Conditional and Time-Varying Betas

A stock's beta, estimated once over the last five years, is a single number that implicitly assumes the stock's relationship to the market never changed over that entire period. In reality, most stocks' true sensitivity to the market shifts with conditions: many stocks become more correlated with the market exactly when the market is falling hardest, and less correlated during calm, range-bound periods. A single fixed beta averages over both regimes and gets each one wrong. A conditional, or time-varying, beta model lets the sensitivity itself move with the state of the market.

The analogy

A crowd at a concert behaves in loose, individual ways during a quiet acoustic song, everyone swaying a little differently. But the moment the band drops into the loudest chorus, the whole crowd suddenly moves as one unit, jumping together in near-perfect unison. Measuring "how correlated is any one person's movement with the crowd" using a single average across the whole concert would badly understate how tightly they move together during the loud parts, and badly overstate it during the quiet parts. A stock's beta to the market behaves the same way, quiet markets, individual stock idiosyncrasies dominate; crisis markets, everything moves together, and beta itself rises.

Building the model

The simplest conditional-beta approach re-estimates beta on a short rolling window instead of one long fixed sample:

$\hat{\beta}_t = \frac{\widehat{\text{Cov}}_t(R_i, R_m)}{\widehat{\text{Var}}_t(R_m)},$

where both the covariance and variance are computed using only the most recent $w$ observations (say, the last 60 trading days), re-estimated fresh at every point in time $t$, rather than once over the full multi-year history. In words: instead of one beta covering the whole sample, compute a new beta every day (or week), using only recent data, so the estimate can drift as market conditions genuinely change. More sophisticated versions use exponentially-weighted covariance and variance (see EWMA Volatility and RiskMetrics) so recent observations count more than older ones within the window, or an explicit regime-switching model that estimates a separate beta for "calm" and "crisis" states and a probability of being in each.

Watch how a mean-reverting process can spend stretches drifting near one level before shifting to another, that regime-like behavior is exactly the intuition behind a beta that sits near one value in calm markets and jumps to another during stress, rather than staying fixed forever.

Worked example 1: rolling beta before and after a shock

A stock's 60-day rolling beta was $0.9$ heading into a sharp two-week market selloff. Recomputed using the 60 most recent days after the selloff (which now includes those crisis weeks), the rolling beta jumps to $1.6$. A hedge sized using the pre-crisis beta of $0.9$ on a $20m position would have shorted $0.9 \times 20 = 18$, i.e. $18m of index futures; the post-crisis, conditionally-correct hedge would have needed $1.6 \times 20 = 32$, i.e. $32m, a $14m shortfall in downside protection during precisely the period it mattered most.

Worked example 2: a regime-switching estimate

A simplified two-state model estimates $\beta_{\text{calm}}=0.85$ and $\beta_{\text{crisis}}=1.55$, with a filter assigning a $30\%$ probability to the crisis regime today. Blended: $0.30(1.55)+0.70(0.85) = 1.06$, above the calm beta, reflecting the elevated chance of stress. If that probability later rises to $70\%$ on fresh bad news, the blend updates to $0.70(1.55)+0.30(0.85)=1.34$, moving substantially even though neither regime beta itself changed.

What this means in practice

Risk desks that hedge tail risk, rather than average-case risk, generally do not trust a single full-sample beta; they use rolling or regime-conditional estimates specifically because the beta that matters for a crash hedge is the crisis beta, not the long-run average, which a full-sample regression blends together and understates.