← Research logMachine Learning
KILLED · scaling
#

A GP with a tuned kernel gives calibrated predictive means and variances for returns, letting us trade only high-confidence (low-variance) forecasts.

Posterior mean and variance at a test point:

$$ \mu_* = k_*^\top (K+\sigma^2 I)^{-1} y,\qquad \sigma_*^2 = k_{**} - k_*^\top (K+\sigma^2 I)^{-1} k_* $$

GP with RBF + periodic kernels on feature windows; trade only when predictive variance is low. Costs applied.

Predictive mean signal≈ 0
Low-variance subset edgenone
$O(n^3)$ inverse costprohibitive
KILLED
The calibrated uncertainty is nice but the mean is still ~zero (same target problem as N-038), and the $O(n^3)$ covariance inverse makes rolling refits impractical. No edge, high cost. Killed.
Well-calibrated uncertainty around a zero-mean forecast is still a zero-mean forecast. Confidence about "no signal" is not tradeable.

We publish the failures too.

This is one of 100+ documented hypotheses. Browse the full lab notebook, or see the strategies that survived into production.