Over the years I have posted about several different methodologies for creating synthetic data and I have recently come across yet another one which readers may find useful.
One of my first posts was Creation of Synthetic Data, which essentially is a random scrambling of historic data for a single time series with an attempt to preserve some of the bar to bar dependencies based upon a bar's position in relation to upper and lower price envelopes, a la Bollinger Bands, although the code provided in this post doesn't actually use Bollinger Bands. Another post, Creating Synthetic Data Using the Fast Fourier Transform, randomly scrambles data in the frequency domain.
Rather than random scrambling of existing data, another approach is to take measurements from existing data and then use these measurements to recreate new data series with similar characteristics. My Hidden Markov Modelling of Synthetic Periodic Time Series Data post utilises this approach and can be used to superimpose known sinusoidal waveforms onto historical trends. The resultant synthetic data can be used, as I have used it, in a form of controlled experiment whereby indicators etc. can be measured against known cyclic prices.
All of the above share the fact that they are applied to univariate time series only, although I have no doubt that they could probably be extended to the multivariate case. However, the new methodology I have come across is Statistical Mechanics of Time Series and its matlab central file exchange "toolbox." My use of this is to produce ensembles of my Currency Strength Indicator, from which I am now able to produce 50/60+ separate, synthetic time series representing, for example, all the Forex pairs, and preserving their inter-relationships whilst only suffering the computational burden of applying this methodology to a dozen or so underlying "fundamental" time series.
This may very well become my "go to" methodology for generating unlimited amounts of training data.
No comments:
Post a Comment