04. Crash Bounds for Single Names (Martin and Shi 2025)#

Optional reading

This guide is optional, and it is not on the exam. Part D of HW 3 needs two of its sections: The data, which describes the query whose WHERE clause you write, and SVIX against Martin’s published series, the check your svix_squared must pass. The Martin-Shi bounds and the thirty years of crash probabilities are there for the curious.

In this guide

A stock’s crash probability needs more than its options. The real-world probability that one stock crashes depends on how it moves with the market in a crash, which no option price reveals, so Martin and Shi bound it instead of estimating it.

A vendor’s surface stops short of the crash. The OptionMetrics surface ends near the 10-delta put, so every one-month tail number built from it rests on an extrapolation convention (guide 01, section 7).

With no exchange benchmark, the check is the published series. No exchange publishes an SVIX for the S&P 500, so the pipeline is validated against the series Martin publishes.

The first three guides worked with the market as a whole. This one asks the question of Martin and Shi (2025), Forecasting Crashes with a Smile: what is the probability that one stock, Apple or AIG, falls 20% in the next month?

Options on the stock give its risk-neutral distribution, exactly as index options gave the market’s in guide 02. But a risk-neutral probability “cries wolf”: it prices crash states at their marginal utility, not their frequency. The correction that worked for the index, weighting each state by the investor’s return, now needs to know how the stock and the market move together in a crash, and no option price reveals that. Martin and Shi’s answer is to bound it. Two marginal distributions pin the joint one down only so far (the Fréchet-Hoeffding theorem), and the two extremes, the stock crashing exactly when the market does and crashing exactly when it does not, give a lower and an upper bound. The lower bound turns out to forecast crashes out of sample with nothing estimated on the stocks.

Sections

  • The data: the OptionMetrics volatility surface, and what it does not cover.

  • The paper’s Figure 1: the bounds for Apple and AIG, 1996 to 2022, and extended to the latest surface.

  • Thirty years of crash fear: the index’s crash probability from the surface, checked against the monthly series Martin (2017) publishes.

  • SVIX from the surface, checked against Martin’s published daily SVIX.

The code is Riley Haas and Marija Jovicic’s, from their Summer 2026 replication of the paper (forecasting_crashes_smile), used with their permission. The computation runs in src/single_name_bounds.py as the single_names task, and the SVIX check in src/calc_svix_vs_martin.py; this notebook loads their output.

from pathlib import Path

import numpy as np
import pandas as pd

import figures
from settings import config

OUTPUT_DIR = Path(config("OUTPUT_DIR"))
DATA_DIR = Path(config("DATA_DIR"))
MANUAL_DATA_DIR = Path(config("MANUAL_DATA_DIR"))

pd.set_option("display.width", 120)
pd.set_option("display.float_format", lambda x: f"{x:,.4f}")

HAVE_RESULTS = (OUTPUT_DIR / "single_name_bounds.parquet").exists() and (
    OUTPUT_DIR / "index_crash_probabilities.parquet"
).exists()
HAVE_SVIX = (OUTPUT_DIR / "svix_vs_martin.parquet").exists()
if not HAVE_RESULTS:
    print(
        "The single-name results are not built. They need the OptionMetrics pull\n"
        "from WRDS, which is not part of the public monitor: run `uv run doit single_names`\n"
        "with your WRDS username in .env, then rebuild. The rest of this notebook is\n"
        "skipped until then."
    )

The data: the OptionMetrics volatility surface#

Guides 01 and 02 used the exchange’s settlement prices, one per listed strike. Single-name options are not on the course’s Databento license, so this part uses what the academic literature uses: OptionMetrics’ IvyDB on WRDS, and specifically its standardized volatility surface, which the vendor fits every day for every security at fixed maturities (30, 60, 91, … days) and fixed deltas (10 to 90, puts and calls). Martin and Shi’s Appendix D filters on the surface’s dispersion, a goodness-of-fit measure, and keys moneyness off impl_strike, both of which exist only in the surface table, so the surface is the paper’s specification, not a shortcut.

Three WRDS tables, pulled by src/pull_optionmetrics.py for the S&P 500, Apple, and AIG at 30 and 91 days: the surface (vsurfd{year}, one table per year), daily closes for the spot (secprd{year}), and the zero curve (zerocd). The homework’s one blank in this part is the WHERE clause of the surface query.

if HAVE_RESULTS:
    surface = pd.read_parquet(DATA_DIR / "optionm_vol_surface.parquet")
    print(f"{len(surface):,} surface rows, {surface['date'].min().date()} to {surface['date'].max().date()}")
    one = surface[(surface["secid"] == 108105) & (surface["days"] == 30)]
    one = one[one["date"] == one["date"].max()].sort_values(["cp_flag", "delta"])
    display(one.head(8))
1,522,316 surface rows, 1996-01-04 to 2025-08-29
secid date days cp_flag delta impl_volatility impl_strike dispersion
1522265 108105 2025-08-29 30 C 10.0000 0.0954 6,713.8630 0.0080
1522266 108105 2025-08-29 30 C 15.0000 0.0965 6,671.0860 0.0051
1522267 108105 2025-08-29 30 C 20.0000 0.0983 6,638.2580 0.0042
1522268 108105 2025-08-29 30 C 25.0000 0.1005 6,610.0140 0.0043
1522269 108105 2025-08-29 30 C 30.0000 0.1030 6,584.0850 0.0046
1522270 108105 2025-08-29 30 C 35.0000 0.1057 6,559.1990 0.0049
1522271 108105 2025-08-29 30 C 40.0000 0.1086 6,534.5580 0.0053
1522272 108105 2025-08-29 30 C 45.0000 0.1119 6,509.5750 0.0060

A surface slice is 34 points: puts from delta -10 to -90 and calls from 10 to 90. Note where the strikes stop. The deepest put on the surface is the 10-delta put, and over the sample its strike sits around 92% of the spot at 30 days:

if HAVE_RESULTS:
    prices = pd.read_parquet(DATA_DIR / "optionm_security_prices.parquet")
    spx = surface[(surface["secid"] == 108105) & (surface["impl_strike"] > 0)].merge(
        prices[prices["secid"] == 108105][["date", "close"]], on="date"
    )
    spx["moneyness"] = spx["impl_strike"] / spx["close"]
    deepest = spx.groupby(["date", "days"])["moneyness"].min().unstack("days")
    coverage = pd.DataFrame(
        {
            "median deepest put (K/S)": deepest.median(),
            "share of days reaching K/S <= 0.80": (deepest <= 0.80).mean(),
            "share of days reaching K/S <= 0.90": (deepest <= 0.90).mean(),
        }
    )
    display(coverage)
median deepest put (K/S) share of days reaching K/S <= 0.80 share of days reaching K/S <= 0.90
days
30 0.9109 0.0098 0.3574
91 0.8512 0.1193 0.9647

A 20% drop is at moneyness 0.80, well beyond the deepest strike on the 30-day surface. Everything the engine does below that strike rests on an extrapolation convention: Appendix D holds the implied volatility flat beyond the last observed point. That is faithful to the paper, and it is the reason guide 01 made such a point of the exchange’s settlement prices, which exist for every listed strike down to a fraction of the spot; guide 01’s last figure shows what the convention costs. Keep it in mind when reading the one-month numbers.

The paper’s Figure 1: Apple and AIG#

For each month-end, each name, and each maturity, single_name_bounds.py builds the risk-neutral CDF of the name and of the index (the engine of guide 02, rnd.risk_neutral_cdf) and applies Result 3 of the paper. In words: read off the stock’s risk-neutral crash probability \(p^* = Q_i(q)\); find the market return \(q_l\) that is exactly as unlikely, \(Q_m(q_l) = p^*\); then the lower bound is the market’s own probability of falling to \(q_l\), corrected for risk aversion as in guide 02. The upper bound is the same construction in the market’s right tail. In symbols:

\[ \frac{E^*[R_m^\gamma\, 1\{R_m \le q_l\}]}{E^*[R_m^\gamma]} \;\le\; P[R_i \le q] \;\le\; \frac{E^*[R_m^\gamma\, 1\{R_m \ge q_u\}]}{E^*[R_m^\gamma]}, \qquad q_l = Q_m^{-1}(Q_i(q)),\; q_u = Q_m^{-1}(1 - Q_i(q)), \]

with \(\gamma = 2\), calibrated once on the market (the paper’s Appendix C). The lower bound is attained when the stock and the market crash together; the upper bound when they move against each other. The risk-neutral probability \(Q_i(q)\) always lies between them. Here is the paper’s Figure 1, the bounds on a 20% drop within a month, over the paper’s 1996 to 2022 sample:

if HAVE_RESULTS:
    bounds = pd.read_parquet(OUTPUT_DIR / "single_name_bounds.parquet")
    figures.fig_single_names(bounds).show()

Takeaway. The two names tell the paper’s story. AIG’s band explodes in the autumn of 2008, when its options priced the crash that was happening; Apple’s spikes are the dot-com bust and 2008, and its bounds are the quieter of the two after 2010. The width of the band is what the options cannot tell you: how the stock and the market move together in a crash. The buttons switch between the paper’s sample and the latest surface.

Stop and discuss

At the end of September 2008, the bounds put AIG’s chance of another 20% fall within a month between about one in three and one in two. What does a band that wide tell a risk manager, and what extra information would narrow it?

The latest month-end, for both names and both maturities:

if HAVE_RESULTS:
    latest = bounds[(bounds["date"] == bounds["date"].max()) & (bounds["threshold"] == 0.80)]
    display(latest.set_index(["ticker", "days_to_maturity"])[["bound_lower", "prob_riskneutral", "bound_upper"]])
bound_lower prob_riskneutral bound_upper
ticker days_to_maturity
AAPL 30 0.0029 0.0040 0.0047
AIG 30 0.0000 0.0000 0.0000
AAPL 91 0.0452 0.0718 0.0873
AIG 91 0.0145 0.0255 0.0325

Thirty years of crash fear, against Martin’s published series#

When the “name” is the market itself, the two returns are comonotonic and the lower bound holds with equality (the paper’s equation 7), so the index’s crash probability is a point, not a band. That is the number guide 02 computed from the exchange’s settlement prices for one session. Here it is from the vendor’s surface for every day since 1996, under the three beliefs of guide 02. Zoom in on 2008 and on March 2020.

if HAVE_RESULTS:
    index = pd.read_parquet(OUTPUT_DIR / "index_crash_probabilities.parquet")
    martin = pd.read_csv(MANUAL_DATA_DIR / "martin_2017" / "crashprob.csv")
    figures.fig_crash_history(index, martin).show()

Martin (2017) publishes exactly the green series, the log investor’s probability of a 20% drop, averaged by month, from January 1996 to January 2012 (data_manual/martin_2017/crashprob.csv, the black dots). His inputs are S&P 500 index options from the same vendor, so this is the check on the single-name half of the pipeline:

if HAVE_RESULTS:
    m1 = index[(index["days_to_maturity"] == 30) & (index["threshold"] == 0.80)]
    ours = m1.assign(year=m1["date"].dt.year, month=m1["date"].dt.month)
    monthly = ours.groupby(["year", "month"])["prob_log_utility"].mean().reset_index()
    merged = monthly.merge(martin[["year", "month", "m1"]], on=["year", "month"])
    merged["date"] = pd.to_datetime(merged[["year", "month"]].assign(day=1))
    gap = (merged["prob_log_utility"] - merged["m1"]).abs()
    print(f"Months in the overlap: {len(merged)}")
    print(f"Correlation with Martin's series: {np.corrcoef(merged['prob_log_utility'], merged['m1'])[0, 1]:.3f}")
    print(f"Median absolute gap: {100 * gap.median():.2f} percentage points; mean: {100 * gap.mean():.2f}")
    display(merged[merged["year"].isin([2008, 2011])].set_index("date")[["prob_log_utility", "m1"]].rename(
        columns={"prob_log_utility": "ours", "m1": "Martin"}))
Months in the overlap: 193
Correlation with Martin's series: 0.956
Median absolute gap: 0.27 percentage points; mean: 0.35
ours Martin
date
2008-01-01 0.0068 0.0083
2008-02-01 0.0052 0.0121
2008-03-01 0.0082 0.0106
2008-04-01 0.0017 0.0042
2008-05-01 0.0004 0.0024
2008-06-01 0.0022 0.0060
2008-07-01 0.0044 0.0067
2008-08-01 0.0008 0.0030
2008-09-01 0.0154 0.0160
2008-10-01 0.0882 0.0562
2008-11-01 0.1069 0.0724
2008-12-01 0.0788 0.0563
2011-01-01 0.0001 0.0005
2011-02-01 0.0004 0.0009
2011-03-01 0.0023 0.0054
2011-04-01 0.0003 0.0073
2011-05-01 0.0002 0.0024
2011-06-01 0.0009 0.0031
2011-07-01 0.0012 0.0068
2011-08-01 0.0289 0.0248
2011-09-01 0.0332 0.0356
2011-10-01 0.0212 0.0237
2011-11-01 0.0170 0.0198
2011-12-01 0.0065 0.0099

Takeaway. Crash fear spikes with every crisis, and ours tracks Martin’s: October 2008, the flash crash of May 2010, August 2011. Ours runs a little below Martin’s in calm months and above it in the worst ones, and the first of those is the surface again: with its strikes stopping near 92% of the spot, the flat extrapolation into the tail understates the price of deep puts in quiet markets, and the probability of a drop that starts beyond the last strike comes out small. Martin works from the raw option quotes, which reach further. The exchange’s settlement prices in guides 01 and 02 are the cure.

Stop and discuss

Ours runs below Martin’s in calm months. Using guide 01’s flat extrapolation figure, explain why a surface that stops near 92% would understate the tail most when markets are calm.

SVIX against Martin’s published series#

The same surface gives SVIX, with the integral of guide 02. Martin publishes his one-month SVIX², daily from January 1996 to January 2012 (data_manual/martin_2017/svix2.csv), computed from S&P 500 index options. No exchange publishes an SVIX for the S&P 500: CME’s CVOL indexes use the same construction, but only for options on its interest-rate, currency, energy, metals, and agricultural futures. So Martin’s series is the check on svix_squared. Both are shown in VIX units, 100 times the square root of SVIX²:

if HAVE_SVIX:
    sv = pd.read_parquet(OUTPUT_DIR / "svix_vs_martin.parquet")
    print(f"Days in the overlap: {len(sv):,} ({sv['date'].min().date()} to {sv['date'].max().date()})")
    print(f"Correlation with Martin's series: {np.corrcoef(sv['svix'], sv['svix_martin'])[0, 1]:.4f}")
    print(f"Median absolute gap: {sv['gap'].abs().median():.2f} vol points; mean gap: {sv['gap'].mean():+.2f}")
    figures.fig_svix_check(sv).show()
else:
    print("The SVIX check is not built: run `uv run doit single_names`.")
Days in the overlap: 4,038 (1996-01-04 to 2012-01-31)
Correlation with Martin's series: 0.9975
Median absolute gap: 0.29 vol points; mean gap: +0.14

Takeaway. The two lines are hard to tell apart. What gap there is comes from the same seam as before: the surface’s strikes stop near the 10-delta put, and holding the volatility flat beyond them puts a little more or a little less area under the out-of-the-money price curve than Martin’s raw quotes do. One of the homework’s tests runs your svix_squared on this surface at every month-end in the sample and requires it to stay within a fraction of a vol point of Martin’s series.

Where this leaves the homework#

Part D asks for one thing: the WHERE clause of the surface query in src/pull_optionmetrics.py, which selects the three securities, the two maturities, and the sample window from the vendor’s yearly tables. Two tests run it against a small in-memory database with the WRDS layout, as HW 1 tested its queries. Then uv run doit single_names pulls the surface and runs the chain that made every chart on this page, and the last test checks the index probability against Martin’s series as above. The same pull feeds Part B’s check of svix_squared against Martin’s SVIX.

Notes on the discussion questions

AIG in 2008. At the end of September 2008 the bounds were about 34% and 50%. A band that wide still says something firm. The lower bound is the case where AIG crashes only when the market does, the case that gives the smallest probability, and even then the chance of another 20% fall within a month was at least one in three. That floor is the number a risk manager needs. What the options cannot say is how AIG and the market move together in a crash; narrowing the band takes information about that joint behavior, such as a model of the dependence, its record in past crises, or the prices of contracts that pay on both at once.

Below Martin’s in calm months. In a calm market implied volatility is low, so the 10-delta put, the surface’s last strike, sits close to the money, and the whole region from there down to a 20% drop is filled by holding the volatility flat. The true smirk keeps rising there (guide 01’s last figure), so the extrapolated deep puts are too cheap and the crash probability too small. Martin worked from the raw option quotes, which reach deeper strikes. In a crisis volatility is high, the 10-delta put moves far out of the money, and less of the tail is extrapolated, so the gap closes.