How it was tested: training data not stated, test data 2021-01→2026-09; out-of-sample design (how it was checked on data it never saw while being built): walk-forward yearly.
Horizon: 5, 21, 63 trading days. Coverage: top-300 US-region single-name equities by trailing 730d median dollar volume (close*volume); >= 2000 daily bars and last print within 30d of 2026-09-03; index ETFs excluded from ranking and supplied separately.
Stage: Deployed (shadow columns). Version v1.2; the hub record is as of 6 Nov 2026.
The verdict
FIT WITH CONDITIONS: EWMA(0.94) beats naive RV20 at h=5d and HAR-RV at h=21d/63d on QLIKE, DM p(BH) < 0.05 at every horizon. QLIKE and MSE disagree on the winner per horizon, so use the QLIKE pick for each. Validated, not deployed.
Development summary: Walk-forward vs naive RV20 (6 folds, 2021-26): EWMA(0.94) wins h=5d on QLIKE (0.567 vs 0.654); HAR-RV wins h=21d (0.306 vs 0.411) and h=63d (0.206 vs 0.325). DM p(BH) < 0.05 at every horizon. MSE ranks winners differently per horizon.
What the review allows
The scoreboard and DM results as an honest out-of-sample comparison on the stated 300-name US sample.
The winning model per horizon, IF the DM test is significant after BH adjustment, as a candidate to wire into VRP/wing pricing.
What the review rules out
Use the model columns for sizing or pricing before the D-008 promotion gate reads eligible or while the monitor reads breach, or read har_* on pool=false rows as validated (use fcst_21).
Extrapolate coefficients fit on this liquid US sample to illiquid names or other markets.
Treat a non-significant DM result as a win — report it as not-beating-naive, per horizon.
What it predicts
Item
Detail
Question it answers
Forecast annualised realised volatility over the next 5/21/63 trading days per symbol, to replace the trailing-20-day proxy several options numbers currently use as if it were a forecast.
Method
HAR-RV vs EWMA vs naive RV20
How it works
Four candidates raced chronologically: trailing-20d naive, RiskMetrics EWMA(0.94), pooled HAR-RV fit on log RV with Duan smearing; GARCH(1,1) skipped (arch not installed).
Per-symbol, per-horizon annualised RV forecast in vol points (same units as `iv`/`hv`), by model
Horizon
5, 21, 63 trading days
Coverage
top-300 US-region single-name equities by trailing 730d median dollar volume (close*volume); >= 2000 daily bars and last print within 30d of 2026-09-03; index ETFs excluded from ranking and supplied separately
Markets
US
Model type
Model · ML/Stat
Stage
Deployed (shadow columns)
Version
v1.2, first recorded 4 Sep 2026
Validation evidence
Training data: not stated. Test data: 2021-01→2026-09. Out-of-sample design: walk-forward yearly.
Headline results from the model's development record, as the hub publishes them.
Measure
Result
Compared with
Note
h=5d QLIKE
0.567 (EWMA 0.94)
naive RV20 0.654
DM p(BH) = 0.00
h=21d QLIKE
0.306 (HAR-RV)
naive RV20 0.411
DM p(BH) = 0.00
h=63d QLIKE
0.206 (HAR-RV)
naive RV20 0.325
DM p(BH) = 2.2e-08
Walk-forward folds
6 (2021-2026)
pooled, chronological
full scoreboard in development.performance
The hub's standard metrics for this model class (only those the record measured).
Metric
Sample / window
Compared with
Value
Range (CI)
Loss vs naive (headline)
WF 2021–26 · h=21 · HAR
1
0.74×
n/a
Diebold-Mariano p
BH-adj. · h=21 · 1,403 dates
0.05
0
n/a
What the validation review found
Is the design sound? Standard Corsi HAR-RV in log space with a proper Duan smearing correction; QLIKE/MSE on the variance scale is the literature-standard pairing; DM+Newey-West with horizon-implied bandwidth and HLN small-sample correction is textbook for overlapping h-step forecasts.
Data quality. Adjusted-close continuity spot-checked across two known splits (no defect found); target requires full h-bar forward windows (no imputation).
Backtest. Anchored walk-forward, one fold per calendar year 2021-2026, embargoed per-symbol against label leakage across the fold boundary — see scoreboard/DM tables in development.performance.
Limitations
Sample = 300 liquid US equities + 6 indices, not all 20 markets or the full options-eligible list; do not apply these coefficients to illiquid or non-US names without re-validating.
GARCH(1,1) skipped — arch not installed; see garch_status in the module for what re-enabling it needs.
"h trading days" are PRINTED bars, not calendar days — a halted or thinly-traded name's forward window can span more calendar time than usual with no explicit flag.
IV cross-check uses only the 621-symbol, ~2-year iv_history panel and is context, not a validated input or a fifth model.
QLIKE and MSE disagree on the ranking at every horizon (at h=5, EWMA wins QLIKE but HAR wins MSE; at h=21/63 it flips — HAR wins QLIKE but EWMA wins MSE): at the longer horizons this is traced to a small number of extreme-volatility outliers in erratic, event-driven names (e.g. BMNR during its 2022 and 2024 spikes, where realised vol pinned at the RV_CAP=300 vol-pt ceiling) that HAR's log-space mean reversion undershoots harder than EWMA's direct variance tracking does — trimming the worst 1% of HAR errors flips HAR to also winning on MSE at h=63, confirming tail behaviour on a handful of names rather than a broad mis-specification. Report both metrics; do not silently pick whichever flatters a given model.
Monitoring (designed 2026-10-16, see `monitoring`): rolling-60-session QLIKE of the forecasts as issued. The thresholds are judgement-set, not calibrated, and 60 sessions of overlapping 21-session windows hold only ~3 independent periods -- act on the persistence rules, not on one line.
Questions
Is the Realized-Volatility Forecaster validated?
Its validation review dated 4 Sep 2026 rated it FIT WITH CONDITIONS: EWMA(0.94) beats naive RV20 at h=5d and HAR-RV at h=21d/63d on QLIKE, DM p(BH) < 0.05 at every horizon. QLIKE and MSE disagree on the winner per horizon, so use the QLIKE pick for each. Validated, not deployed.
What does the Realized-Volatility Forecaster predict?
It answers: Forecast annualised realised volatility over the next 5/21/63 trading days per symbol, to replace the trailing-20-day proxy several options numbers currently use as if it were a forecast. Output: Per-symbol, per-horizon annualised RV forecast in vol points (same units as `iv`/`hv`), by model. Horizon: 5, 21, 63 trading days. Coverage: top-300 US-region single-name equities by trailing 730d median dollar volume (close*volume); >= 2000 daily bars and last print within 30d of 2026-09-03; index ETFs excluded from ranking and supplied separately.
How was the Realized-Volatility Forecaster tested?
Training data not stated, test data 2021-01→2026-09, out-of-sample design: walk-forward yearly. h=5d winner (QLIKE): ewma, 0.867x naive QLIKE, DM p(BH)=0; h=21d winner (QLIKE): har, 0.745x naive QLIKE, DM p(BH)=0; h=63d winner (QLIKE): har, 0.635x naive QLIKE, DM p(BH)=2.244e-08.
Out-of-sample: results measured on data the model never saw while it was built or tuned.
Walk-forward: the model is refit on data up to a date and tested on the period after it, then the window rolls forward, so every test is out of sample.
Confidence interval: the range that likely contains the true value (95% unless stated); [a, b] on this page is that range.
Shadow: scored and tracked on paper only; no list is published and nothing is acted on.
BH correction: the Benjamini–Hochberg adjustment for running many tests at once, so chance hits are not counted as findings.
QLIKE: a loss score for volatility forecasts that punishes under-forecasting; lower is better.
Diebold–Mariano test: a test of whether one forecast is significantly more accurate than another.
Implied volatility (IV): the volatility that an option's price implies.
VRP: volatility risk premium: implied volatility minus the volatility later realised.
HAR and EWMA: two volatility forecasters: HAR blends daily, weekly and monthly realised volatility; EWMA weights the most recent days most.
Calibration: whether predicted probabilities match how often things actually happen.