UQO Working Papers · WP07 · v1.2
The Options-Implied Information Content for Cross-Asset Return and Volatility Prediction: Evidence from 3.8 Billion Option Contracts
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
The option surface forecasts variance, not returns
Implied features explain 0.3% of 5-day stock returns and only one of fourteen strategies survives a 10 bp cost, while the surface lifts the HAR-RV R² from 0.421 to 0.515 at the weekly horizon in every subperiod.
Abstract
This paper measures the information content of the implied volatility surface for returns and realized variance in a panel of 69 liquid US underlyings (49 large stocks, 17 ETFs and three equity indices) over 2010–2025, built from 3.83 billion end-of-day option records and 5-minute prices. Option-implied features predict the cross-section of weekly returns in a statistically detectable but economically marginal way: they explain 0.3% of the variation in 5-day stock returns, 0.6% for ETFs and 1.0% for indices; under two-way clustered standard errors only the 25-delta skew remains significant, with a positive sign consistent with compensation for skewness risk; and of fourteen long-short strategies implemented in calendar time with measured turnover, only a sort on the implied volatility level is profitable after a 10 basis-point cost, with a t-statistic of 2.2 that does not survive a multiple-testing correction.
The surface is more informative about second moments. On a common sample it raises the in-sample R² of the HAR-RV model from 0.415 to 0.457 for next-day realized variance and from 0.421 to 0.515 for weekly realized variance, in every one of nine subperiods; out of sample the gain is reliable under mean squared error but not under QLIKE, and at the weekly horizon the level of implied volatility alone is the best forecaster under both losses. The ratio of implied to realized correlation among S&P 500 constituents predicts stress episodes at 5- to 20-day horizons beyond the VIX (t of 2.3–4.1 with overlap-robust standard errors). At-the-money implied volatility Granger-causes realized variance in every ticker, but implied-volatility shocks account for 20–29% of its 20-day forecast-error variance.
Open-interest concentration does not attract the underlying price, and tree-based learners trained on the SPX surface do not beat HAR-RV out of sample. The revision corrects a target-alignment error in the previous version, whose weekly return results (R² of 5–19% and a long-short Sharpe ratio of 2.33) are withdrawn.
Key findings
3.83 billion
end-of-day option records behind the panel
69
liquid US underlyings, 264,383 ticker-days, 2010–2025
0.3%
R² of implied features for 5-day stock returns
0.421 → 0.515
HAR-RV R² for weekly realized variance with the surface
1 of 14
long-short strategies profitable after 10 bp costs
20–29%
20-day RV forecast-error variance due to IV shocks
Scientific contributions
- Return predictability is marginal and mostly non-tradable
- Weekly R² is 0.29% for stocks, 0.61% for ETFs and 1.01% for indices; only the 25-delta skew survives two-way clustering (+19 bp per standard deviation, t = 2.91), and daily predictability is absent.
- Calendar-time strategies with measured turnover
- Fourteen long-short sorts are implemented à la Jegadeesh–Titman with Newey–West t-statistics; only the ATM implied-volatility sort keeps a positive net Sharpe ratio (0.34 at 10 bp, break-even cost 27 bp, gross t = 2.19).
- The surface improves HAR-RV in every subperiod
- On a common 119,093-observation sample the in-sample R² rises by 9.9% at one day and 22.3% at five days; out of sample the median MSE ratio is 0.967 (better for 14 of 30 tickers) but the QLIKE ratio is 1.17 (worse for 15).
- Implied correlation signals market stress
- The implied-to-realized correlation ratio among S&P 500 constituents predicts stress at 5, 10 and 20 days beyond the VIX, with Newey–West t-statistics of 3.27, 4.13 and 2.33 over February 2016–December 2025.
- Granger causality without dominance
- ATM implied volatility Granger-causes realized variance in 100% of tickers (mean F = 62.4), yet Cholesky-orthogonalised FEVD attributes only 20.2–28.5% of the 20-day RV variance to IV shocks.
- A corrected and fully regenerated paper
- The 5-day targets of v1.1 were trailing windows; with strictly forward targets the weekly R² of 5–19%, the kurtosis-sort Sharpe of 2.33 and the R²_OOS of 0.99 disappear, and all 27 tables are now generated from the result files.
Methodology
Methodology (5)
Cross-sectional return regressions
Pooled OLS with HC1, Newey–West and two-way clustered errors plus Fama–MacBeth with Newey–West lags, on strictly forward 1- and 5-day returns; quintile and decile sorts with overlap-robust t-statistics.
Calendar-time portfolio strategies
Overlapping-cohort long-short portfolios for every sort at 1- and 5-day horizons with measured turnover, net Sharpe ratios at 5–50 bp and break-even costs; Bonferroni adjustment across strategies.
HAR-RV versus implied-volatility forecasting
HAR-RV with daily, weekly and monthly components against IV-augmented models on a common sample; rolling out-of-sample windows (500/250), MSE, QLIKE and Diebold–Mariano tests with Newey–West variance for 30 tickers.
Correlation divergence and VAR analysis
Equal-weight implied correlation versus 22-day realized correlation in three separate stress specifications with Newey–West errors; per-ticker Granger tests and bivariate VAR(5) with Cholesky-orthogonalised IRF and FEVD under both orderings.
Robustness battery
Nine subperiods, VIX regimes, leave-one-year-out, winsorisation, alternative lag lengths, Spearman information coefficients, controls and a within-ticker permutation placebo (R² 0.0004 versus 0.0031).
Research data
- Option chains and intraday prices
- 3.83 billion end-of-day option records with Greeks for 11,077 underlyings and 11.5 billion 1- and 5-minute OHLCV rows (2010–2025) from the HF Market Data lake, stored in DuckDB.
- Analysis panel
- 264,383 ticker-days for a fixed ex-post universe of 69 liquid names (119,081 with all ten predictors): ATM IV at 30 and 90 days, term slope, 25-delta skew, smile-based skewness and kurtosis proxies, put–call ratios, net gamma, and 5-minute realized variances; five derived parquets shipped.
Reproducibility
- Parquet-only pipeline (scripts 02–14) reproducible in under two minutes via make pipeline; WP7_LEGACY_TARGETS=1 reproduces the mis-aligned v1.1 targets for audit
- All 27 LaTeX table fragments generated by scripts/14_make_paper_tables.py; no hand-typed numbers except the frozen raw-data inventory
- v1.1 result CSVs preserved in results/_pre-revision-20261017/; Crossref verification logs for all 97 references in docs/verify/
- Full LaTeX source with an anonymised version (main_anon.pdf) for journal submission; pre-revision PDF and review report in docs/
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.2current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 53 p. |