Volatility Disagreement in the Options Market

Bibliographic record. Follow the original-source link for the publication.

Field	Value
Primary domain	Hedging Exposure Risk
Other domains	—
Methods	Financial Ml
Facets	Structure Straddle
Authors	Turan G. Bali, Bryan T. Kelly, Mathis Mörke
Published	2026-07
Source	NBER Asset Pricing
Identifiers	doi:10.3386/w35500, nber:w35500
URL	Open original source

Editorial synthesis

Why it matters

The paper adds dispersion in ML variance forecasts to single-stock option research, offering a candidate second-moment signal beyond return disagreement. VDIS measures simulated forecaster dispersion, not directly observed investor beliefs or a demonstrated executable edge. (full_text:S33, full_text:S186, full_text:S187, full_text:S198, full_text:S257, full_text:S151)

Main author claims

• The authors construct VDIS from 100 random-forest forecasters with half-feature subsets and heterogeneous hyperparameters. They fit holding-period log realized variance, convert forecasts to the variance scale, and measure their cross-sectional dispersion. (full_text:S193, full_text:S194, full_text:S198, full_text:S211, full_text:S213, full_text:S226, full_text:S227)
• The authors report: The reported equal-weighted high-minus-low VDIS decile spread is −5.14% monthly in delta-hedged straddle excess returns (Newey–West t = −7.20), with OCAPM alpha of −6.68%. Conditional sorts also retain negative spreads across several established option-return predictors. (full_text:S257, full_text:S265, full_text:S267, full_text:S362, full_text:S363, full_text:S365, full_text:S1019, full_text:S1021)
• The authors report: With stock and month fixed effects but no additional controls, the authors report about 30% higher opening volume per standard deviation of VDIS (β = 0.26). Adding controls reduces β to 0.12 (t = 9.80). (full_text:S574, full_text:S575, full_text:S576, full_text:S578, full_text:S579, full_text:S580, full_text:S581, full_text:S1090)

Data, method, or discussion scope

This NBER working paper, not a peer-reviewed journal article, reports a 2001–2022 US single-stock option sample with 5,678 underlyings, using OptionMetrics, CRSP, and Jensen characteristics, excluding financials, utilities, and stocks below $5. It describes five-year rolling training and annual refits, prior-calendar-month stock inputs, and pre-formation-Friday option inputs. Opening-volume data cover four Nasdaq exchanges from May 2005 to February 2021. Neither the empirical results nor underlying PIT data were independently audited here. (full_text:S6, full_text:S7, full_text:S127, full_text:S130, full_text:S131, full_text:S132, full_text:S133, full_text:S135, full_text:S205, full_text:S206, full_text:S214, full_text:S217, full_text:S221, full_text:S1127)

Main limitations

Investor wealth shares are unobserved and proxied with past variance. Simulated dispersion, conditional sorts and historical-volatility/GARCH benchmarks do not identify mispricing causes or exclude omitted risk. Returns use initial straddle value; margin/arbitrage-cost sorts are not net funded-capital returns after spreads, hedging and financing. AFD-orthogonalized returns differ between body and appendix (−4.09% versus −3.76%), with further caption inconsistencies; these unresolved details are not verified robustness evidence. (full_text:S535, full_text:S536, full_text:S539, full_text:S187, full_text:S592, full_text:S599, full_text:S603, full_text:S151, full_text:S657, full_text:S669, full_text:S674, full_text:S675, full_text:S334, full_text:S1156, full_text:S1167, full_text:S1168, full_text:S1169)

Relationships

• None recorded.