Causal Identification

A factor with IC 0.04 could be a real effect or a confounded association. Without a DAG and refutation tests, you cannot tell which. Conditioning on the wrong variables - mediators, colliders, post-treatment - can create or destroy apparent signal.

The Problem

"Kitchen sink regression" - conditioning on every available variable - is the default in ML pipelines. But including a collider (e.g., fund flows driven by both momentum and returns) induces spurious correlation (~-0.25 between independent variables). Including a mediator (the channel through which the treatment operates) attenuates the true effect. Including a post-treatment variable introduces bias of unknown sign. The DAG determines which variables are admissible controls.

The Pattern

WRONG

import numpy as np
from sklearn.linear_model import Ridge