An arXiv preprint reports that the speed at which S&P 500 stocks change the way they move together is associated with a higher value in a particular volatility-premium measure. In the study’s three-month index, the estimated coefficient was +0.195, with a t-statistic of +5.40 and a 95% confidence interval from 0.124 to 0.266, across 365 observations. The result is noncausal: it shows a relationship, not that reconfiguration creates the premium.
The finding applies to the paper’s own premium object, which it explicitly says is not the canonical variance risk premium. The measure compares one month of implied variance—a market-implied volatility measure—with the slow, broad realized-volatility environment of the constituent cross-section.
A measure of changing co-movement
The analysis used monthly total returns for S&P 500 constituents, including names later delisted, from June 1994 through December 2025. After filtering, the panel contained 379 months and 430 names.
The index, called REC, is defined as the mean squared sine of the principal angles between two subspaces. Put more plainly, it measures how much the underlying directions of co-movement rotate between the periods being compared.
REC was weakly related to level measures: no level correlation exceeded 0.32, and the implied-correlation surface spanned at most 6.7% of the index. Much of the measure was therefore absent from the traded correlation surface rather than simply duplicating a broad volatility or correlation reading.
When one-month and 12-month rotation measures were entered together, the one-month measure retained a t-statistic of +4.76, while the 12-month measure fell to +1.38. The pattern favors the pace of structural revision over cumulative distance as the component associated with the premium.
The premium moves on two clocks
The two legs of the premium moved on different clocks. The relationship was concentrated in implied variance at the same time as reconfiguration; realized variance built later, and the premium turned negative, reaching a trough at eight months before fading at month 11.
In a forward test against a simulated no-lead–lag benchmark, REC remained associated with broad realized variance from one through 10 months ahead, peaking around months seven to eight. The peak increase in log realized variance was 0.155; by months 11 and 12, the association was no longer detectable. The benchmark preserved contemporaneous correlation while removing lead–lag predictability, and the 10-month result was sensitive to the autoregressive specification.
The implied-correlation curve also showed a shape effect. The difference between its three-month and one-month measures was inversely associated with smoothed REC, with a t-statistic of −3.46 and a coefficient of −0.910 index points per standard deviation, even though the curve’s level did not span the index.
What the index did not deliver
The forward relationship did not translate into an advantage in the study’s timing comparison. Across 329 months, the equal-volatility paired comparison favored the unconditional position, with a paired t-statistic of −1.85; the reported Sharpe ratios were 1.49 for the unconditional position and 1.36 for the scaled one.
Nor did the index support a crash-hedge interpretation. In the worst 5% of months for short-variance positions, mean standardized REC was −0.47, so rotation was below average when those positions suffered most.
The downside test could not isolate a separate structural effect. With 12-month windows and roughly six down observations per name in each window, the analysis could not distinguish downside co-movement from downside intensity at monthly resolution. That is a limit of identification in this design, not evidence that the two features can never be separated.
The economic attribution was similarly narrower than a trading claim. In the highest-rotation quartile, mean payoff was +0.0410, capture was 52.7%, and P&L share was 64%. The paper says this payoff is nonstandard and not a tradeable variance-swap backtest.
Taken together, the findings describe an association within the paper’s own premium object, not evidence of a causal effect, timing advantage or crash hedge. The manuscript is an arXiv preprint, version 1, dated 20 August 2026.
Paper data and sources
Original title: The Reconfiguration Premium: Co-movement Structure as an Unspanned Dimension of the Variance Risk Premium
Authors: Lucas Carvalho
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text