UQO Working Paper No. 7· 29 pages
The Options-Implied Information Content for Cross-Asset Return and Volatility Prediction: Evidence from 3.8 Billion Option Contracts
What 3.8 Billion Option Contracts Know About Tomorrow
Do option-implied moments predict returns and volatility across assets? Weekly returns yes, daily no — and implied volatility dominates every realized-volatility benchmark.
Simon-Pierre Boucher — contact@spboucher.ai
3.83B
Option contracts analyzed (11,077 underlyings, 2010–2025)
2.33
Sharpe ratio of the implied-kurtosis long/short strategy (t = 19.8)
+23.3%
Gain in 1-day RV forecasting R² from adding the IV surface to HAR
100%
Share of tickers where IV Granger-causes realized volatility (F = 62.4)
73.8%
RV forecast-error variance explained by IV shocks at 20 days (FEVD)
47.0%
Max-OI "price magnet" hit rate — below 50%, hypothesis rejected
Abstract
This paper asks five questions about the information content of equity option markets, answered with 3.83 billion option contracts on 11,077 underlyings (2010–2025) merged with 11.5 billion intraday OHLCV observations. The estimation panel covers 264,383 ticker-days across 69 tickers.
Implied moments predict cross-sectional returns at the weekly horizon (R² of 4.8–19.3%) but not daily; a long/short portfolio sorted on implied kurtosis earns a Sharpe ratio of 2.33 (t = 19.8). Adding the IV surface to a HAR model raises 1-day realized-volatility forecasting R² by +23.3%, robust across all nine subperiods, and IV Granger-causes RV for 100% of tickers.
The implied-minus-realized correlation ratio predicts market stress at 5–20 day horizons (t-statistics 3.91–8.30). The popular max-open-interest "price magnet" hypothesis is rejected (47.0% hit rate, below the 50% coin flip), and on the SPX surface a simple HAR-RV model beats random forests, gradient boosting, and the VIX out of sample.
Key Results
Weekly, not daily, return predictability
Implied moments predict cross-sectional returns at the weekly horizon (R² 4.8–19.3%) but carry no daily signal; the kurtosis long/short portfolio delivers a 2.33 Sharpe ratio.
IV surface dominates HAR and GARCH
Augmenting HAR-RV with implied-volatility surface features improves 1-day forecasting R² by +23.3%, a gain that survives in all nine subperiods examined.
Implied–realized correlation as a stress gauge
The ratio of implied to realized correlation predicts market stress at 5–20 day horizons, with t-statistics ranging from 3.91 to 8.30.
Price-magnet hypothesis rejected
Prices gravitate toward maximum-open-interest strikes only 47.0% of the time — below chance — refuting a widespread piece of options-market folklore.
Simple models beat ML on the SPX surface
For SPX realized-volatility forecasting, HAR-RV beats random forests, gradient boosting, and the VIX out of sample; two-week ATM IV alone captures 50.8% of feature importance.
Placebo-validated information content
A placebo design confirms the results are not mechanical: shuffled predictors yield R² of 0.0008 against 0.053 for the actual implied moments.
Data
Equity options database
3.83 billion option contracts across 11,077 underlyings, 2010–2025, stored in external DuckDB stores and distilled into five derived parquet files.
Intraday OHLCV database
11.5 billion intraday open-high-low-close-volume observations merged with the options data to construct realized-volatility measures.
Estimation panel
264,383 ticker-days over 69 tickers, 2010–2025 — the analysis sample behind all cross-sectional and time-series results.
Methodology
Fama-MacBeth cross-sectional regressions
Panel regressions of daily and weekly returns on option-implied moments, with Newey-West, clustered-SE, and quantile robustness checks.
HAR-RV and GARCH forecasting horse race
Realized-volatility forecasts from HAR and GARCH benchmarks compared against IV-surface-augmented models across nine subperiods and rolling windows.
Granger causality, VAR, IRF and FEVD
Vector autoregressions establish that IV leads RV for every ticker; forecast-error variance decompositions attribute 73.8% of 20-day RV variance to IV shocks.
Portfolio sorts and double sorts
Single and double sorts on implied moments, including decile sorts and Spearman information coefficients, generate the long/short trading strategies.
Machine learning on the SPX surface
Random forests and gradient boosting trained on implied-volatility surface features, benchmarked against HAR-RV and the VIX out of sample.
Extended robustness battery
Winsorization and Newey-West lag sensitivity, leave-one-year-out stability, and a placebo test with shuffled predictors validate the headline findings.
Reproducibility
- Fully scripted pipeline: 13 numbered Python entry points (make pipeline | figures | paper) with a reusable src/wp7 package.
- All 28 regenerable result files were re-run and verified on 2026-08-05: 10 byte-identical, 15 equal to floating-point noise, 1 explained deviation, 2 newly regenerated.
- Every table in the 29-page paper derives from one of 41 shipped CSV result tables; raw-dependent steps skip gracefully when the DuckDB stores are absent.
- Reproduction evidence (rerun logs and a CSV comparator) ships in _verify/, with a full audit trail in AUDIT.md and CHANGES.md.
- LaTeX source is modular (one file per section, 45 BibTeX references, all cited) and builds with latexmk via make paper.