currency pairs --- modeling regime transitions between up-trending, down-trending, and ranging states. And of course there's Kinlay's articles mentioned earlier (the website is slow but it loads).

So, yeah I do. But I have to cite other people's work cause I'm a little over my head. Just a few videos into mathematicalmonk's ML course. But the regime-switching idea seems really intuitive to me. Processes that are non-random have real-world dependencies, often hidden from view, and real-world things tend to exhibit persistence, whether they be interest rate differentials, MM agendas, lack thereof, the weather, or what-have-you. There's a persistence about the real world. My laptop isn't perpetually popping in & out of existence, and Mario Draghi isn't changing his mind back and forth 100x per second. Taking advantage of this the way HMMs do seems to be beneficial.

Thanks babelproofre for the explanation. My question then is, why would we use the bootstrap with replacement? It would not keep the statistical information such as total up and down during the time. Wouldn't it make more sense to use the random permutation in our case to keep the statistical distribution equal?

In my tests the stops did not improve the results significantly but add one more parameter to test for bias. Daniel and the Asirikuy communitiy favor the all in (either long or short) method. Are your results with the stops significantly better, or do you have other reasons for it?

Well, it would depend on what you want to test, but generally a bootstrap with replacement would be used to measure the variability of a statistic of interest, e.g. confidence intervals. See https://en.wikipedia.org/wiki/Bootst...8statistics%29 for a more detailed description.

10 years (1994-2014) simulation with EURUSD with +0.8ATR =1 and -0.8*ATR =0
This with SVC with 104 samples and gamma = 0.006.
I encode my inputs with [-2,-1,0,1,2] with 6 days as inputs.
I use python Pandas with every tick to get this.

:-)

In my opinion it would be interesting to use the sample entropy instead of ATR to set the SL/TP thresholds.

Sample entropy to set the SL/TP ?That's intereting.Sample entropy alternative approach to measure random and deterministic elements in financial time series.Do you use Matlab code or can you give a link for sample entropy?

14 Years tick with SVC and 104 samples EUR_USD
I need a super computer !

If your code is able to split the calculations by a big number of threads, this could be a solution: https://www.parallella.org/

I'm just making the first steps in ML, so I will not bore this thread with beginner questions. By the way, I have a server with 2x xeon quadcore, if you see it useful for your study, let me know.

I get very good result with EUR_USD but nothing with others pairs ?

I dont have any explanation ?

GPU is another solution, these device are build for paralleling calculation.
https://developer.nvidia.com/pycuda

A good read on testing for data mining bias here:-
http://www.adaptrade.com/Newsletter/NL-GoodOrLucky.htm

subscribed!!!

an honest question to machine learning gurus; does anyone trade a machine learning algo with success in real trading and for few months?

I personally tried machine learning a while back but due to the complexity and tiredness that it gets you into; i resorted to traditional method of doing research; and haven't looked back since.

Exactly. With high performance GPU is excepted to have even better performance, but it require an initial investment considerable. With parallela you can start with few cards and increase the cluster as you go. This site, have almost a R framework for ML trading implementations (API for datafeed, order placement...) and also some useful libraries in case of computing with parallela. http://censix.com/downloads.html

For now, i'm running my base ideas in MATLAB, just for training some base concepts of ML, but R timeseries analyes and functions seams to be more adequate. I need to learn R...

Starting from page 33, some more good reading on data mining bias
https://r-forge.r-project.org/scm/vi...f?root=blotter

if you doubt machine learning doesnt work for trading, see these recent headlines about HFs pushing in this area - http://www.bloomberg.com/news/articl...in-public-data

the type of stuff we discussed in this thread is already being done for years.. the new push is using more unstructured data...

Little bit off the topic.I want to maximize the correlation coefficiency of last 2 bars of an indicator to past 500 bars.What would be the best method to achive this?
Neural network(what type would be suitable?)? genetec algo?GA+NN?ARIMA ?Random forest?

also, SVM ,Levenberg-Marquardt_nonlinear_least_squares, perceptrons, distributing kernel regression ... and so on ??

Hello may i ask you something, do your input using amount of candle? if yes how many candle did you use? Thank you

I am using 12 candles in inputs.With encoding like [-3,-2,-1,0,1,2,3] depending of height of each candle