Horizon Design

An arbitrary 1-day horizon forces daily rebalancing, which costs 2-5% annually in transaction costs. If the signal's IC peaks at 20 days, you are paying for turnover that destroys the edge.

The Problem

The prediction horizon determines everything downstream: label construction, feature relevance, turnover, and whether transaction costs leave any alpha. Choosing it arbitrarily - or defaulting to "1 day because that is what everyone uses" - misaligns the model with the actual signal dynamics.

The Pattern

WRONG

import numpy as np

# Arbitrary 1-day horizon - no evidence this matches the signal
labels = np.roll(returns, -1)  # forward 1-day return as label
# Result: high turnover, transaction costs eat the edge