and

The function outputs are normalised price and price delayed by a quarter and a half of the cycle period (in this case 20.) The trend slopes have been chosen to exemplify the difference in the features between up and down trends. The function outputs are identical for the unseen cyclic prices to the left.

When the sets of function outputs are plotted as 3D phase space plots typical results are

with the green, blue and red phase trajectories corresponding to the cyclic, up trending and down trending portions of the above synthetic price series. The markers in this plot correspond to points in the phase trajectories at which turns in the underlying price series occur. The following plot is the above plot rotated in phase space such that the green cyclic price phase trajectory is horizontally orientated.

It is clearer in this second phase space plot that there is a qualitative difference in the phase space trajectories of the different, underlying market types.

As a final check of feature relevance I used the Boruta R package, with the above turning point markers as classification targets (a turning point vs. not a turning point,) to assess the utility of this approach in general. These tests were conducted on real price series and also on indices created by my currency strength methodology. In all such Boruta tests the features are deemed "relevant" up to a delay of five bars on daily data (I ceased the tests at the five bar mark because it was becoming tedious to carry on.)

In summary, therefore, it can be concluded that features derived using Taken's theorem are useful on financial time series provided that:

- the underlying time series are normalised (detrended) and,
- the embedding delay (Tau) is set to the theoretical optimum of a quarter (and multiples thereof) of the measured cyclic period of the time series.

## No comments:

Post a comment