FINM7405 — Tutorial 2: Volatility and Financial Risk

Model Answers

Alexander Cameron

1. What is volatility?

(a) Volatility measures the dispersion of returns around their mean, typically estimated as
the standard deviation of a return series over a chosen lookback window. It quantifies
how much prices fluctuate over time.
(b) Because high dispersion implies greater uncertainty about future prices, volatility is a
natural proxy for risk. Its limitations include: it treats upside and downside symmetri-
cally; it assumes stationarity; and it ignores distributional shape (skewness, kurtosis).
√
(c) Daily volatility = 20%/ 252 ≈ 1.26% per day.

2. Volatility clustering

(a) Volatility clustering: large changes in returns tend to be followed by large changes,
and small changes by small changes. While returns themselves may be approximately
uncorrelated, their squares (or absolute values) exhibit positive autocorrelation.
(b) A constant-volatility model would severely underestimate risk during the turbulent
period (days 100–200), leading to inadequate capital buffers, underpriced options, and
VaR breaches.
(c) ARCH models conditional variance on past squared residuals. GARCH additionally
includes lagged conditional variance, allowing shocks to produce a more persistent ef-
fect. GARCH is therefore better suited to capturing the slow decay of volatility seen
2
empirically; the key structural difference is the β σt−1 term.

3. Historical vs. implied volatility

(a) Historical volatility is the sample standard deviation of log returns

√ over a chosen window
(e.g., 20 or 60 days), typically annualised by multiplying by 252. The choice of window
creates a bias–variance trade-off: shorter windows are more responsive but noisier.
(b) Implied volatility is the value of σ that, when input into an option pricing model (e.g.,
Black–Scholes), equates the model price to the observed market price. It is extracted
from traded options prices.
(c) Implied volatility reflects market expectations of future uncertainty. At the onset of
a crisis, investors demand insurance (put options), driving up option prices and hence
implied volatility, even before realised losses accumulate. This makes implied volatility
a leading rather than coincident risk indicator.

4. The VIX

(a) The VIX is the CBOE’s measure of 30-day implied volatility of the S&P 500, computed
model-free from a portfolio of out-of-the-money calls and puts across a range of strikes.
Unlike single-option implied volatility it does not depend on a specific pricing model.

1
 (b) It is called the “fear index” because spikes tend to coincide with market stress and
investor anxiety. The label can be misleading because elevated VIX can also reflect
genuine economic uncertainty rather than irrational fear, and because it measures ex-
pected volatility, not directional risk.
(c) Two mechanisms: (i) a large skew trade or options positioning could widen the put–call
spread without a directional move in the underlying; (ii) a scheduled event with binary
outcomes (e.g., central bank decision, geopolitical announcement) raises uncertainty
about future volatility even if current prices are unmoved.
5. Correlation during crises
(a) In the crisis panel the scatter tightens around the 45-degree line, indicating that the
two assets move almost one-for-one. The correlation coefficient rises toward 1.
(b) Diversification reduces variance only when correlations are below 1. If correlations ap-
proach 1 during a crisis, diversification benefits vanish. This is “correlation breakdown”:
the strategy fails precisely when risk is highest and protection most needed.
(c) A liquidity channel: when investors face margin calls or redemptions they sell liquid
assets indiscriminately, causing co-movement across otherwise unrelated assets.
(d) Not necessarily. Portfolio variance is σp2 = i j wi wj σi σj ρij . Even if ρij rises, vari-
P P
ance could remain stable if individual σi fall. In practice, crises raise both correlations
and individual volatilities simultaneously, so portfolio variance typically increases sub-
stantially.
6. Skewness and kurtosis
(a) Negative skewness means the left tail is heavier than the right: extreme losses are
more frequent and larger in magnitude than extreme gains of equal probability. The
distribution is asymmetric, with the mean below the median.
(b) Excess kurtosis (leptokurtosis) means more probability mass resides in the tails than
a normal distribution with the same variance. Extreme returns therefore occur more
often than a normal model would predict.
(c) Both features cause underestimation of tail risk. A normal-distribution VaR under-
weights large losses (fat left tail from negative skew) and underestimates the frequency
of extreme events (excess kurtosis). The true loss at a given confidence level will exceed
the model’s estimate — i.e., the normal VaR is systematically too optimistic.
7. Excel exercise
Indicative answers; exact values depend on the dataset provided.
√
(a) Annualised mean = r̄ × 252; annualised volatility = s × 252.
(b) TSLA or NVDA typically exhibit the highest annualised volatility, consistent with their
high-growth, speculative characteristics and sensitivity to earnings surprises.
(c) TSLA often shows pronounced negative skewness and excess kurtosis. Technology
stocks generally deviate more from normality than defensive stocks such as JNJ and
PG.
(d) AAPL–MSFT or GOOGL–MSFT are typically highly correlated (similar sector expo-
sure). JNJ–PG tend to be least correlated with technology names, reflecting defensive
versus growth characteristics.
(e) Portfolio variance = w⊤ Σ w. When correlations are low, off-diagonal covariance terms
are small and equal-weighting substantially reduces variance relative to the average indi-
vidual variance. Diversification provides the least benefit when all pairwise correlations
are high — as occurs during a crisis.

2
8. ARCH model
(a)
σ22 = 0.5 + 0.6 × (0.3)2 = 0.5 + 0.6 × 0.09 = 0.554
σ32 = 0.5 + 0.6 × (−3.0)2 = 0.5 + 0.6 × 9.0 = 5.9
(b) Covariance-stationarity requires α1 < 1. Here α1 = 0.6 < 1, so the condition is satisfied.
The unconditional variance is σ̄ 2 = α0 /(1 − α1 ) = 0.5/0.4 = 1.25.
(c)
σ42 = 0.5 + 0.6 × (0.5)2 = 0.5 + 0.15 = 0.65
Variance drops sharply from 5.9 to 0.65 in a single period, illustrating that ARCH(1)
has very limited persistence: once the large shock leaves the single-lag window, the
estimated variance returns almost immediately to its long-run level.
9. GARCH model
(a)
σt2 = 0.20 + 0.08 × (−3.0)2 + 0.85 × 1.06
= 0.20 + 0.72 + 0.901
= 1.821
(b) The unconditional (long-run) variance is obtained by setting σt2 = σt−12 = σ̄ 2 in the
GARCH equation:
ω 0.20 0.20
σ̄ 2 = = = ≈ 2.86
1−α−β 1 − 0.08 − 0.85 0.07
Covariance-stationarity requires α+β < 1. Here 0.08+0.85 = 0.93 < 1, so the condition
is satisfied.
(c) The β σt−12 term means today’s variance depends directly on yesterday’s variance, cre-
ating a memory that propagates across many periods. A large shock raises σt2 , which
2 , and so on, producing the gradual mean-reversion observed empirically.
flows into σt+1
ARCH(1) has no such channel: once the lagged squared shock is replaced, variance
reverts within a single period.
10. ARCH vs. GARCH
(a) ARCH(1) reverts to the long-run level within one or two periods because it depends
only on the most recent squared shock ε2t−1 . GARCH(1,1) decays gradually because
the lagged variance term β σt−12 carries forward accumulated information, producing
the slow geometric decline visible in the figure.
(b) For option pricing: persistent volatility implies that shocks affect the entire remaining
life of the option, raising premiums for longer-dated contracts. For VaR: ignoring
persistence leads to underestimation of multi-day risk during turbulent periods. For
portfolio rebalancing: knowing that volatility will remain elevated justifies reducing
position sizes or increasing hedges rather than waiting for calm to return.
(c) Better GARCH fit implies that variance shocks accumulate over time rather than being
immediately absorbed — i.e., the data exhibit genuine, slow-decaying volatility clus-
tering consistent with the GARCH data-generating process. An ARCH(1) model is
misspecified for such data because it cannot capture the multi-period persistence.
(d) The 2008–09 global financial crisis: the VIX remained above 40 for approximately five
months (October 2008 – February 2009), compared to a long-run average near 20.
Volatility normalised gradually as interbank lending markets stabilised following gov-
ernment and central bank interventions (TARP, Federal Reserve asset purchases).