Disliked{quote} How important is the choice of the distance (metric) is nearest neighbor? L1, L2, L∞, Mahalanobis.... Do you have a preference?Ignored
- #63
- Dec 9, 2014 1:48pm Dec 9, 2014 1:48pm
I am also interested in machine learning and I have read probably almost all accessible papers on NN and SVM implementation for financial predictions. Just a heads up, a lot of papers are written about prediction of indices and me trying to apply the same system in a forex environment lead to poor results.
I just happen to be working on my own implementation of a Neural Network which is much more closely modeled to the human brain than common NN's are.
I haven't had time to take a deeper look into the thread yet but since this is one of the topics I am really interested in I might join and catch up soon
I just happen to be working on my own implementation of a Neural Network which is much more closely modeled to the human brain than common NN's are.
I haven't had time to take a deeper look into the thread yet but since this is one of the topics I am really interested in I might join and catch up soon
{quote} So looking at your result it seems you're a step ahead than Daniel, it's really interesting that your NN could be backtested with the same outcome for different runs! About mining bias problem , I don't know if it's a concept that could be applied to NN, btw a nice question to ask DanielCheers, Skyline
- #66
- Dec 9, 2014 1:59pm Dec 9, 2014 1:59pm
- Joined Feb 2014 | Status: ember | 1,268 Posts
{quote} 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 methods is much more complicated. Sure, machine learning methods are simply glorified functions (input to target converters) but sometimes achieving a similar result while understanding everything "behind the scenes" (all variable interactions) is difficult. It is also worth considering that machine...
Cheers,
k
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. Those of you who are keen observers may have noticed that the SVM has two parameters that the K-NN does not have (C and gamma) which are related to how the model classifies data. To make things simple, C is related to how hard the SVM is trained to classify the data and gamma is related to how fine the grain of classification actually is, how much "attention to detail" (those of you versed in statistics might want to give a simple definition that is more accurate if you wish to do so). The bigger the C the harder the training and the larger the gamma the coarser the detail. Using C=10, gamma=0.01 is a good starting point but it might not be the best to obtain the best results. By improving training (increasing C to 20) we actually can obtain better results for our SVM strategy.
Using larger values of C is problematic in this case because the simulation reaches a point where training does not converge (we are demanding too much separation). With this system we finally have a case where the maximum drawdown is not reached within the past 2 years but in 1997. This means that we have an algorithm which does seem to have deteriorated in predictive qualities during the past few years. The SVM used here also uses a Gaussian kernel, which means it can find more complex non-linear relationships, using different Kernel types can also change the distribution of predictions for the SVM.
A question that remains is data-mining bias. What is the probability that an SVM reaches the same level of profitability as above given the same degrees of freedom and a data series that contains no causal past-to-future correlations?
Using larger values of C is problematic in this case because the simulation reaches a point where training does not converge (we are demanding too much separation). With this system we finally have a case where the maximum drawdown is not reached within the past 2 years but in 1997. This means that we have an algorithm which does seem to have deteriorated in predictive qualities during the past few years. The SVM used here also uses a Gaussian kernel, which means it can find more complex non-linear relationships, using different Kernel types can also change the distribution of predictions for the SVM.
A question that remains is data-mining bias. What is the probability that an SVM reaches the same level of profitability as above given the same degrees of freedom and a data series that contains no causal past-to-future correlations?
- #68
- Edited 2:47pm Dec 9, 2014 2:34pm | Edited 2:47pm
@PipMeUp: Thanks for the reference, I will obtain and study the book as soon as I am convinced about the path. (Also, I have a feeling that for the understanding/description of the 6-month drop in eurusd one should probably include the data dating back 2 or 3 years.)
check this reddit for good pointers. with your background, i think it wont take much effort to get started.
http://www.reddit.com/r/MachineLearn..._well_written/
Think : Trade : Live Life
- #71
- Dec 9, 2014 4:47pm Dec 9, 2014 4:47pm
- Joined Aug 2011 | Status: Trader | 1,327 Posts
AlgoTrader how do you deal with multimodal predictions?
No greed. No fear. Just maths.
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 parameters are exactly the same on both currency pair tests. We can see that the results on the USDCHF are also good, although they were initially better and later started to go flat around the year 2007.
It is however interesting to see that the USDCHF strategy has been in fact going up steadily (although slowly) since the 35% drawdown it suffered in 2007. Although there has been heavy intervention by the SNB since 2011 the strategy did not tank, but continued to perform in a rather poor but stable manner (no deeper drawdown reached). On the other hand, a test on the GBPUSD (below) shows that the system behaves badly on this pair.
There is something fundamentally different when comparing the GBPUSD and the USDCHF. Do you think that it is a matter of tuning the model parameters and inputs? Do you think there is something that makes the GBPUSD fundamentally more difficult to predict?
It is however interesting to see that the USDCHF strategy has been in fact going up steadily (although slowly) since the 35% drawdown it suffered in 2007. Although there has been heavy intervention by the SNB since 2011 the strategy did not tank, but continued to perform in a rather poor but stable manner (no deeper drawdown reached). On the other hand, a test on the GBPUSD (below) shows that the system behaves badly on this pair.
There is something fundamentally different when comparing the GBPUSD and the USDCHF. Do you think that it is a matter of tuning the model parameters and inputs? Do you think there is something that makes the GBPUSD fundamentally more difficult to predict?
- #75
- Dec 10, 2014 1:29am Dec 10, 2014 1:29am
- Joined Sep 2012 | Status: Edge,Phsycology And Money Managemen | 1,011 Posts
I am what Many Dream to be but only a few can achieve, im a part of the 1%
- #76
- Dec 10, 2014 3:30am Dec 10, 2014 3:30am
- Joined Aug 2011 | Status: Trader | 1,327 Posts
No greed. No fear. Just maths.
- #77
- Dec 10, 2014 5:17am Dec 10, 2014 5:17am
- Joined Aug 2011 | Status: Trader | 1,327 Posts
No greed. No fear. Just maths.