Volatility Regime Adjustment in Risk Models

A standard factor risk model estimates each factor's volatility from a long history of returns — often a year or more — because long histories give statistically stable estimates. The tradeoff is responsiveness: when markets suddenly turn turbulent, a model built on a long trailing window keeps predicting yesterday's calmer volatility for weeks, understating risk exactly when it matters most.

Volatility regime adjustment (VRA) fixes this with a multiplicative overlay computed separately from the main model. It compares how large realized market moves actually were, over a recent short window, against how large the model predicted they should be — a ratio sometimes called the bias statistic. If realized volatility has been running, say, 40% hotter than the model's forecast over the last month, the VRA scales every factor volatility (and often the whole covariance matrix) up by a factor tracking that gap, so the model's risk numbers catch up to current conditions without needing the slow-moving long-window estimates to be rebuilt from scratch.

This is a patch on top of the model rather than a redesign of it: it does not change the model's factor exposures or correlations, only rescales overall volatility, so it responds fast to shocks but shouldn't be relied on to fix a genuinely broken covariance structure.