"Trading is statistics and time series analysis." This blog details my progress in developing a systematic trading system for use on the futures and forex markets, with discussion of the various indicators and other inputs used in the creation of the system. Also discussed are some of the issues/problems encountered during this development process. Within the blog posts there are links to other web pages that are/have been useful to me.
Following on from the successful C++ .oct coding of the Gaussian and Binary units, I thought I would conduct a simple visual test of the conditional restricted boltzmann machine, both as a test of the algorithm itself and of my coding of the .oct functions. For this I selected a "difficult" set of price bars from the EURUSD forex pair, difficult in the sense that it is a ranging market with opportunities for the algorithm to fail at market turns, and this is shown below:
For this test there was no optimisation whatsoever; I simply chose a model order of 3, 50 hidden units for both the Gaussian and binary units, the top 500 training examples from my Cauchy-Schwarz matching algorithm, 100 training epochs and 3 sets of features each to model the next day's open, the 3 day maximum high and the 3 day minimum low. These training targets are different from the targets I presented earlier, where I modelled the next 3 OHLC bars individually, because of the results of a very simple analysis of what I will be trying to predict with the CRBM.
The video below presents the results of this visual test. It shows a sliding window of 9 bars moved from left to right over the bars shown in the chart above. In this window the first 6 bars are used to initialise the CRBM, with the 6th bar being the most recent, and the 7th, 8th and 9th bars are the "future" bars for which the CRBM models the open price of the 7th bar and the 3 bar max high and min low of the 7th, 8th and 9th bars. The open price level is shown by the horizontal black line, the max high by the blue line and the min low by the red line.
In viewing this readers should bear in mind that these levels will be the inputs to my MFEMAE indicator and so the accuracy of the absolute levels is not as important as the ratios between them. However, that said, I am quite impressed with this unoptimised performance and I am certain than this can be improved with cross validated optimisation of the various parameters. This will be my focus for the nearest future.
1 comment:
with this code you can do a regression for simple multiple back propagation NN and compare to your CRBM to see if CRBM is really superior.
http://sourceforge.net/projects/mbp/
From my experience RBMs are not superior at all for financial TS
Post a Comment