volrec

Measuring the volatility risk premium with data I collect myself.

The question

Options are priced off an implied volatility — the market’s forecast of how much a stock will move. There’s a well-documented finding in the literature that implied volatility systematically exceeds the volatility that subsequently shows up, meaning options tend to be expensive. That gap is called the volatility risk premium.

I wanted to test whether it appears in data I gathered myself, rather than take it on authority.

Why collect rather than buy

Historical options data with a full implied-volatility surface is the expensive, gated part of options research. Vendors like OptionMetrics and ORATS price it for institutions. Live quotes, by contrast, are close to free.

So this trades time for money: record a daily snapshot, and after six months I own six months of proprietary history for a fixed universe. Started 4 September 2026.

Method

record.py runs every weekday at 8:30am Pacific via GitHub Actions. For each of 109 tickers it:

Fetches the underlying’s spot price
Pulls the option chain filtered to strikes within ±8% of spot and expiries 21–45 days out
Selects the call closest to 30 days to expiry, then closest to at-the-money
Records spot, strike, expiry, bid/ask, implied volatility, and greeks

One row per ticker per day, appended to data/iv_history.csv. Same-day reruns are deduplicated, and weekends and exchange holidays are skipped outright (checked against the market calendar), so no row is written for a day the market never traded.

At-the-money and ~30 days is deliberate: that’s where gamma is highest and where implied volatility is most reliably quoted, which keeps snapshots comparable as contracts roll.

The universe

109 tickers chosen for spread, not count — 109 correlated tech names would be one observation repeated 109 times. Started at 52; expanded to 109 on 5 September 2026, one day into collection, when the cost of uneven history was a single day. Candidates were screened live against the API and kept only if implied volatility, all five greeks and a two-sided quote came back with a spread under 60% of mid.

group	why
Index ETFs (SPY, DIA, QQQ, IWM, MDY)	Baseline, ascending volatility by market cap
Non-equity (GLD, SLV, TLT, IEF, USO, UNG, FXE, EEM, EFA)	Volatility driven by different forces — tests whether the premium is equity-specific
Sector ETFs (10)	Regime coverage without single-name earnings noise
Mega-cap tech (5)	The liquid, heavily-traded middle
High-vol growth (NVDA, TSLA, AMD, PLTR, COIN, MSTR, SMCI)	Top of the volatility range
Defensives (JNJ, PG, KO, PEP, WMT, MCD, VZ)	Anchors the bottom — without these everything clusters 20–45%
Financials, energy, industrials	Sector variety

First run confirmed the spread is real: implied vol ranged from 4.2% (FXE) and 5.1% (IEF) at the low end up to 72.1% (MSTR). The expansion widened it further, adding LQD at 4.9% and MARA at 82%.

The analysis (from ~November 2026)

For each snapshot, compare the implied volatility recorded that day against the volatility the underlying actually delivered over the following 30 days.

If realized consistently comes in below implied, the premium is present. The questions worth asking:

Does it hold across asset classes, or only in equities?
Is it larger for index ETFs than for single names? The literature says yes, and attributes the gap to a correlation risk premium.
Does it scale with the volatility level, or is it flat?
Is it large enough to survive bid-ask costs? (The bid/ask columns exist precisely so this can be answered rather than assumed.)

A null result is a real result. If the premium doesn’t clear costs in this sample, that is worth writing up as it stands.

One caution that shapes the whole analysis: sampling daily while looking forward 30 days means consecutive observations share almost the same realized path, and all 109 tickers on a given day share a market factor. A pooled t-test across every row would report significance on pure noise — simulated at 65–89% false positives against this exact design. The effective sample size is roughly the number of non-overlapping windows, not the number of rows. HANDOFF.md §4 carries the corrected method.

Companion tool

gamma-lab.html is a delta-hedging simulator built alongside this. It shows why the question matters: a long option, delta-hedged, profits when realized volatility exceeds implied. Version 2 adds transaction costs, stochastic implied vol, and jumps, and sweeps rebalance frequency to find where net profit peaks.

Data notes and limitations

Feed: Alpaca’s free indicative options feed for quotes, and the sip consolidated tape for stock closes. Options quotes are indicative, not exact NBBO. Fine for daily snapshots; worth stating in any write-up. Stock closes moved from iex to sip on 7 Sep 2026: iex is a single venue carrying a low single-digit share of volume, so its close is not the official closing print, and its free history is both shallower and ragged.
Snapshot timing: one reading per day at a fixed time, so intraday volatility is invisible. The schedule is fixed at 15:30 UTC, which is 11:30am ET while daylight saving is in effect but 10:30am ET once it ends on 1 November 2026. Snapshots from November onward therefore sit an hour earlier in the session than those before it — worth controlling for, since implied vol is not flat across the trading day.
Expiry drift: MDY and FXE lack weekly options and fall back to ~42-day expiries. The dte column records this so it can be controlled for.
Wide quotes on FXE and XLU: on day one FXE quoted 0.63/1.27 and XLU 0.45/0.79 — spreads of 67% and 55% of the mid respectively. Wide enough that the mid is unreliable for both. Under review.
Survivorship: the universe is fixed as of Sept 2026 and doesn’t adjust for future delistings or index changes.
Vendor-computed IV: implied volatility and greeks are calculated by Alpaca from its free indicative feed rather than from exact NBBO quotes, using a documented Black–Scholes solver but an undocumented risk-free rate and dividend treatment. Levels should be read as approximate.
At-the-money only, and calls only: implied vol varies by strike, and that shape carries information this dataset doesn’t capture. An ATM Black–Scholes IV also approximates the volatility swap rate rather than the model-free variance swap rate that VIX-style measures use — which understates the premium, and understates it more for index ETFs than for single names.

Two samples, one codebase

The binding constraint here is not data, it is independent observations. By late October there will be ~40 trading days but only about six independent episodes of the market factor, because most days move together. analyze.py --simulate demonstrates the consequence: under a true null with zero premium by construction, a pooled test across tickers rejects 63.8% of the time, while a non-overlapping test sits near its nominal 5%.

No additional column fixes that. Only additional observations do. So there are two samples running identical code:

Sample A (samples/long/) uses free Cboe volatility indices as the implied-vol measure, 2016-2026, ~127 non-overlapping episodes per pair. This is where the method is validated. Tested 7 Sep 2026: the premium is significantly positive on 9 of 11 underlyings, VIX/SPY at +3.52 volatility points, t = 5.14.
Sample B (samples/panel/) is data/iv_history.csv, N ≈ 6. The same frozen code runs on it and the result is reported with its sample size stated.

Sample A is a different estimand and is never presented as a result about the collected panel: Cboe’s indices are variance-swap-style and integrate the whole strike surface, while this project records at-the-money implied vol. The measured gap across four matched pairs on 4 Sep 2026 was 3.61 volatility points, and that gap is the skew premium.

The strike surface, and what free data actually costs

Recording at-the-money implied volatility answers how large the premium is. It cannot answer where in the strike surface it sits, which needs the whole smile and, to say anything about liquidity, the volume at each strike.

Since 14 September 2026 a second recorder captures that: eight underlyings, two expiries, forty strikes spanning +/-30% of spot, with bid, ask, implied volatility, all five greeks, volume and open interest per contract. It costs no additional API calls, because the original recorder was already fetching those strikes and discarding all but one.

The measurement question this makes possible. Cboe publishes the authoritative model-free volatility index for five of those underlyings (SPY/VIX, QQQ/VXN, IWM/RVX, GLD/GVZ, USO/OVX). So a model-free estimate built from a free, indicative, non-OPRA feed can be checked against the authoritative number, same underlying, same day. Nobody with a research budget measures what free data costs, because they buy OptionMetrics instead.

Calibrated 11 September 2026, implementing Cboe’s own methodology:

strike band	mean absolute gap vs Cboe	worst
+/-10% x 20 strikes	3.97 vol pts	-12.37
+/-20% x 30 strikes	1.31	-4.39
+/-30% x 40 strikes	0.59	-1.84

At the widest band SPY comes in at 15.83 against a published VIX of 15.84.

The gap was never data quality. It was strike truncation, and it bites hardest where volatility is highest: a fixed +/-10% band spans about 2.2 standard deviations on a 16-vol name but only 0.6 on 59-vol crude oil, which is why USO was twelve points light. Any study using a fixed percentage strike band is therefore most biased on exactly the high-volatility names most likely to be interesting.

Measuring the premium across strikes, without cancelling it

The obvious approach does not work, and the arithmetic is worth stating. If the premium at strike K is IV(K)^2 - RV and RV is one number shared by every strike that day, then differencing two strikes cancels the realised term exactly and leaves the shape of the implied volatility surface, which is the volatility smile, which Bollen & Whaley published in 2004.

So the outcome here is strike-specific: the delta-hedged gain of Bakshi & Kapadia (2003). Buy the option, short delta shares, rebalance daily, see what is left. The hedging path depends on the contract’s own gamma, so it does not cancel. A negative hedged gain means the buyer paid for more movement than arrived, which is a positive variance risk premium.

Repo structure

record.py                  the daily ATM recorder, 109 tickers
surface.py                 the daily strike-surface recorder, 8 underlyings
analyze.py                 shared estimators, used by both samples
modelfree.py               Cboe's variance methodology, and the gap against it
hedged.py                  per-contract delta-hedged P&L
data/iv_history.csv        the ATM panel, append-only, never written elsewhere
data/iv_history.pre-17col.csv   one-time migration backup, immutable
data/surface.csv           the strike surface
hypotheses/                pre-registered, dated, committed BEFORE the test
samples/long/              Sample A
samples/panel/             Sample B
tools/pressure_test.py     49 read-only integrity checks
tools/test_hedged.py       18 hand-computed cases for the hedging math

The rule that holds the rest together: nothing becomes a model variable without a hypothesis in hypotheses/ dated before the join that tests it.

Running it

pip install requests
export ALPACA_KEY="..."
export ALPACA_SECRET="..."
python record.py

python record.py --probe SPY dumps a raw API response for debugging.

Automated via .github/workflows/record.yml; keys live in repository secrets.