Copulas
If you are familiar with how to predict and you have noticed the zstreams that are created, you might consider some more advanced functionality. The /copula API serves two purposes:
- Simultaneous publication of multiple streams
- Simultaneous publication of multiple streams with associated z2 or z3 `copula’ streams
The following family of streams illustrates the pattern.
| Type of stream | Examples |
|---|---|
| Regular #1 | c5_bitcoin.json |
| Regular #2 | c5_ethereum.json |
| Regular #3 | c5_cardano.json |
| Z-scores stream #1 | z1~c5_ethereum~3555 |
| Bivariate copula #1,#2 | z2~c5_bitcoin~c5_ethereum~3555 |
| Trivariate copula (70s) | z3~c5_bitcoin~c5_cardano~c5_ethereum~70 |
| Trivariate copula (1 hr) | z3~c5_bitcoin~c5_cardano~c5_ethereum~3555 |
To reiterate there are NO MULTIVARIATE streams on microprediction.org. However there are univariate streams whose values represent a point in \(R^2\) or $$R^3$. The mapping from one dimension to two or three is implied. The former is used by streams prefixed by z2~, the latter by streams prefixed by z3~.
Python
Very similar to the use of set()
from microprediction import MicroWriter
mw = MicroWriter(write_key='YOUR WRITE KEY HERE')
names = ['my_stream_a.json','my_stream_b.json','my_stream_c.json', 'my_stream_d.json']
values = [5,6,4,9]
res = mw.cset(names=names,values=values)
Python examples
See /submission_examples_copulas.
The meaning of values in z2~ streams
Probably you’ve already grok’d zscores, and the bivariate and trivariate counterparts are not dissimilar:
\[z(x_1,x_2) = \Phi^{-1} \left( H\left( F^1_{70}(x_1), F^2_{70}(x_2) \right) \right)\]where:
- $F^1_{70}(x_1)$ is the community distributional transform applied to a published data point $x_1$ for one stream (using predictions for the 70 second horizon);
- similarly for $F^2_{70}(x_2)$;
- The function $H:[0,1]^2 \rightarrow [0,1]$ is the inverse of a space-filling curve described presently; and
- as before $\Phi$ is the standard normal distribution function.
The choice of space-filling curve H
As for H, a Morton Z-curves are employed. Actually the most up to date reference for these embeddings is the code, so see zcurve_conventions. At time of writing these docs, the heart of this calculation is:
pymorton.interleave(prctls)
where ‘prctls’ comprisese the pair \(F^1_{70}(x_1), F^2_{70}(x_2)\) in the bivariate case, and similarly for trivariate.
-+-
Documentation map
Edit this page.