Signal Decay and the Cadence of Model Refresh

Alpha is perishable

In systematic research, a signal’s useful life often begins to shorten the moment it is discovered. Competitors adapt, market microstructure shifts, and the statistical edge that looked clean in sample can decay out of sample—sometimes slowly enough that teams keep trusting a model that no longer earns its risk budget.

Sophron treats signal decay as an operating assumption, not a surprise. Sophron AI helps monitor live feature drift and prediction quality; people decide when a model is paused, resized or retired.

A practical refresh cadence

1. Discovery hygiene — document the economic story, not only the t-stat.
2. Holdout discipline — keep a sealed evaluation window that research cannot “peek” into casually.
3. Live monitoring — track information coefficient, turnover and capacity against pre-agreed floors.
4. Kill criteria — write the retirement rule before go-live, so emotion does not extend a dead edge.

What we refuse to automate

Refreshing a model is not the same as letting an agent continuously rewrite production logic without review. Automation can propose candidates and stress them; promotion into a live book still requires an accountable owner and a written invalidation path.

Quant research that ignores decay is marketing. Quant research that plans for decay is institutional process.