Ang, Hodrick, Xing & Zhang (2006) — The Cross-Section of Volatility and Expected Returns

asset pricing
volatility
cross-section
The idiosyncratic-volatility puzzle: high-idio-vol stocks earn abnormally LOW returns. Notes, critique, and how it held up.
Author

David Maguire

Published

June 28, 2026

Citation. Ang, A., Hodrick, R. J., Xing, Y., & Zhang, X. (2006). The Cross-Section of Volatility and Expected Returns. The Journal of Finance, 61(1), 259–299.

One-line takeaway. Stocks with high idiosyncratic volatility earn abnormally low average returns — the opposite of what a risk-based story predicts, and robust enough to launch a decade of follow-up work.

What the paper claims

Two related results. First, stocks with high sensitivity to aggregate volatility innovations (proxied by changes in the VIX) earn low average returns — consistent with aggregate volatility being a priced state variable investors will pay to hedge. Second, and more famously, stocks with high idiosyncratic volatility relative to the Fama–French three-factor model earn low average returns. The second result is the puzzle: idiosyncratic risk is diversifiable, so standard theory says it shouldn’t be priced at all — and if anything, under-diversification would predict a positive premium, not a negative one.

How they show it

Idiosyncratic volatility is measured as the standard deviation of residuals from a Fama–French 3-factor regression on daily returns within each month. Stocks are sorted into quintile portfolios on this measure and held the next month (value-weighted); the strategy is rebuilt monthly. Aggregate volatility risk is measured by sorting on stocks’ loadings on \Delta \text{VIX} (a factor they label FVIX). The headline is a spread portfolio: quintile 5 (high idio-vol) minus quintile 1 (low).

Key numbers

The high-minus-low idiosyncratic-volatility quintile spread is about −1.06% per month, and it stays large and significant — on the order of −1.3% per month — after adjusting for the FF3 factors. It survives controls for size, book-to-market, momentum, liquidity, volume, turnover, and leverage. The aggregate-volatility result is a similarly-sized negative premium on the \DeltaVIX loading.

What I’d push on

  • Measurement fragility. The result is sensitive to design choices. Bali & Cakici (2008) show that switching between value- and equal-weighting, using different breakpoints (e.g. NYSE vs all stocks), or screening out low-priced / micro-cap names materially weakens or removes the spread. A replication has to fix these knobs before looking at returns.
  • Realised vs expected idio-vol. The sort uses last month’s realised idiosyncratic volatility as the conditioning variable. Fu (2009) argues that expected idiosyncratic volatility (EGARCH-modelled) is positively priced, and that the negative result is partly a mechanical consequence of using lagged realised vol when idio-vol is time-varying. This is the critique I find most interesting and would test first.
  • Microstructure. Han & Lesmond (2011) attribute much of the effect to liquidity and bid-ask bounce contaminating the idio-vol estimate — a bias, not a premium.
  • Is it really idio-vol, or lottery demand? Bali, Cakici & Whitelaw (2011) show a closely related MAX effect (low returns to stocks with extreme recent daily returns); controlling for MAX substantially reduces the idio-vol spread, pointing to a behavioural lottery-preference channel rather than volatility per se.
  • Arbitrage asymmetry. Stambaugh, Yu & Yuan (2015) reconcile the sign: idio vol proxies for arbitrage risk, and because shorting is harder than buying, overpricing dominates — so high-idio-vol stocks are net overpriced and earn low returns, strongest among hard-to-short names and in high-sentiment periods.

How I’d replicate it

  1. CRSP daily returns → monthly FF3 residual volatility per stock; form value-weighted quintiles; compute the 5−1 spread and its FF3 alpha.
  2. Robustness grid: value vs equal weight × NYSE vs all-stock breakpoints × with/without a $5 price and micro-cap screen. Report the full grid, not the best cell — this is exactly the multiple-testing trap from my backtest-overfitting note.
  3. Horse-race realised idio-vol against an EGARCH expected-idio-vol measure (Fu) and against MAX (Bali–Cakici–Whitelaw) in the same regression.