📈 In one sentence (revised, peer-reviewed version 2.0) — How much does the implied volatility surface add to HAR-RV, and does the implied-to-realized correlation gap predict market stress? On a 2010–2025 panel of 69 liquid US underlyings (275,448,124 option records) with strictly post-origin targets, the surface raises the within-ticker in-sample R² for next-day variance from 0.319 to 0.360 but does not lower out-of-sample loss (median MSE ratio 1.001, QLIKE ratio 1.284), and the stress association (t = 2.50 at 10 days) does not survive multiple-testing adjustment.
🔬 Peer review at a glance
This paper went through the full multi-referee review of the UQO working-paper series: 6 independent referee reports, 98 review items, an editorial decision (major revision), re-execution of the analysis, a complete revision of text, figures and tables, and an independent verification before sign-off. The revised manuscript and every change live on the branch review-20261103 of this repository (26 commits ahead of main, head 1d90c97, 2026-11-04).
| Dimension | Submitted | Revised | Δ |
|---|---|---|---|
| Originality | 45.5 | 48.0 | +2.5 |
| Rigor | 53.7 | 72.0 | +18.3 |
| Identification | 53.3 | 68.0 | +14.7 |
| Data & reproducibility | 47.5 | 58.0 | +10.5 |
| Robustness | 62.0 | 78.0 | +16.0 |
| Writing | 65.3 | 74.0 | +8.7 |
| Figures & tables | 59.7 | 76.0 | +16.3 |
| Literature | 57.7 | 74.0 | +16.3 |
| Weighted total | 53.0 | 65.3 | +12.3 |
Scientific integrity: PASS · Reproducibility: R2 · partially reproduced · Maturity: grade B
What the review changed
Peer review refocused the paper on two questions, future variance and market stress, and corrected the timing of the targets, the stress indicator and the winsorization, all of which had used information not available at the forecast origin. The out-of-sample forecasting gain and the stress result of the earlier version did not survive; return predictability, price magnets and machine learning were moved to an exploratory appendix, and the paper was retitled.
- Out-of-sample MSE ratio to HAR-RV of 0.967 with 14 tickers better and none worse (30 tickers, full-sample winsorization) → 1.001 with 6 better and 19 worse on 67 tickers (none better after Holm): information without a loss gain
- Stress prediction t-statistics of 3.27, 4.13 and 2.33 → 1.55, 2.50 and 1.73 with a post-origin stress definition; Holm p-values 0.49, 0.10 and 0.42 and bootstrap intervals covering zero: claim not supported
- “Evidence from 3.8 billion option contracts” → 275,448,124 option records actually enter the panel; the larger figure describes the data lake and was removed from the title
- Return predictability, open-interest price magnets and tree learners as research questions → an exploratory appendix; no return predictor survives a family-wise adjustment
- In-sample gain restated on a within-ticker basis: R² for next-day variance 0.319 → 0.360
- VIX series completed for every ticker-day; the stress sample now covers 2010–2025 (3,678 days) rather than starting in 2016
Limitations stated in the revised paper
- The 69 underlyings are a fixed list chosen for liquidity in 2025, not a point-in-time screen, so survivorship bias affects the exploratory return results in particular.
- The stress indicator is defined from the VIX, which is a function of the same index implied volatility that enters the correlation ratio; the alternative based on realized returns gives no result.
- The revision-specific inputs are derived from a licensed data lake and are not committed, so a replication without lake access reproduces the legacy files only.
- An earlier version reporting the withdrawn weekly return results is still distributed on SSRN; the sample covers a single market and sixteen years dominated by an equity bull market.
Where to find the reviewed material
| What | Where |
|---|---|
| Revised manuscript (PDF, 87 pages) | paper/main.pdf on review-20261103 |
| Revised sources, code, figures and tables | branch review-20261103 |
| Commit-by-commit revision history | review-20261103 commits |
All branches (submitted main vs revised) |
branches |
| Interactive reader (figures, tables, citations) | wp.spboucher.ai/papers/wp7/read |
| Paper page, abstract and citation formats | wp.spboucher.ai/papers/wp7 |
🏆 Headline results (revised version)
| Result | What it is |
|---|---|
| 0.319 → 0.360 | within-ticker in-sample R² for next-day realized variance when five surface features are added to HAR-RV, on a common sample of 119,093 ticker-days; the gain is positive for every one of the 69 tickers (median +13.1%) |
| 1.001 / 1.011 | median out-of-sample MSE ratio of the surface-augmented model to HAR-RV for next-day and 5-day variance, on 67 tickers with rolling 500/250-day windows: no loss reduction |
| 6 vs 19 | tickers for which the Diebold–Mariano test favours the augmented model versus HAR-RV; after a Holm adjustment no ticker is improved and 9 are worse |
| 1.284 | median QLIKE ratio: the augmented model is worse for 38 tickers and better for none |
| 45 of 67 | tickers for which the Clark–West statistic exceeds its 5% critical value: the surface contains nested-model information that estimation noise prevents from becoming a forecasting gain |
| +4.20 pts, t = 2.50 | higher 10-day stress probability per standard deviation of the implied-to-realized correlation ratio (3,678 days); Holm p = 0.10, Romano–Wolf p = 0.37 and episode-bootstrap intervals covering zero leave the claim unsupported |
| 18.8–27.8% | share of the 20-day forecast-error variance of realized variance attributed to implied-volatility shocks in a bivariate VAR, depending on the ordering; implied volatility Granger-causes realized variance in every ticker, the reverse in 31.9% |
📄 Abstract
How much does the implied volatility surface add to the HAR-RV model, and does the gap between option-implied and realized correlation among S&P 500 constituents predict market stress? On a 2010–2025 panel of 69 liquid US underlyings (275,448,124 option records, 5-minute prices), with strictly post-origin targets and training-window winsorization, the surface raises the within-ticker in-sample R² for next-day variance from 0.319 to 0.360 and carries nested-model information (Clark–West) for 45 of 67 tickers, but it does not lower out-of-sample loss: the median MSE ratio to HAR-RV is 1.001, the QLIKE ratio 1.284, and no ticker’s next-day forecast improves after a Holm adjustment. A one-standard-deviation higher implied-to-realized correlation ratio is associated with a 4.20 percentage-point higher 10-day stress probability (t = 2.50), but Holm and Romano–Wolf p-values of 0.10 and 0.37 and episode-bootstrap intervals that cover zero leave the stress claim unsupported. Return predictability, price magnets and tree learners are reported as exploratory results.
🧩 Contributions
- Information separated from forecasting gain — On one panel of stocks, ETFs and indices, the surface is evaluated in sample and out of sample under MSE and QLIKE with Clark–West, Giacomini–White and model-confidence-set inference. It finds information without loss reduction; the single 30-day at-the-money level is the only augmentation that stays in the 90% model confidence set for almost every ticker under both losses.
- A correctly timed design — Every target is strictly post-origin (forward realized variance over days t+1 to t+h; stress events on days t+1 to t+h), and winsorization thresholds are estimated on each 500-day training window. Both features correct choices of earlier versions that used information not available at the forecast origin.
- The stress claim tested and not established — With the post-origin definition on the full 2010–2025 series, the coefficients at 5, 10 and 20 days are 1.77, 4.20 and 3.78 percentage points (t-statistics 1.55, 2.50 and 1.73); a stress definition based on realized market returns gives negative and insignificant coefficients. The ratio is reported as a candidate indicator only.
- Stability across periods — The in-sample gain of the surface is present in each of nine subperiods and sixteen leave-one-year-out estimations, while the out-of-sample loss ratio stays between 0.958 and 1.116 across subperiods.
- Earlier claims reported as exploratory, with their decisive tests — No return predictor survives a family-wise adjustment; the only long-short strategy with a factor alpha after costs draws 89.5% of its cumulative return from five names and has a deflated Sharpe ratio probability of 0.00; price moves toward the maximum-open-interest strike are indistinguishable from a permutation null; no tree learner beats HAR-RV.
🗄️ Data
- End-of-day option records — Option chains licensed from FirstRate Data for 69 underlyings from 4 January 2010 to 31 December 2025: 434,957,531 records exist for these underlyings and 275,448,124 pass the feature filters (positive ask, 7 to 180 days to expiration).
- Analysis panel — 264,383 ticker-days for 49 large-capitalization stocks, 17 ETFs and three cash indices (SPX, NDX, RUT), combining daily option features with realized measures from 5-minute bars. The list is fixed and chosen for liquidity, not by a point-in-time screen.
- VIX and correlation series — The daily VIX close is available on every ticker-day of the panel; the implied correlation is an equal-weight equi-correlation on 29 S&P 500 constituents, compared with a 22-day pairwise realized correlation over 3,678 days.
⚙️ Methodology
- HAR-RV against surface-augmented models — HAR-RV is augmented with five surface features (30- and 90-day at-the-money levels, 25-delta skew and two wing-based proxies); pooled within-ticker estimates use Driscoll–Kraay standard errors.
- Out-of-sample evaluation without look-ahead — Rolling 500/250-day windows on 67 tickers with training-window winsorization, MSE and QLIKE losses, Diebold–Mariano, Clark–West and Giacomini–White tests, model confidence sets and a Monte Carlo of the nested comparison under the null.
- Implied correlation and stress regressions — The log ratio of implied to realized correlation predicts VIX-defined stress events within 5, 10 and 20 days, with Newey–West standard errors, an episode bootstrap and an alternative stress definition based on realized market returns.
- Multiple-testing control — Holm and Romano–Wolf adjustments over the nine headline stress tests and Holm adjustment over tickers for the forecast comparisons.
- Lead–lag relations and robustness — Per-ticker heteroskedasticity-robust Granger tests and a bivariate VAR variance decomposition under both orderings with bootstrap bands; nine subperiods, leave-one-year-out estimation, VIX regimes and minimum detectable effects.
♻️ Reproducibility
- Five analysis files are versioned in the repository with checksums, together with every result file and every generated table fragment.
- The inputs specific to the revised analyses (complete VIX series, rebuilt correlation series and other re-extractions) are regenerated from the data lake by logged read-only queries but are not committed; a replication without lake access reproduces the legacy result files, not the revision inputs.
- FirstRate Data licenses the data for private use: raw bars and option records may not be redistributed, and lake access requires a key issued by the operator.
- The original stores of the first extraction were lost; the lake used for the revision is a later vintage, and two frozen exhibits were reproduced on it within 0.018 and 0.008 in R².
- A note documents the differences from versions 1.1 and 1.2, including the earlier out-of-sample scheme reproduced side by side with the corrected one.
Keywords implied volatility surface · realized variance · HAR-RV · implied correlation · market stress · out-of-sample evaluation · multiple testing
JEL G12, G13, G14, G17, C53, C58
📚 How to cite
@techreport{boucher2026wp7,
author = {Simon-Pierre Boucher},
title = {Option-Implied Information about Future Variance and Market Stress: Correctly Timed Evidence from 69 Liquid US Underlyings, 2010–2025},
institution = {Université du Québec en Outaouais},
type = {UQO Working Paper No. 7},
year = {2026},
note = {Version 2.0, November 2026, revised after peer review},
url = {https://wp.spboucher.ai/papers/wp7}
}ℹ️ Below: documentation of the
mainbranch (the version submitted to review). Numbers and claims that the review corrected are superseded by the block above and by the branchreview-20261103.
📈 The Options-Implied Information Content for Cross-Asset Return and Volatility Prediction
UQO Working Paper No. 7 — Evidence from 3.8 Billion Option Contracts
-b31b1b?logo=latex&logoColor=white){kind=icon}
TL;DR — A daily panel of option-implied features and 5-minute realized variances for 69 liquid US underlyings (49 large stocks, 17 ETFs, 3 indices; 264,383 ticker-days, 2010–2025) is built from 3.83 billion option records. Option-implied features predict weekly returns in a statistically detectable but economically marginal way (R² 0.3 % for stocks, 0.6 % ETFs, 1.0 % indices; only the 25δ skew survives two-way clustering; one strategy of fourteen is profitable after 10 bps, with t = 2.2). The IV surface raises the HAR-RV R² from 0.415 to 0.457 (1-day) and 0.421 to 0.515 (5-day) on a common sample, in every subperiod; out of sample the gain holds under MSE but not under QLIKE. The implied/realized correlation ratio predicts stress at 5–20 days (t 2.3–4.1); ATM IV Granger-causes RV in 100 % of tickers but explains 20–29 % of its 20-day FEVD. Price magnets rejected; tree-based ML does not beat HAR-RV.
⚠️ Version 1.2 (2026-10-17) corrects a target-alignment error in v1.1: the 5-day "forward" targets were trailing windows (r_{t−3}+…+r_{t+1}). The v1.1 headline results (weekly R² 5–19 %, kurtosis Sharpe 2.33, R²_OOS 0.99) are withdrawn. See
docs/CHANGELOG-REVISION.md.
📌 Status
| Item | Value |
|---|---|
| Version | 1.2 (October 2026) — revised 2026-10-17 (editorial-board review; docs/REVIEW-REPORT.md, docs/CHANGELOG-REVISION.md) |
| SSRN | 7626643 — uploaded 2026-10-17 with the v1.1 results; must be replaced by paper/main.pdf (v1.2) |
| Journal | target Journal of Futures Markets (double-anonymous; paper/main_anon.pdf); fallback JBF / JFM. Not yet submitted |
| Code & data | parquet-only pipeline reproducible in < 2 min (make pipeline); raw DuckDB stores no longer available |
🎯 What this paper does
Five research questions on the information content of equity option markets, answered on a panel of 69 liquid underlyings:
| RQ | Question | Result (v1.2) |
|---|---|---|
| RQ1 | Do implied moments predict cross-sectional returns, and is it tradable? | Weekly R² 0.3 % (stocks) → 1.0 % (indices); daily ≈ 0; only Skew₂₅δ robust (t_DC 2.91, positive sign); best net strategy (ATM IV sort) Sharpe 0.34 at 10 bps |
| RQ2 | IV surface vs HAR-RV for RV forecasting? | In-sample +9.9 % (1-day) / +22.3 % (5-day) on a common sample, positive in 9/9 subperiods; OOS: better under MSE (14/30 tickers), worse under QLIKE (15/30); IV level alone best at 5 days |
| RQ3 | Implied − realized correlation → market stress? | Ratio predicts stress at 5/10/20 d (t_NW 3.27 / 4.13 / 2.33), Feb 2016–Dec 2025 |
| RQ4 | Greeks by DTE; max-OI "price magnets"? | RV fit rises with DTE (3.8 % → 9.5 %); magnet share 47.0 % < 50 % (frozen output) |
| RQ5 | ML on the SPX surface vs the VIX? | HAR-RV (R²_OOS 0.437) beats VIX (0.336), RF (0.143), GBM (0.095); 2-week ATM IV = 50.8 % of RF importance (frozen output) |
🏆 Headline results
| Object | Estimate |
|---|---|
| Option records / underlyings / period | 3.83 B / 11,077 / 2010–2025 |
| Analysis panel | 264,383 ticker-days, 69 tickers (119,081 with all ten predictors) |
| 5-day return R², stocks / ETFs / indices / all | 0.29 % / 0.61 % / 1.01 % / 0.31 % |
| Only predictor significant under two-way clustering (5-day) | Skew₂₅δ: +19 bps per SD, t_DC = 2.91 |
| Best long-short sort (5-day, calendar time) | ATM IV: 13.6 %/yr, Sharpe 0.54 gross (t_NW 2.19), 0.34 net of 10 bps, break-even 27 bps |
| HAR-RV → HAR-RV + IV surface, 1-day / 5-day R² (common sample) | 0.415 → 0.457 (+9.9 %) / 0.421 → 0.515 (+22.3 %) |
| OOS (30 tickers): median MSE ratio / QLIKE ratio | 0.967 (DM>1.96: 14, <−1.96: 0) / 1.171 (DM>1.96: 0, <−1.96: 15) |
| Implied/realized correlation ratio → stress, t_NW (5/10/20 d) | 3.27 / 4.13 / 2.33 |
| IV → RV Granger causality | 100 % of tickers (mean F = 62.4) |
| RV FEVD explained by IV shocks at 20 d (RV first / IV first) | 20.2 % / 28.5 % |
| Placebo R² (within-ticker permutation) | 0.0004 vs 0.0031 actual |
🗂 Repository layout
options-implied-cross-asset/
├── data/
│ ├── raw/ # external DuckDB stores (not shipped — see data/raw/README.md)
│ └── processed/ # 5 derived parquets (unchanged; forward targets re-derived at load time)
├── src/wp7/ # analysis library (config, data_io, econometrics, forecasting, portfolio)
├── scripts/ # numbered entry points 01–14 (table below)
├── results/ # CSV result tables (every paper table derives from these)
│ └── _pre-revision-20261017/ # v1.1 CSVs kept for audit
├── figures/ # supplementary figures (the paper itself is tables-only)
├── docs/ # REVIEW-REPORT.md, CHANGELOG-REVISION.md, AUDIT.md, CHANGES.md, verify/
├── Makefile # make pipeline | figures | paper | all
└── paper/ # main.tex + preamble.tex + sections/ + tables/ (generated) + appendix/ + references.bib → main.pdf🚀 Quick start
python3 -m pip install -r requirements.txt # scripts bootstrap src/ themselves
make pipeline # scripts 02–13 (raw-dependent steps skip themselves gracefully)
python3 scripts/14_make_paper_tables.py # regenerate paper/tables/*.tex from results/
make figures # supplementary figures from the result CSVs
cd paper && make # latexmk → main.pdf (latexmk -pdf main_anon.tex for the anonymous version)| Script | Question | Needs raw stores? |
|---|---|---|
01_extract_data.py |
raw → processed parquets | yes (entirely) |
02_rq1_return_predictability.py |
RQ1 panel + Fama–MacBeth (NW) | no |
03_rq2_rv_forecasting.py |
RQ2 HAR-RV vs IV surface, OOS, DM (30 tickers) | no |
04_rq3_correlation_divergence.py |
RQ3 correlation divergence (3 specifications, NW) | stage 1 only |
05_rq4_greeks_decay_magnets.py |
RQ4 Greeks decay + magnets | part A; part B has a parquet fallback |
06_rq5_ml_rv_forecast.py |
RQ5 ML vs VIX (code corrected, not re-run) | yes (entirely) |
07_descriptive_stats.py |
descriptive panels A–H | panel H only |
08_subperiod_regime.py |
subperiods, VIX regimes, rolling R² | no (VIX from the RQ3 parquet) |
09_portfolio_sorts.py |
sorts, double sorts, calendar-time strategies + costs | no |
10_robustness.py |
HC1 / NW(5,22) / double-clustered SE / controls / quantile / ticker R² | no |
11_granger_var.py |
Granger, VAR, orthogonalized IRF / FEVD | no |
12_make_figures.py |
supplementary figures | no |
13_extended_robustness.py |
winsorization / NW lags / Spearman ICs / deciles / leave-one-year-out / placebo | no |
14_make_paper_tables.py |
all LaTeX table fragments (paper/tables/) |
no |
Environment variables: WP7_RAW_DATA_DIR (raw DuckDB stores, default data/raw/), WP7_RESULTS_DIR (default results/), WP7_LEGACY_TARGETS=1 (reproduce the mis-aligned v1.1 targets, audit only).
🗃 Data
Options chains with Greeks (3.83 B records, 11,077 underlyings, 2010–2025) and 1-minute/5-minute OHLCV bars (11.5 B rows) from the HF Market Data lake, stored as DuckDB; the five derived parquets in data/processed/ (daily implied features, surface features, realized measures, correlation panels, SPX magnet data) are shipped and sufficient for every table of the paper except the frozen RQ4-A and RQ5 outputs. The raw stores are no longer on this machine (docs/AUDIT.md §6.1). The analysis universe is a fixed, ex-post list of liquid names (survivorship bias acknowledged in the paper).
🔬 Methodology
- Implied features (ATM IV 30/90 d, term slope, 25δ skew, smile-based skewness and kurtosis proxies, put–call ratios, net gamma) per ticker-day (
src/wp7/data_io.py,scripts/01). - Forward targets: strictly forward sums
Σ_{j=1..5} x_{t+j}(data_io.fix_forward_targets). - Return predictability: pooled OLS (HC1, NW, two-way clustering), Fama–MacBeth with NW, quintile/decile sorts with NW t-stats, calendar-time overlapping-cohort strategies with measured turnover and net-of-cost Sharpe ratios (
econometrics.py,portfolio.py). - RV forecasting: HAR-RV (RV_t, RV^(w), RV^(m)) vs IV models on a common sample; rolling OOS (500/250), MSE, QLIKE, R²_OOS, DM with NW variance (
forecasting.py). - Correlation divergence: equal-weight implied correlation (CBOE/Skintzi–Refenes analogue), 22-day realized correlation, LPM with NW(h).
- Greeks decay & price magnets: by DTE bucket; max-OI attraction test.
- Granger / VAR: per-ticker F-tests; bivariate VAR(5) with Cholesky-orthogonalized IRF/FEVD under both orderings.
- Robustness: subperiods, VIX regimes, controls, quantiles, ticker R², winsorization, NW lags, Spearman ICs, deciles, leave-one-year-out, placebo.
🖼 Figures & tables inventory
The paper is tables-only: 27 table fragments generated by scripts/14_make_paper_tables.py into paper/tables/ (plus the frozen raw-data inventory table typed in sections/data.tex). Supplementary figures from scripts/12_make_figures.py are in figures/.
⚠️ Limitations
- Fixed ex-post universe of 69 liquid names (survivorship bias); results do not speak to small caps, where the literature locates most option-implied return predictability.
- "Implied skewness/kurtosis" are smile-based proxies, not model-free risk-neutral moments; results are sensitive to tail trimming.
- Long-short strategies are equal-weighted, not factor-adjusted; costs are proportional assumptions.
- RQ4-A (Greeks by DTE) and RQ5 (ML) are frozen outputs of the original run (raw stores unavailable); only their 1-day components are reported.
- RQ3 stress indicator is defined from the VIX, which is also a regressor; realized correlation available from Feb 2016 only.
📝 Citation
@techreport{boucher2026options,
author = {Boucher, Simon-Pierre},
title = {The Options-Implied Information Content for Cross-Asset Return and Volatility Prediction: Evidence from 3.8 Billion Option Contracts},
institution = {Universit\'e du Qu\'ebec en Outaouais, D\'epartement des sciences administratives},
type = {Working Paper},
number = {7},
year = {2026},
month = {October},
note = {Version 1.2}
}👤 Author & contact
Simon-Pierre Boucher — Département des sciences administratives, Université du Québec en Outaouais (UQO), Gatineau, QC. 📧 simon-pierre.boucher@uqo.ca
🗒 Notes
- Revision 2026-10-17 (v1.2): full editorial-board review —
docs/REVIEW-REPORT.md(diagnosis, methodology verdicts, literature, maturity) anddocs/CHANGELOG-REVISION.md(every corrected number, scripts re-run, items not re-executable). Pre-revision state: git tagpre-revision-20261017,docs/main_pre-revision-20261017.pdf,results/_pre-revision-20261017/. Bibliography 45 → 97 entries, all verified on Crossref (docs/verify/crossref/). - Reproduction audit of v1.0 (2026-08-05):
docs/AUDIT.md,docs/CHANGES.md, evidence indocs/verify/(v1.1 logs and CSVs). - The May 2026 flat version is archived in
Recherche-UQO/_archives/versions-mai-2026/options-implied-cross-asset--_old/; the original project is in…/options-implied-cross-asset--hf-wp7-20260519/. - Harmonisation 2026-10-17: canonical title page, bibliography style
aer, pinnedrequirements.txt.