to AlphaNova mofos

With the package, you can run:

# Full walk-forward + city/novelty metrics
python runner.py demo.py --gauge-fix

You can also run this using your own predictor.

The output includes a city-space novelty metric, which estimates how close your signal is to the existing signals we provide.

A signal can be thought of as a cross-sectionally demeaned vector time series:

x_t \in \mathbb{R}^{20}

with

\sum_{i=1}^{20} x_{t,i} = 0

for each time step ( t ).

After demeaning, normalization, and a few additional geometric transformations, we map the signal into a sequence of points on the unit sphere:

p_t \in S^{18}

where

S^{18} = \left\{ x \in \mathbb{R}^{19} : |x| = 1 \right\}

Intuitively, the original 20-dimensional instrument space loses one degree of freedom from demeaning, and another from normalization.

A city is simply the normalized average of these spherical points:

c = \frac{\sum_t p_t}{\left|\sum_t p_t\right|}

This gives a compact geometric representation of the entire signal.

The similarity between two cities is measured using cosine similarity:

\mathrm{sim}(c_1, c_2) = c_1^\top c_2

or equivalently by the angular distance:

\theta = \cos^{-1}(c_1^\top c_2)

Empirically, if your signal is sufficiently far from the provided cities in this space, then there is a good probability that it will also exhibit low correlation with those signals in practice.

As a rough rule of thumb, you generally want your city to be at least:

\theta \geq 60^\circ

away from existing cities, corresponding to:

c_1^\top c_2 \leq 0.5

The package also includes the raw ambient coordinate city representations in a parquet file. Those can be explored directly, potentially with the assistance of an agentic workflow for automated novelty search or signal discovery.

to AlphaNova mofos

2 Replies