Yes, you're correct, this is "deeper" meaning it is more complex, not in the sense of the field of deep learning (as in the machine learning field). Nonetheless see that I am assigning inputs dynamically on this post, looking at the input structure ...

Machine Learning with algoTraderJo

I was waiting for someone to suggest something along these lines. We haven't used the High/Low values for anything yet. What input/target structures would you suggest to use them? They do provide additional information but is this information ...

Machine Learning with algoTraderJo

It seems that going into the use of a deeper learning procedure was a bit off-putting (no one posted in a week!). So I will instead go into some simpler yet extremely interesting stuff related with the use of neural network models. So far we haven't ...

Machine Learning with algoTraderJo

For those worried about negative balances. Some brokers have a "No Debit Balance Policy". This is the case for FXCM: url . They will not hold any clients liable for negative balances. I was out of all CHF pairs, so I had no problems today. Thought ...

The CHF tsunami - 15 January 2015

These are the results when using a PF filter of 1.2 instead of 1. As it would be expected the trading frequency changes significantly (becomes much lower) which makes compounding less effective and hence total profits lower. The daily Sharpe ratios ...

Machine Learning with algoTraderJo

The challenge is to come up with ways to improve the last results based (solely) on the information provided on the last post (I don't see how that information would lead you to consider that multiple timeframes could improve results). The exercise ...

Machine Learning with algoTraderJo

This is indeed a good point. A PF of 1.2 is the median of the PF distribution. I will upload this when I resume posting on Monday. Have a nice week-end!

Machine Learning with algoTraderJo

Going back to the idea I exposed on my previous post, I have ran a test using the following idea. The SVM always uses the direction of return of the next 5 days as target and the number of examples and number of inputs (directions of previous bars ...

Machine Learning with algoTraderJo

Cool Some differences are probably due to the feed you're using which is the Asirikuy data (I presume), it is difficult to compare when using different time periods as well (difficult graphically I mean). However make sure you use C=10 ...

Machine Learning with algoTraderJo

I'm moving now into some more complex and insightful experiments about machine learning. Picture all the experiments we have done up until now, we have been using a model that retrains every day but we have been choosing the number of examples and ...

Machine Learning with algoTraderJo

For those of you using the Asirikuy framework you can display your graphs in R instead by using this simple loading script (you need to have the PerformanceAnalytics library installed): library(PerformanceAnalytics) temp<- ...

Machine Learning with algoTraderJo

Good inputs are an important part of the problem but not the whole problem. You can see for example that sometimes using more inputs has apparently helped (look at the examples where I used data from other symbols to predict EU values and obtained ...

Machine Learning with algoTraderJo

The reversing on trailing on the same signal and reversion on contrary is implemented automatically on the F4 framework if you use the openOrUpdateLongEasy/openOrUpdateShortEasy functions for entering trades. These functions automatically implement ...

Machine Learning with algoTraderJo

You only include complete examples when building your examples for training. If you are training an algorithm that makes a prediction that needs 4 bars you simply cannot use the last 3 examples for training because, as you say, they are incomplete.

Machine Learning with algoTraderJo

I use at least 4 different data sources. I use 1M data 2000-2014 from Asirikuy, 15M 1986-2014 data from forex-historical, Dukascopy tick data from 2007-2014 and Oanda tick data from 2007-2014. I always confirm that all 4 data sources give similar ...

Machine Learning with algoTraderJo

You might be interested in this free and open source automated system generator.

Robot Lab Project

If you remember last time we looked at target results changes we were using a C=1 and a gamma=1 using 30 examples and 8 bar inputs. It is interesting to revisit the target comparison now using the configuration of the last SVM results. Using 145 ...

Machine Learning with algoTraderJo

I would like to point out that the SVM model using 5 bar targets already has a 22 year annualized sharpe ratio of 1.39 while the sharpe ratio of the SPY is 1.20, this is already a strategy that has better risk adjusted returns vs the broad market.

Machine Learning with algoTraderJo

Drawdowns are caused mainly by changes in information which take time to digest by the models. If the local distribution of returns of the time series starts to change significantly it takes time for the model to be able to predict something in line ...

Machine Learning with algoTraderJo

We can also make classifiers more complex to attempt to filter predictions that are too small to be of practical interest. Suppose we want to enter trades only when the total net movement within the next 5 bars is predicted to be larger than 30% of ...

Machine Learning with algoTraderJo