Volatility risk premium study Evidence release 1.2
Aggregate evidence and calculation code are public without sign-in. Raw archived market-data inputs are excluded; full reproduction requires those exact files.
Author: Tomasz Dobrowolski. Publisher: FlashAlpha. Origin coverage: 2017-03-01 to 2026-10-01; forward input coverage through October 30, 2026. Symbol start dates vary.
How often did ATM implied vol exceed forward realized vol?
It did on 65% of 23,104 symbol-days; the median difference was +1.4 volatility points.
Sample
23,104 overlapping symbol-days across 11 instruments; origin dates 2017-03-01 to 2026-10-01; starts vary by symbol.
Method
30-calendar-day ATM IV minus the next 21 sessions' intraday-benchmark realized vol, annualized with 252.
Uncertainty
Pooled frequency and median are descriptive. Per-symbol mean intervals use session-aware HAC25 and HAC63; 11 of 11 volatility-mean HAC25 intervals include zero.
Limitation
The ATM-based proxy and overlapping windows are not independent trades, an option-strip premium, or investment returns.
Adding RV lowered observed forecast MSE on 3 of 11 symbols; 0 positive improvements had a pointwise 95% HAC25 interval entirely above zero.
Sample
20,126 forecast symbol-days; common origins across models within each symbol.
Method
Expanding OLS with an intercept, at least 252 eligible training outcomes, original-session embargo and outcomes ending before each forecast origin.
Uncertainty
Paired squared-error differences, pointwise normal-approximation HAC25 and HAC63 intervals; conditional on the forecast path, not formal nested-model superiority tests.
Limitation
This bounded historical comparison is exploratory; it does not establish that RV is universally useless or validate trading profitability.
Did the tested high-reading rule separate forward outcomes?
High-minus-low ATM variance differences were positive on 8 of 11 symbols. The unadjusted test was significant for 1 of 11, compared with 0 after Holm correction.
Sample
21,274 overlapping symbol-days; ranks use the preceding 252 sessions and require at least 189 available prior readings.
Method
Rank against the preceding 252 sessions with at least 189 prior readings; top third versus bottom third; centered conditional-mean influence scores, original-session Newey-West25, two-sided normal p-values and Holm across 11 symbols.
Uncertainty
The Holm result addresses this family of 11 tests, not a search across every strategy specification.
Limitation
This tests one percentile rule and a measurement proxy; it is not a z-score entry/sizing prescription or an options backtest.