02. Replicating SVIX and the Probability of a Crash#
In this guide
Option prices are a probability distribution. The slope of the call
price curve across strikes is the market’s risk-neutral CDF (Breeden and
Litzenberger). You write this step in rnd.risk_neutral_cdf.
SVIX is that distribution’s variance. SVIX² is the variance of the
market’s return under the risk-neutral distribution, not the real-world
one. It is a sum of out-of-the-money option prices with the same weight on
every strike. The VIX leans on the low strikes, so the gap between the VIX
and SVIX measures the left tail. You write this in svix.svix_squared.
The market’s crash probability is a price, not a forecast. The
risk-neutral probability of a crash overstates the real one. Martin’s log
investor shrinks it by the price of a put, and more risk aversion shrinks it
further. You write this in crash_prob.crash_probability.
Ian Martin’s What is the Expected Return on the Market?
(QJE 2017; the paper is in papers/) builds two numbers from S&P 500
option prices and publishes both on his website:
SVIX, a volatility index like the VIX with different weights on the
strikes, and the probability of a 20% decline in the S&P 500 over the
coming month. This guide rebuilds both for one session, from the E-mini
settlements of guide 01. There are three formulas, the three functions
Part B asks you to complete, and each gets a picture. The pictures are the
part to remember.
Sections
From option prices to a distribution.
SVIX: add up the option prices.
Three beliefs about a crash.
from pathlib import Path
import numpy as np
import pandas as pd
import crash_prob
import figures
import monitor
import pull_es_options
import pull_fred
import rnd
import smile
import svix
from settings import config
DATA_DIR = Path(config("DATA_DIR"))
pd.set_option("display.width", 120)
options = pull_es_options.load_es_options()
fred = pull_fred.load_fred()
# The newest session the monitor computed a row for (a session it had to
# skip, a rare chain with too few usable strikes, is not shown here either).
computed = set(monitor.load_history()["date"]) & set(options["date"])
latest = max(computed) if computed else options["date"].max()
session = options[options["date"] == latest]
rate = pull_fred.rate_on(fred, latest)
near, far = smile.bracketing_expiries(session)
month_expiry = smile.closest_expiry(session) # the expiration closest to 30 days
cdf_near = rnd.risk_neutral_cdf(smile.smile_for(options, latest, near, rate), rate)
cdf_far = rnd.risk_neutral_cdf(smile.smile_for(options, latest, far, rate), rate)
cdf_month = {near: cdf_near, far: cdf_far}[month_expiry]
print(f"Session {latest.date()}; expirations {pd.Timestamp(near).date()} and {pd.Timestamp(far).date()} bracket 30 days; "
f"one-month Treasury yield {100 * rate:.2f}%")
Session 2026-10-30; expirations 2026-11-30 and 2026-12-31 bracket 30 days; one-month Treasury yield 4.06%
Two kinds of probability#
There are two distributions of the market’s return a month from now. The real-world distribution says how often each outcome will happen. The risk-neutral distribution, written \(P^*\) and \(E^*\), is the set of probabilities under which every asset’s price equals its expected payoff, discounted at the risk-free rate. A call struck at \(k\) costs \(c(k) = E^*[(R_T - k)^+] / R_f\), where \(R_T\) is the gross return of the index (its level at expiration over today’s) and \(R_f\) the gross risk-free rate.
The two differ because a dollar is worth more in a crash than in a boom. A price weights each state by that worth as well as by its likelihood, so the risk-neutral distribution puts more weight on the bad states than the real world does. Option prices reveal the risk-neutral distribution directly. The risk-neutral CDF \(Q(k) = P^*(R_T \le k)\) is the price-implied chance that the index ends below \(k\) times today’s level. Sections 1 and 2 read the risk-neutral distribution; section 3 is the step back toward real-world probabilities.
1. From option prices to a distribution#
A call struck at \(k\) pays \((R_T - k)^+\). Raise the strike a little and the payoff falls by that little in every state where the call finishes in the money, and not at all elsewhere. So the slope of the call price in the strike is minus the risk-neutral probability of finishing above \(k\), discounted:
This is Breeden and Litzenberger’s result: the CDF is the slope of the call curve, and the density is its curvature. To take a slope we need a call price at every strike, not only at the listed ones, so the engine interpolates the smile of guide 01 onto a fine grid and prices a call at every point (the recipe of Martin and Shi’s Appendix D). The spot is normalized to 1, so a strike and a gross return are the same number \(k\).
The figure has the three steps, one panel each. Use the menu to change the expiration.
figures.fig_distribution(options, session, latest, rate).show()
Takeaway. The call curve falls and flattens as the strike rises, its slope is the CDF, and the CDF’s slope is the density. Short expirations give narrow distributions, long ones wide; the height of the CDF at 0.8 is the risk-neutral probability of a 20% drop.
p_star_20 = crash_prob.risk_neutral_crash_prob(cdf_month, 0.80)
p_star_10 = crash_prob.risk_neutral_crash_prob(cdf_month, 0.90)
print(f"Risk-neutral probability of a 20% drop by {pd.Timestamp(month_expiry).date()}: {100 * p_star_20:.2f}%; "
f"of a 10% drop: {100 * p_star_10:.2f}%")
Risk-neutral probability of a 20% drop by 2026-11-30: 0.80%; of a 10% drop: 3.54%
Stop and discuss
The density has a long, fat left tail. Is that what the market expects, or what it pays to avoid? From option prices alone, could you tell the two apart?
One cleaning step. Why interpolate at all? Take the same slope on the settlement prices directly, strike by strike, and check the one property every CDF has: it never falls.
for expiry in [near, month_expiry]:
spot = smile.smile_for(options, latest, expiry, rate)["spot_price"].iloc[0]
calls = session[(session["expiration"] == expiry) & (session["cp_flag"] == "C") & (session["settle"] > smile.MIN_SETTLE)].sort_values("strike")
k = calls["strike"].to_numpy() / spot
tau = calls["days_to_expiry"].iloc[0] / 365
q_from_prices = 1.0 + np.exp(rate * tau) * np.gradient(calls["settle"].to_numpy() / spot, k)
falls = np.diff(q_from_prices) < 0
print(f"{calls['days_to_expiry'].iloc[0]:>3} days: the 'CDF' from settlement prices falls at {falls.sum()} of {len(k)} strikes, "
f"by up to {100 * (-np.diff(q_from_prices)).max(initial=0):.1f} percentage points")
31 days: the 'CDF' from settlement prices falls at 29 of 152 strikes, by up to 0.5 percentage points
31 days: the 'CDF' from settlement prices falls at 29 of 152 strikes, by up to 0.5 percentage points
Takeaway. Settlements are marked to the tick, strike by strike, so their differences are noisy, and a CDF read off them falls in places, which no probability can do. The engine differentiates the smooth grid instead, then an isotonic regression (the closest non-decreasing curve) and a clip to \([0, 1]\) make the result a proper CDF. A vendor’s fitted surface hides this problem; working from the exchange’s prices, you see it and fix it.
2. SVIX: add up the option prices#
Martin defines SVIX² as the variance of the market’s return under the risk-neutral distribution of section 1, scaled by the risk-free rate and annualized (his equation 13):
The star matters: this is not a forecast of how volatile the market will be, but the variance that option prices imply. Martin’s paper uses it to put a lower bound on the expected return of the market, the question in its title.
Why is a variance a sum of option prices? Any smooth payoff can be built from a bond, a forward, and out-of-the-money options, with the option at strike \(K\) weighted by the payoff’s curvature there (Carr and Madan 1998). A variance is the price of a squared return, whose curvature is the same at every strike, so every strike gets the same weight (Martin’s equation 12, spot normalized to 1):
Puts below the forward, calls above it. The VIX prices a log payoff instead, whose curvature is \(1/K^2\), so it leans on the low strikes. The figure shows each index’s weights, as a share of its total:
figures.fig_svix_weights(cdf_month, latest).show()
svix2_near, svix2_far = svix.svix_squared(cdf_near), svix.svix_squared(cdf_far)
svix2_30 = svix.interpolate_variance(svix2_near, cdf_near.maturity_years, svix2_far, cdf_far.maturity_years)
# FRED publishes the VIX close a day later, so take the latest close on or before the session.
vix_obs = fred.dropna(subset=["vix"])
vix_obs = vix_obs[vix_obs["date"] <= latest].iloc[-1]
print(f"SVIX at {int(round(cdf_near.maturity_years * 365))} days: {svix.svix_index(svix2_near):.2f}; "
f"at {int(round(cdf_far.maturity_years * 365))} days: {svix.svix_index(svix2_far):.2f}")
print(f"Interpolated to 30 days, as the VIX is: SVIX {svix.svix_index(svix2_30):.2f}")
print(f"The VIX closed at {vix_obs['vix']:.2f} on {vix_obs['date'].date()}")
SVIX at 31 days: 17.99; at 62 days: 18.57
Interpolated to 30 days, as the VIX is: SVIX 17.95
The VIX closed at 19.42 on 2026-10-30
Takeaway. SVIX sits below the VIX because the VIX’s \(1/K^2\) weights lean on the deep puts, the most expensive options in volatility terms (the smirk of guide 01). The VIX-style number in the legend, computed from the same settlements, lands close to the VIX Cboe published that day: a check that the options are being read right. Two expirations rarely sit exactly 30 days out, so, like the VIX, the monitor interpolates SVIX² times \(\tau\) between the two that bracket 30 days.
Stop and discuss
If returns were lognormal, SVIX would sit slightly above the VIX at a one-month horizon. Martin found it below the VIX in every year of his sample, and it is below today. What does that gap say about how the options market prices a fall?
3. Three beliefs about a crash#
\(Q(0.8)\) overstates the chance of a crash, because option prices weight a crash by what a dollar is worth then. Martin’s way back (his Result 2) is to ask what an investor with log utility who holds the market must believe to be content holding it. Their marginal utility is \(1/R_T\), which turns the risk-neutral probability into
the risk-neutral probability, shrunk by the price of the put struck at the crash level \(\alpha\). Martin and Shi (2025) do the same for power utility with risk aversion \(\gamma\), weighting each state by \(R_T^\gamma\): \(P = E^*[R_T^\gamma 1\{R_T \le \alpha\}] / E^*[R_T^\gamma]\). At \(\gamma = 1\) it is Martin’s formula, and they calibrate \(\gamma = 2\).
The figure reads all three off the same distribution, for every size of drop. Drag the slider to change the risk aversion.
figures.fig_crash_beliefs(cdf_month, latest).show()
rows = []
for label, alpha in [("20% drop", 0.80), ("10% drop", 0.90)]:
rows.append({
"event": f"{label} by {pd.Timestamp(month_expiry).date()}",
"risk-neutral, P*": crash_prob.risk_neutral_crash_prob(cdf_month, alpha),
"log investor (Martin 2017)": crash_prob.crash_probability(cdf_month, alpha),
"power utility, gamma = 2 (Martin and Shi 2025)": crash_prob.crash_probability_power(cdf_month, alpha),
})
pd.DataFrame(rows).set_index("event").style.format("{:.2%}")
| risk-neutral, P* | log investor (Martin 2017) | power utility, gamma = 2 (Martin and Shi 2025) | |
|---|---|---|---|
| event | |||
| 20% drop by 2026-11-30 | 0.80% | 0.55% | 0.38% |
| 10% drop by 2026-11-30 | 3.54% | 2.92% | 2.43% |
Takeaway. The market charges for a crash at a probability well above what an investor who holds the market would assign to it, and the more risk-averse the investor, the larger the share of an option’s price that is fear rather than likelihood. At a risk aversion of 1 the power-utility line lands exactly on the log investor’s.
Stop and discuss
A risk committee asks you for “the probability of a 20% drop next month.” Which of these three numbers do you report, and what do you say about the other two?
The homework#
Part B asks for the three formulas of this guide, in src/rnd.py,
src/svix.py, and src/crash_prob.py. Each function’s docstring repeats
its formula, and uv run pytest src/test_svix.py src/test_rnd.py src/test_crash_prob.py
checks them offline: under a flat smile, returns are lognormal and every
answer has a closed form. With the three in place, uv run doit computes
this session, and every session in the pull, into the monitor’s history,
which is guide 03.
Notes on the discussion questions
What the market expects, or what it pays to avoid? From option prices alone you cannot tell. The risk-neutral density is the real-world density reweighted by how much a dollar is worth in each state, so a fat left tail can mean that crashes are likely, that investors badly want dollars in a crash, or both. Separating the two takes an assumption about preferences, which is what section 3 makes, or a long record of how often crashes actually happen.
SVIX below the VIX. The two indexes add up the same option prices with different weights. Under a lognormal distribution the weightings nearly agree, and SVIX comes out slightly above. The VIX’s \(1/K^2\) weights magnify the deep puts, so the VIX ends up above SVIX when those puts cost more than a lognormal distribution says they should, that is, when the market prices a left tail fatter than the normal curve’s. The gap measures how much the market pays for crash insurance beyond what volatility alone explains.
Which number for the risk committee? It depends on the question they are asking, and a good report says so. The risk-neutral probability is the price of insurance: what it costs to hedge a 20% drop, and an overstatement of how likely the drop is. The log investor’s and the power-utility investor’s numbers are forecasts, but each rests on an assumed utility (Martin and Shi calibrate risk aversion at 2). Report a forecast, name the assumption behind it, and give the risk-neutral number as the cost of protection.