FEDS Paper: Capturing Heterogeneity: Machine Learning Approaches to Implied Volatility Forecasting¶
Bibliographic record. Follow the original-source link for the publication.
| Field | Value |
|---|---|
| Primary domain | Volatility |
| Other domains | — |
| Methods | Financial Ml, Research Methods |
| Facets | Horizon Short Dated |
| Authors | Hyung Joo Kim, Dong Hwan Oh |
| Published | 2026-07-07 |
| Source | Federal Reserve FEDS |
| Identifiers | doi:10.17016/FEDS.2026.049 |
| URL | Open original source |
Editorial synthesis¶
Why it matters¶
The study tests whether replacing global SHAR coefficients with local moneyness–maturity dynamics improves conditional IV-surface forecasts. A shared first-stage surface fit focuses the comparison on the second-stage forecasting rule; forecast-error evidence is not evidence of utility, hedging P&L or investability. (full_text:S104, full_text:S105, full_text:S106, full_text:S142, full_text:S154, full_text:S231, full_text:S247, full_text:S419)
Main author claims¶
- The authors report: With the AHBS first stage, the authors report the lowest pooled OOS RMSE for Boosted-SHAR at 1, 5 and 22 days: 1.852, 3.250 and 5.124 IV percentage points. The 22-day SHAR value is 5.910; the displayed values imply a 13.3% relative reduction, consistent with the text's approximately 13% for this specific comparison. (
full_text:S94,full_text:S247,full_text:S257,full_text:S258,full_text:S259,full_text:S419) - The authors' DM/MCS results retain only Boosted-SHAR in the 95% MCS at all three horizons; the AHBS 22-day SHAR versus Boosted-SHAR DM statistic is 92.48. These are source-reported significance results, not inference independently verified by this review. (
full_text:S270,full_text:S272,full_text:S275,full_text:S422,full_text:S425) - The authors report larger gains in deep-OTM and short-maturity regions, with exceptions in some near-the-money years and longer horizons. Boosted-SHAR remains better in pooled SVI results, but the 22-day reduction from SHAR 5.614 to 5.133 is about 8.6%; in 2020 the 5-day RW value 6.105 beats Boosted-SHAR 6.171. Aggregate superiority is not superiority in every subgroup. (
full_text:S285,full_text:S306,full_text:S312,full_text:S322,full_text:S324,full_text:S426,full_text:S486)
Data, method, or discussion scope¶
The sample is daily OptionMetrics S&P 500 OTM calls/puts from 2011-01 through 2023-08, screened for 20-240-day maturity, moneyness S/K 0.8-1.6, IV, quotes and trading activity. Targets are 1-, 5- and 22-day IV conditional on future moneyness–maturity points, not unconditional fixed-contract forecasts. The authors describe expanding-window OOS evaluation in 2018-2023, re-estimating coefficients and partitions using only data through each origin, with a common AHBS or SVI first stage. Appendix B uses chronological training/validation within the first half, requires both origin and target to be in validation, excludes second-half OOS from tree-size selection, and shares selected horizon-specific leaf counts across Tree, Bagged and Boosted models. (full_text:S104, full_text:S105, full_text:S118, full_text:S119, full_text:S120, full_text:S145, full_text:S216, full_text:S218, full_text:S219, full_text:S231, full_text:S232, full_text:S233, full_text:S235, full_text:S459, full_text:S460, full_text:S463, full_text:S464, full_text:S465, full_text:S471, full_text:S480, full_text:S487)
Main limitations¶
The authors restrict splits to moneyness/maturity and leave other option characteristics and state variables for future work. Interpretation uses single Tree-SHAR because the winning Boosted-SHAR is difficult to summarize with one partition and parameter set. Reader-inferred limits: this filtered single-index-option sample and IV RMSE do not establish cross-asset validity, net-cost hedging superiority or executable investment outcomes. Regional/year exceptions and the different SVI gain limit universal-best language. Investor-clientele explanations of local coefficients are author conjectures, not identified causal mechanisms. DM/MCS loss aggregation and dependence-handling implementation are insufficiently specified. Table 5's Newey-West 10 lags apply only to coefficient t-statistics, not a documented DM specification; headline significance remains source-reported, not independently verified. (full_text:S149, full_text:S150, full_text:S151, full_text:S216, full_text:S218, full_text:S219, full_text:S247, full_text:S270, full_text:S285, full_text:S312, full_text:S322, full_text:S325, full_text:S331, full_text:S337, full_text:S362, full_text:S419, full_text:S422, full_text:S425, full_text:S426, full_text:S441, full_text:S486, full_text:S490)
Relationships¶
- None recorded.