Monday 16 September 2019

The Ideal Tau for Time Series Embedding?

In my Preliminary Test Results of Time Series Embedding post I got a bit ahead of myself and mistakenly quoted the ideal embedding length (Tau) to be half a period for cyclic prices. This should actually have been a quarter of a period and I have now corrected my earlier post.

This post is about my further investigations, which uncovered the above, and is written in a logical progression rather than the order in which I actually did things.

It was important to look at the Tau value for ideal sinusoidal data to confirm the quarter period, which checked out because I was able to easily recreate the 2D Lissajous curves in the Novel Method for Topological Embedding of Time-Series Data paper. I then extended this approach to 3D to produce plots such as this one, which may require some explaining.
This is a 3D Phase space plot of adaptive Max Min Normalised "sinusoidal prices" normalised to fall in the range [ -1 +1 ] and adaptive to the known, underlying period of the sine wave. This was done for prices with and without a "trend." The circular orbits are for the pure sine wave, representing cyclic price action plotted in 3D as current price, price delayed by a quarter of a cycle (Tau 1) and by half a cycle (Tau 2). The mini, blue and red "constellations" outside the circular orbits are the same pure sine wave(s) with an uptrend and downtrend added. The utility of this representation for market classification etc. is obvious, but is not the focus of this post.

Looking at Tau values for ideal prices with various trends taken from real market data, normalised as described above, across a range of known periods (because they were synthetically generated) the following results were obtained.
This is a plot of periods 10 to 47 (left to right along the x-axis) vs the global tau value (y-axis) measured using my Octave version of the mdDelay.m function from the mdembedding github. Each individual, coloured line plot is based on the underlying trend of one of 66 different, real time series. Each point per period per tradeable is the median value of 20 Monte Carlo runs over that tradeable's trend + synthetic, ideal sine wave price. The thick, black line is the median of the median values per period. Although noisy, it can be seen that the median line(s) straddle the theoretical, quarter cycle Tau (0.25) quite nicely. It should be remembered that this is on normalised prices. For comparison, the below plot shows the same on the unnormalised prices.
where the Tau value starts at over 0.5 at a cyclic period of 10 at the left and then slowly decreases to just over 0.1 at period 47.

More in due course.

No comments: