10. Variance and Standard Deviation
• Variance = σ2 = E[(X – E(X))2], properties:
1. If c is constant: Var(c) = 0 and VaR(cX)=c2 x Var(X)
2. If c and a are constants: Var(aX+c ) = a2 x Var(X)
3. If c and a are constants: E(cX+a) = cE(X)+a
4. If X and Y are independent random variables:
Var(X+Y) = Var(X‐Y) = Var(X) + Var(Y)
• Standard Deviation = Square root of Variance
= σ = {E[(X – E(X))2]}1/2
12. Covariance
• Covariance: A measure of how to variables move together.
Cov(X,Y) = E[(X‐E(X))(Y‐E(Y))] = E(XY)‐E(X)E(Y)
• Interpretation:
– Values range from negative to positive infinity.
– Positive (negative) covariance means when one variable has
been above its mean the other variable has been above (below)
its mean.
– Units of covariance are difficult to interpret which is why we
more commonly use correlation (next slide)
• Properties:
– If X and Y are independent then Cov(X,Y) = 0
– Covariance of a variable X with itself is Var(X)
– If X and Y are NOT independent:
• Var(X+Y) = Var(X) + Var(Y) + 2(Cov(X,Y)
• Var(X‐Y) = Var(X) + Var(Y) ‐ 2(Cov(X,Y)
43. • Is the variance of a banks trading book returns = 0.16%?
– H0: σ2 = 0.16% vs HA: σ2 ≠ 0.16% || Type of test: Two tailed
– Facts: n (months) = 24, s2 (sample average standard deviation) = 0.1444%
• Steps:
1. Select test statistic (Chi‐square)
2. Specify significance level (5%)
3. Determine Critical Values
4. Calculate Test Statistic (below)
5. Fail to Reject H0: μ ≤ 0.016
= (23 x .14%) / .16% = 20.75
Chi‐Square test of population variance
df 0.975 0.025
22 10.982 36.781
23 11.689 38.076
24 12.401 39.364
Chi Square Table
Chi Square PDF Distribution
11.689 ‐ Critical Values – 38.076
X2 = 20.75
Intuition: X2 is close to n
(24) because hypothesized
σ2 is close to observed s2
48. Factor Models
• Factor models can be used to define correlations between normally distributed
variables.
• Equation below is for a one‐factor model where each Ui has a component
dependent on one common factor F in addition to another idiosyncratic factor Zi
that is uncorrelated with other variables.
• Steps to construct:
1. Create the SND common factor F
2. Choose a weight α for each Ui
3. Create correlations with F (previous slide)
4. Draw i number of SND idiosyncratic factors Z
5. Calculate U using equation to right
• Advantages of Single Factor Models:
– Covariance matrix is positive‐semidefinite
– Number of correlation estimations is reduced to N from [Nx(N‐1)]/2
• Capital Asset Pricing Model (CAPM) is well known example of Single Factor Model
Common
Factor
Idiosyncratic
Factor
52. OLS – Popular method of Linear Regression
• Linear relationship between dependent and independent variables
• Assumptions
– Independent variable uncorrelated with error term
– Expected value of error term is zero
– Variance of error term is constant
– Error term is normally distributed
Error
Term
Independent
Variable
Slope
Coefficient
Intercept
Dependent
Variable
Sum of Squared
Residuals (SSR)
Explained Sum
of Squares (ESS)
Total …
(TSS)
Analysis of Variance (ANOVA)
55. Confidence Intervals
β1 +/‐ tc x SE
• tc = Critical t‐value using two‐tails
with n‐2 degrees of freedom
• SE = Standard Error as previously
defined
Y +/‐ tc x sf
• tc = is same as for coefficient
• sf = Standard Error of Forecast
2 1
1
1
Coefficient Predicted Value
• Standard Error of the Regression (SER) = Standard deviation of the error terms of
the regression. Measures the degree of variability of the actual Y‐values relative to
the estimated Y‐values. The smaller the SER the greater the accuracy
• s2 = Variance of independent variable X
58. Multicollinearity
• Multicollinearity occurs when two or more X variables are highly
correlated with each other.
• Effects:
– Inflated standard errors, reduces t‐stats
– Fail to reject null hypothesis too often (Type II Error)
– Variables incorrectly look unimportant
• Detection:
– Significant F‐stat overall but insignificant t‐stats
– High correlation between X variables (if only two Xs). If more than two Xs, low
correlations alone cannot rule out multicolinearilty because linear
combinations may still be highly correlated.
• Correcting for multicollinearity is typically accomplished by omitting one
or more independent variables. However, choosing the correct one(s) to
omit can be challenging. Stepwise regression is one commonly used
method.
60. Model Selection Criteria
• MSE is a common metric for comparing models. A ranking of
models by MSE will be identical (but reversed) to that of R2.
• MSE does not increase with more variables, which causes
downward bias in out‐of‐sample variance making it “inconsistent”
• S2 provides a simple method for reducing this bias via a penalty.
• Akaike information criterion (AIC) and Schwartz information
criterion (SIC) provide two other methods for reducing this bias.
• Penalty factors for each bias correction method is shown below.
SIC places the
greatest penalty
while s2 places
the smallest
T = number of observations
k = number of explanatory variables
63. • Detecting heteroskedasticity is most easily accomplished using Scatter Plots which
plot of residuals against each independent variable and against time.
• Correcting for heteroskedasticity is most commonly accomplished using “Robust
Standard Errors”. These can be calculated using “White‐corrected standard errors”
in the estimation.
Heteroskedasticity
Residuals
on Y‐Axis
Independents
on X‐Axis
65. White Noise
• Strong White Noise
– Unconditional mean and variance are constant
– Serially uncorrelated and independent
– Conditional and unconditional mean/variance are same
• White Noise Process is the same as above, but allows for serial dependence.
• Normal White Noise is strong white noise that is normally distributed.
• Testing for White Noise: A Q‐Statistic is often used to test for white noise by
evaluating the overall statistical significance of the autocorrelations. The most
common is the Ljung‐Box Q‐stat (left) where n is the sample size, is the
sample autocorrelation at lag k, and h is the number of lags being tested.
• The Box‐Pierce Q‐stat (right) is the same except that it uses a simple
summation (instead of the weighted sum above) which
67. AutoRegressive Moving Average (ARMA)
Autoregressive Model
• Modeling a series as a function of
past values
• Gradual Decay: Autocorrelation has
long memory because current y is
correlated with all previous y, albeit
with decreasing strength
• ACF will show significant lags beyond
that of PACF.
• Only stationary if ‐1 < φ < 1
Moving Average Model
• Modeling a series as a function of
past residuals
• Autocorrelation Cutoff: “Very short
memory“ because y is only correlated
with a (generally) small, number of
previous y
• PACF will show significant lags
beyond that of ACF
• Stationary for any value of ϴ
AR(1)AR(p) MA(1)MA(q)
ARMA(1,1)• An ARMA includes both AR and MA terms
72. • Covariance matrix must be positive‐semidefinite (PS). What
does that mean?
• Two examples:
– PS
– Not PS
• Small changes to a small PS matrix will likely still be PS, but
small changes to a large PS matrix may can it to not be PS.
Covariance Consistency
Variance Covariance Matrix of
dimensions n x n
Any vector of n real numbers
Transpose of w
74. Simulation Modeling
• Incorporating Correlations – Common Approaches
– Correlations of inputs are implicitly introduced by generating joint
scenarios of input variables
– Samples of historical data are used to define the correlations between
input variables in the model
– Correlation matrix can be specified as an input
• Accuracy
– More simulations (i.e. observations, trials) can increase accuracy (see
formula for Standard Error of the Sample Mean)
– Estimator bias can be introduced via discretization error; the practice of
breaking the simulation into fixed time periods (ex, months, years). This
can be reduced by using shorter time periods, but this also increases
cost of computation