The purpose of this thread is not crystal clear to me but I have some general ideas. I want it to be the following things: A journal to keep details about basic machine learning experiments I run so that I can take a look at them in the future ...

Machine Learning with algoTraderJo

It is very important for the data to be GMT +1 non-DST and GMT +2 DST. It is also important for your data to start Monday at 4 and finish on Friday at 19. The timestamp alignment of the data is very important for any of these methods to give you the ...

Machine Learning with algoTraderJo

It is also interesting to note that the normalization of the data (using returns instead of absolute price differences) has little effect when using this linear regression model although the effect is not zero. Below you can see the results for the ...

Machine Learning with algoTraderJo

Here are two plots showing you the actual returns on the EURUSD and the returns that are predicted by the linear regression model at each point: image image As you can see the predictions made by the linear regression model fall in line with the ...

Machine Learning with algoTraderJo

We're looking at daily data here. If you compare the bar return and range distributions of 1991-2000 daily data with those of 2000-2014 daily data you will find that they don't change that much. Also consider that I am not giving any preference to ...

Machine Learning with algoTraderJo

All simulations contain a spread of 3 pips per trade.

Machine Learning with algoTraderJo

Yes! This is a mechanism to improve results. We will get into that with time.

Machine Learning with algoTraderJo

Yes the 0.5 x ATR20 is part of the strategy parameters (perhaps I am not understanding very well what you mean?). About the ATR, I use it because it is the simplest best proxy I have found for future volatility. The ATR20 has a very good chance of ...

Machine Learning with algoTraderJo

Good questions. Carry out the experiment of running a random system that trades with the same frequency (approximately 3700 trades across 20 years) with the same stoploss and see the probability that you will come up with a better sharpe. Running ...

Machine Learning with algoTraderJo

Till now we have used only classifiers with binary inputs and outputs for the creation of trading systems. Nonetheless we can also use models that instead of predicting a category (in our case we were predicting a binary category) attempt to predict ...

Machine Learning with algoTraderJo

I personally don't like Java very much, in my own trading things generally come down to being able to do a lot of intensive computing in a short span of time and therefore I prefer to use C/C++ for building machine learning models. I have also found ...

Machine Learning with algoTraderJo

Nifty resources! Hope you enjoy the thread!

Machine Learning with algoTraderJo

It's not that straightforward! Results can be very different for two pairs, even if they are somewhat correlated some of the time. Note that the EURUSD and USDCHF are not always so tightly correlated (a 30 day rolling window correlation measure ...

Machine Learning with algoTraderJo

You train your K-NN or SVM using the past N examples, all of which have X bar directions as inputs and the target bar's direction as output. Using just the data you have posted you would build the examples: 0 0 => 0 0 0 => 1 0 1 => 0 0 1 => 1 1 0 => ...

Machine Learning with algoTraderJo

Depends. In my opinion it is first important to determine if the two modes do offer some predictive advantage (do their separate predictions have value or not) then if they both have value it is worth testing whether their intersection performs ...

Machine Learning with algoTraderJo

It is also rather interesting to take the same SVM technique and put it to a test on a symbol different than the EURUSD. For example we can test it on the USDCHF and see what results we get (both results are showed now for comparison). All strategy ...

Machine Learning with algoTraderJo

Whatever coding solution you consider it is extremely important that you consider how you will simulate and live trade from the start. A mistake I made in the very beginning was to only concern myself with the building of models and simulations ...

Machine Learning with algoTraderJo

On our last construction of a strategy we used a portfolio of two models, one trading using a K-NN and the other using an SVM. The combination of both models gave something which in the end had a better performance than both separate strategies. ...

Machine Learning with algoTraderJo

Outside of the trading world there are many. For example think about face and audio recognition. A neural network can find many non-linear relationships that make the prediction of a face or an audio track possible while doing this with transparent ...

Machine Learning with algoTraderJo

Data-mining bias is a concept that applies to all system searches, regardless of how you are searching for strategies (it applies to machine learning, rule based strategies, etc). I remember a discussion between Eric and Daniel on the Asirikuy forum ...

Machine Learning with algoTraderJo