3. Measuring and Monitoring Volatility
3.1 Deviations from Normality
3.1.1 Asset Return Distribution
Asset return distributions tend to deviate from the normal distribution.
Fat-tailed: a distribution has more probability of observations in its tails relative to the normal distribution.
Non-symmetrical(mainly left skewed): refers to a higher probability of extreme losses and frequent small gains.
Unstable: varying market conditions result in unstable parameter values such as volatility and mean. This situation can lead to either fat tail or skewed.
3.1.2 Unconditional and conditional distribution
Unconditional distribution: the same distribution exists at any given point of time with constant mean and standard deviation for the asset return regardless of market condition.
Conditional distribution: different distribution exists with conditional mean, conditional volatility and other conditional parameters due to changing market condition.
- Fat tails: mixed normal distributions with same mean and different standard deviations. (Time varying volatilities).
- Skewness(Asymmetrical distributions): mixed normal distribution with both different means and standard deviations.
3.1.3 Slow changes vs. Regime switching
It is often reasonable to suppose that volatility changes slowly
Sometimes volatility can suddenly jump because of unforeseen events or government actions. When markets calm down, they could suddenly rebound. This phenomenon is called regime switching.
3.2 Volatility estimation approach
3.2.1 Equally Weighted Standard Deviation
Denote the return on day
i
i
i by
r
i
r_i
ri. Assuming the asset provides no income:
r
i
=
(
S
i
−
S
i
−
1
)
/
S
i
−
1
r_i=(S_i-S_{i-1})/S_{i-1}
ri=(Si−Si−1)/Si−1
The usual formula for calculating standard deviations from sample data would give the volatility estimated for day n n n from the return on the m m m previous days as follow, and the r ‾ \overline{r} r is the average return over the m m m previous days.
σ n 2 = 1 m − 1 ∑ i = 1 m ( r n − i − r ‾ ) 2 \sigma_n^2=\sqrt{\frac{1}{m-1}\sum^m_{i=1}(r_{n-i}-\overline{r})^2} σn2=m−11i=1∑m(rn−i−r)2
In risk management, we usually simplify the formula in two ways: replace m − 1 m-1 m−1 with m m m (reasonable when m m m is big) and make r ‾ = 0 \overline{r}=0 r=0 (reasonable as the mean of daily return approximates 0 0 0). Each day, the forecast is updated by adding the most recent day and dropping the furthest day.
σ n 2 = 1 m ∑ i = 1 m ( r n − i ) 2 \sigma_n^2=\frac{1}{m}\sum^m_{i=1}(r_{n-i})^2 σn2=m1i=1∑m(rn−i)2
Estimation window m m m of different lengths will impact the forecasts of standard deviation.
- If m m m is too large, the estimate of the current volatility will not be reliable because it will use data too long age.
- If m m m is too small, and the standard error of the estimate will be too big. Meanwhile, the adding and deleting of the extreme values due to rolling window will result in sharp changes of the estimation.
3.2.2 Exponential Smoothing(EWMA)
Exponentially weighted moving average(EWMA) gives more weight to more recent information and places exponentially declining weights on distant information(rather than using the equal weight).
ω + ω λ + ω λ 2 + ω λ 3 + ⋯ + ω λ k − 1 = 100 % → ω = 1 − λ \omega+\omega\lambda+\omega\lambda^2+\omega\lambda^3+\dots+\omega\lambda^{k-1}=100\% \to \omega=1-\lambda ω+ωλ+ωλ2+ωλ3+⋯+ωλk−1=100%→ω=1−λ
- λ ( 0 < λ < 1 ) \lambda(0<\lambda<1) λ(0<λ<1) is introduced as a smoothing parameter.
- k k k: days of data.
- ω \omega ω: the weight applied to the most recent return.
{ σ n 2 = ω λ 0 r n − 1 2 + ω λ 1 r n − 2 2 + ω λ 2 r n − 3 2 + … λ σ n − 1 2 = ω λ 1 r n − 2 2 + ω λ 2 r n − 3 2 + ω λ 3 r n − 4 2 + … → σ n 2 = ( 1 − λ ) r n − 1 2 + λ σ n − 1 2 \begin{cases} \sigma_n^2=\omega\lambda^0 r^2_{n-1}+\omega\lambda^1 r^2_{n-2}+\omega\lambda^2 r^2_{n-3}+\dots \\ \lambda\sigma_{n-1}^2=\omega\lambda^1 r^2_{n-2}+\omega\lambda^2 r^2_{n-3}+\omega\lambda^3 r^2_{n-4}+\dots \end{cases} \to \sigma^2_n=(1-\lambda)r_{n-1}^2+\lambda\sigma^2_{n-1} {σn2=ωλ0rn−12+ωλ1rn−22+ωλ2rn−32+…λσn−12=ωλ1rn−22+ωλ2rn−32+ωλ3rn−42+…→σn2=(1−λ)rn−12+λσn−12
- Relatively little data needs to be stored
- High λ \lambda λ respond relatively slowly to new information provided by the daily percentage changes.
- RiskMetric found that λ = 0.94 \lambda=0.94 λ=0.94 to be a good choice.
Suppose today’s volatility estimate is 2 % 2\% 2% per day and we observe a return of − 1 % -1\% −1% today. Assuming λ = 0.94 \lambda=0.94 λ=0.94, what is the new estimate of volatility based on the exponential smoothing approach?
( 1 − 0.94 ) × ( 1 % ) 2 + 0.94 × ( 2 % ) 2 = 0.000382 (1-0.94)\times(1\%)^2+0.94\times(2\%)^2=0.000382 (1−0.94)×(1%)2+0.94×(2%)2=0.000382
The new estimate of volatility 0.000382 = 1.95 % \sqrt{0.000382}=1.95\% 0.000382=1.95%
3.2.3 GARCH
The GARCH model can be regarded as an extension of EWMA.
In GARCH(1,1), we not only give some weight to the most recent variance rate and the latest squared return, but also give some weight to a long run average variance rate.
σ n 2 = γ V L + α r n − 1 2 + β σ n − 1 2 \sigma^2_n=\gamma V_{L}+\alpha r^2_{n-1}+\beta\sigma^2_{n-1} σn2=γVL+αrn−12+βσn−12
- Since weights must sum to 1 1 1, so γ + α + β = 1 \gamma+\alpha +\beta=1 γ+α+β=1
- The EWMA model is a particular case of the GARCH(1,1) model where α = 1 − λ \alpha=1-\lambda α=1−λ, β = λ \beta=\lambda β=λ
Setting ω = γ V L \omega=\gamma V_{L} ω=γVL, the GARCH(1,1) model is
σ n 2 = ω + α r n − 1 2 + β σ n − 1 2 \sigma^2_n=\omega+\alpha r^2_{n-1}+\beta\sigma^2_{n-1} σn2=ω+αrn−12+βσn−12
V L = ω / ( 1 − α − β ) , α + β < 1 V_{L}=\omega/(1-\alpha-\beta) ,\; \alpha+\beta<1 VL=ω/(1−α−β),α+β<1
Suppose σ n 2 = 0.000002 + 0.13 r n − 1 2 + 0.86 σ n − 1 2 \sigma_n^2=0.000002+0.13r_{n-1}^2+0.86\sigma_{n-1}^2 σn2=0.000002+0.13rn−12+0.86σn−12 and the current estimate volatility is 1.6 % 1.6\% 1.6% per day and the most recent return is 1 % 1\% 1%. What is the long-term volatility and what is the new estimate of volatility?
The long run volatility is 1.41 % 1.41\% 1.41% per day and the new estimate of volatility is 1.53 % 1.53\% 1.53%.
V L = 0.000002 / ( 1 − 0.13 − 0.86 ) = 0.0002 V_{L}=0.000002/(1-0.13-0.86)=0.0002 VL=0.000002/(1−0.13−0.86)=0.0002
σ L = 0.0002 = 1.41 % \sigma_L=\sqrt{0.0002}=1.41\% σL=0.0002=1.41%
σ n 2 = 0.000002 + 0.13 × ( 1 % ) 2 + 0.86 × ( 1.6 % ) 2 = 0.00023336 \sigma_n^2=0.000002+0.13\times(1\%)^2+0.86\times(1.6\%)^2=0.00023336 σn2=0.000002+0.13×(1%)2+0.86×(1.6%)2=0.00023336
σ n = 0.00023336 = 1.53 % \sigma_n=\sqrt{0.00023336}=1.53\% σn=0.00023336=1.53%
Mean reversion: the
V
L
V_L
VL term provides a ''pull" toward the long-run average mean. We can prove that:
E
(
σ
n
+
t
2
)
=
V
L
+
(
α
+
β
)
t
(
σ
n
2
−
V
L
)
E(\sigma^2_{n+t})=V_L+(\alpha+\beta)^t(\sigma^2_n-V_L)
E(σn+t2)=VL+(α+β)t(σn2−VL)
When t t t is large, the E ( σ n + t 2 ) E(\sigma^2_{n+t}) E(σn+t2) approximates to V L V_L VL
α
+
β
\alpha+\beta
α+β is called the persistence level, which defines the speed at which shocks to the variance revert to their long-run value. The bigger the persistence level, the slower the variance revert to their long-run value.
Suppose that the current estimated daily volatility is 3 % 3\% 3% and the lone-run estimated average daily volatility is 2 % 2\% 2%. What are the volatility estimates in 10 10 10 days and 100 100 100 days in a GARCH(1,1) model where α = 0.04 \alpha=0.04 α=0.04, β = 0.94 \beta=0.94 β=0.94?
E ( σ 10-days ) = 0.0 2 2 + ( 0.04 + 0.98 ) 10 ( 0.0 3 2 − 0.0 2 2 ) = 2.84 % E(\sigma_{\text{10-days}})=\sqrt{0.02^2+(0.04+0.98)^{10}(0.03^2-0.02^2)}=2.84\% E(σ10-days)=0.022+(0.04+0.98)10(0.032−0.022)=2.84%
E ( σ 100-days ) = 0.0 2 2 + ( 0.04 + 0.98 ) 100 ( 0.0 3 2 − 0.0 2 2 ) = 2.16 % E(\sigma_{\text{100-days}})=\sqrt{0.02^2+(0.04+0.98)^{100}(0.03^2-0.02^2)}=2.16\% E(σ100-days)=0.022+(0.04+0.98)100(0.032−0.022)=2.16%
Long horizon volatility: given short term volatility, square root rule(平方根法则) is applied to forecast for long term volatility.
σ J-periods = σ 1-period × J \sigma_{\text{J-periods}}=\sigma_{\text{1-period}}\times\sqrt{J} σJ-periods=σ1-period×J
Assuming the volatility is mean reverting.
- If today’s volatility is higher than the long-term mean, using square root may overstate the volatility / VaR \text{VaR} VaR.
- If today’s volatility is lower than the long-term mean, using square root rule may understate the volatility / VaR \text{VaR} VaR.
3.2.4 Exponential Smooth(EWMA)-correlation
EWMA can also be used to update correlation. Assue the return for asset X X X and Y Y Y are x n x_n xn and y n y_n yn, the updated covariance between the asset X X X and Y Y Y is
c o v n = λ c o v n − 1 + ( 1 − λ ) x n − 1 × y n − 1 cov_n=\lambda cov_{n-1}+(1-\lambda)x_{n-1}\times y_{n-1} covn=λcovn−1+(1−λ)xn−1×yn−1
If we use EWMA to estimate the standard deviation of the rate of return, we can estimate the correlation coefficient.
ρ n = c o v n σ x , n × σ y , n \rho_n=\frac{cov_n}{\sigma_{x,n}\times \sigma_{y,n}} ρn=σx,n×σy,ncovn
Suppose the returns for X X X and Y Y Y on day n n n are both 2 % 2\% 2% and λ = 0.94 \lambda=0.94 λ=0.94 and the current covariance of X X X and Y Y Y is 0.00004 0.00004 0.00004. What is the updated coefficient of correlation if the updated volatilities of X X X and Y Y Y are 1.086 % 1.086\% 1.086% and 2 % 2\% 2%?
The updated covariance between X X X and Y Y Y is
c o v n = 0.94 × 0.00004 + 0.06 × 0.02 × 0.02 = 0.0000616 cov_n=0.94\times0.00004+0.06\times0.02\times0.02=0.0000616 covn=0.94×0.00004+0.06×0.02×0.02=0.0000616
The updated coefficient of correlation is then:
ρ n = 0.0000616 / ( 0.01086 × 0.02 ) = 0.28 \rho_n=0.0000616/(0.01086\times0.02)=0.28 ρn=0.0000616/(0.01086×0.02)=0.28
3.2.5 Implied Volatility
Implied volatility approach: obtains an implied volatility based on current market data using the Black-Scholes-Merton option pricing model rather than historical data.
- Implied volatilities are forward looking, whereas the volatilities calculated from historical data (e. G., EWMA, GARCH) are backward looking.
- Evidence indicates that implied volatilities give better estimates of realized volatility than historical
- Reliable implied volatilities for an asset are sometimes not available as options are not traded actively.
- VIX is an index of the implied volatilities of 30-day options on the S&P 500.
3.3 Historical simulation with different weighting schemes
3.3.1 Historical Simulation with declining weights
Scenario | Loss(USD millions) | Days ago | Weight | Cumulative Weight |
---|---|---|---|---|
490 | 7.8 | 11 | 0.20% | 0.20% |
492 | 6.5 | 9 | 0.20% | 0.40% |
2 | 4.6 | 499 | 0.20% | 0.60% |
23 | 4.3 | 478 | 0.20% | 0.80% |
48 | 3.9 | 453 | 0.20% | 1.00% |
367 | 3.7 | 134 | 0.20% | 1.20% |
235 | 3.5 | 266 | 0.20% | 1.40% |
… | … | … | … | … |
E S daily ( 1 % ) = ( 7.8 + 6.5 + 4.6 + 4.3 ) / 4 = 5.8 million ES_{\text{daily}}(1\%)=(7.8+6.5+4.6+4.3)/4=5.8\;\text{million} ESdaily(1%)=(7.8+6.5+4.6+4.3)/4=5.8million
V a R daily ( 1 % ) = 3.9 million VaR_{\text{daily}}(1\%)=3.9\;\text{million} VaRdaily(1%)=3.9million
Using exponentially declining weights can be useful when calculating VaR \text{VaR} VaR or E S ES ES, as scenarios calculated from recent data are more relevant than scenarios calculated using data from many days ago.
N N N represents the number of scenarios and m represents the scenario from m days ago. (m ranges from 1 1 1 to N N N)
ω 1 = ( 1 − λ ) × λ 0 , ω 2 = ( 1 − λ ) × λ 1 , ω N = ( 1 − λ ) × λ N − 1 \omega_1=(1-\lambda)\times\lambda^0,\; \omega_2=(1-\lambda)\times\lambda^1,\; \omega_N=(1-\lambda)\times\lambda^{N-1} ω1=(1−λ)×λ0,ω2=(1−λ)×λ1,ωN=(1−λ)×λN−1
Scenario | Loss(USD millions) | Days ago | Weight | Cumulative Weight |
---|---|---|---|---|
490 | 7.8 | 11 | 0.90% | 0.90% |
492 | 6.5 | 9 | 0.92% | 1.83% |
2 | 4.6 | 499 | 0.01% | 1.83% |
23 | 4.3 | 478 | 0.01% | 1.84% |
48 | 3.9 | 453 | 0.01% | 1.85% |
367 | 3.7 | 134 | 0.26% | 2.12% |
235 | 3.5 | 266 | 0.07% | 2.19% |
… | … | … | … | … |
Introduce λ = 0.99 \lambda=0.99 λ=0.99 to assign weights to scenarios
Weight 490 = ( 1 − 0.99 ) × 0.9 9 10 = 0.009 \text{Weight}_{490}=(1-0.99)\times0.99^{10}=0.009 Weight490=(1−0.99)×0.9910=0.009
ES daily ( 1 % ) = 7.8 × 0.9 + 6.5 × 0.1 = 7.67 million \text{ES}_{\text{daily}}(1\%)=7.8\times0.9+6.5\times0.1=7.67\;\text{million} ESdaily(1%)=7.8×0.9+6.5×0.1=7.67million
VaR daily ( 1 % ) = 6.5 million \text{VaR}_{\text{daily}}(1\%)=6.5\;\text{million} VaRdaily(1%)=6.5million
3.3.2 Multivariate Density Estimation (MDE)
MDE is another approach used to determine weights of scenario when calculating VaR \text{VaR} VaR and E S ES ES. With MDE, an analysis is carried out to determine which periods in the past are most similar to the current period.
Macro economic factors like GDP growth rate, interest rate level etc.are often used to determine how similar one period is to another.