Question 1

1. What stylized features of financial data cannot be explained using linear time series models?

Linear time series models, such as ARMA (Autoregressive Moving Average), assume constant variance and are typically inadequate for modeling certain stylized facts of financial time series data, including:

\(\bullet\) Volatility Clustering: Financial time series often show periods of high volatility followed by periods of low volatility, which linear models cannot capture.

\(\bullet\) Leptokurtosis: Financial returns tend to exhibit “fat tails,” meaning extreme events occur more frequently than predicted by normal distribution.

\(\bullet\) Leverage Effects: Negative shocks often increase volatility more than positive shocks of the same magnitude.

2. Which of these features could be modeled using a GARCH(1,1) process?

A GARCH(1,1) model is designed to handle some of the stylized features that linear models cannot:

\(\bullet\) Volatility Clustering: The GARCH(1,1) model specifically captures time-varying volatility, which explains periods of volatility clustering.

\(\bullet\) Leptokurtosis: GARCH models, due to their ability to model conditional heteroskedasticity, can account for fat tails in return distributions.

3. Why, in recent empirical research, have researchers preferred GARCH(1,1) models to pure ARCH(p)?

Researchers prefer GARCH(1,1) over pure ARCH(p) models because:

\(\bullet\) Parsimony: GARCH(1,1) is more parsimonious than ARCH(p) models, which require many parameters to capture the same effect.

\(\bullet\) Stability and Interpretation: GARCH(1,1) models provide a more stable framework for modeling and interpreting volatility dynamics since they include both lagged values of squared returns and past variances.

4. Describe two extensions to the original GARCH model. What additional characteristics of financial data might they be able to capture?

Two common extensions of the GARCH model are:

\(\bullet\) EGARCH (Exponential GARCH): Captures the asymmetric effects of shocks on volatility, meaning it can model how negative shocks might have a different impact on volatility than positive shocks.

\(\bullet\) GJR-GARCH: Another model that captures the leverage effect, where negative shocks increase volatility more than positive shocks. These extensions better capture the asymmetric responses in volatility and leverage effects, which are important features in financial time series data.

5. Consider the following GARCH(1,1) model

\[y_t=\mu +u_t , \space u_t \sim N(0, \sigma^2_t) \\ \sigma^2_t=\alpha_0+\alpha_1 u_{t-1}^2+ \beta \sigma_{t-1}^2\]

If \(y_t\) is a daily stock return series, what range of values are likely for the coefficients \(\mu,\alpha_0, \alpha_1\ and\ \beta ?\)

For daily stock return series:

\(\bullet\) \(\mu\): The mean return is typically close to zero, as daily returns are often mean-reverting and the long-term average is small.

\(\bullet\) \(\alpha_0\): This is the constant term in the variance equation and should be small and positive to ensure that the variance is always positive.

\(\bullet\) \(\alpha_1\): Represents the short-term impact of past squared returns on current volatility. Values typically range from 0 to 0.2.

\(\bullet\) \(\beta\): Represents the persistence in volatility, often taking values close to but less than 1 (e.g., 0.7 to 0.98), indicating high persistence in volatility.

6. Suppose that a researcher wanted to test the null hypothesis that \(α_1\) + β = 1 in the equation for part (e). Explain how this might be achieved within the maximum likelihood framework.

To test the null hypothesis \(\alpha_1 + \beta=1\)

\(\bullet\) Likelihood Ratio Test: The researcher can use a likelihood ratio test by estimating the GARCH(1,1) model with the unrestricted model (without constraints) and compare it to the restricted model (where \(\alpha_1+\beta=1\)).

\(\bullet\) Maximization: The model parameters are estimated by maximizing the log-likelihood function. Under the null hypothesis, constraints are imposed during maximization.

\(\bullet\) Wald Test: Alternatively, the researcher could use a Wald test to directly test whether the sum \(\alpha_1+\beta=1\)

7. Suppose now that the researcher had estimated the above GARCH model for a series of returns on a stock index and obtained the following parameter estimates: \(\mu = 0.0023, \alpha_0 =0.0172, \beta = 0.9811, \alpha_1=0.1251\). If the researcher has data available up to and including time T, write down a set of equations in \(\sigma_t^2\) and \(\u_{t}^{2}\) their lagged values, which could be employed to produce one-, two-, and three-step-ahead forecasts for the conditional variance of \(y_t\).

The model is given by:

\(\sigma_t^2=\alpha_0+\alpha_1 u_{t-1}^2+\beta \sigma_{t-1}^2\)

where the parameters are:

\(\mu=0.0023, \alpha_0=0.0172, \beta=0.9811, \alpha_1=0.1251\)

One-step-ahead forecast (time T+1)

\(\sigma_{T+1}^2= \alpha_0+\alpha_1u_T^2+ \beta \sigma_T^2\)

Two-step-ahead forecast (time T+2)

\(\sigma_{T+2}^2=\alpha_0+\alpha_1E[u_{T+1}^2]+\beta\sigma_{T+1}^2\)

where \(E[u_{T+1}^2]=\sigma_{T+1}^2\) because \(u_t\) has a conditional variance of \(\sigma_t^2\). Therefore:

\(\sigma_{T+2}^2=\alpha_0+\alpha \sigma_{T+1}^2+\beta\sigma_{T+1}^2=\alpha_0+(\alpha_1+\beta)\sigma_{T+1}^2\)

Three-step-ahead forecast (time T+3)

\(\sigma_{T+3}^2=\alpha_0+\alpha_1E[u_{T+2}^2]+\beta\sigma_{T+2}^2\)

Since \(E[u_{T+2}^2]=\sigma_{T+2}^2\)

\(\sigma_{T+3}^2=\alpha_0+\alpha_1\sigma_{T+2}^2+\beta\sigma_{T+2}^2= \alpha_0+(\alpha_1+\beta)\sigma_{T+2}^2\)

Question 1: Page 271

1. Present two examples in finance and two in econometrics (ideally other than those listed in this chapter!) of situations where a simulation approach would be desirable. Explain in each case why simulations are useful.

1) Finance:

• Option Pricing (Monte Carlo Simulation):

When pricing complex financial derivatives, such as exotic options, where analytical solutions may not exist or are difficult to derive, Monte Carlo simulation can be employed. It allows the estimation of the option’s value by simulating the various paths the underlying asset might take, taking into account factors like volatility, interest rates, and time to maturity. Simulations are useful because they can handle complex, path-dependent features that are challenging for traditional analytical methods.

• Portfolio Risk Management (VaR Calculation):

Value at Risk (VaR) is a commonly used risk metric in finance that estimates the maximum loss a portfolio could face over a given time period with a certain confidence level. Simulations, particularly Monte Carlo methods, are useful for VaR calculation because they can model the joint distribution of portfolio returns, incorporating the complexities of correlation between assets and non-linearities, which are difficult to address analytically.

2) Econometrics:

• Simulating the Impact of Policy Changes: In econometrics, simulations can be used to predict the effects of changes in policy, such as tax reforms or changes in interest rates. For example, a government might want to understand the potential impact of a new tax policy on household income distribution. Simulations allow the modeling of various scenarios and their potential outcomes, incorporating complex interactions within the economy that are hard to capture analytically.

• Relative Merits:

Useful when the underlying model is well understood and can be accurately specified.

Allows for the simulation of hypothetical scenarios that may not be present in historical data.

Can be computationally intensive, especially with complex models.

• Time Series Forecasting (Stochastic Simulation):

When forecasting economic indicators like GDP, inflation, or unemployment rates, stochastic simulations can be used to account for uncertainty in model parameters and future shocks. By simulating multiple possible future paths for a time series, economists can generate a distribution of forecasts rather than a single point estimate, providing a more comprehensive understanding of potential future outcomes.

• Relative Merits:

Does not require strong assumptions about the distribution of the data.

Useful for estimating the sampling distribution of a statistic when the sample size is small or when the data does not meet the assumptions required for parametric methods.

Limited by the quality and size of the original dataset; it cannot simulate scenarios outside the scope of the existing data.

2. Distinguish between pure simulation methods and bootstrapping. What are the relative merits of each technique? Therefore, which situations would benefit more from one technique than the other?

• Pure Simulation Methods:

These involve generating data or outcomes based on a predefined probabilistic model. For example, in Monte Carlo simulations, random numbers are used to simulate the behavior of a financial asset or an economic variable according to its underlying distribution. Pure simulation methods rely heavily on assumptions about the underlying distributions and model structure.

• Bootstrapping:

Bootstrapping, on the other hand, involves resampling from an existing dataset to create new samples. It is a non-parametric method that does not assume any specific distribution of the data. The idea is to generate multiple samples by drawing with replacement from the original dataset, which can then be used to estimate the distribution of a statistic.

3. What are variance reduction techniques? Describe two such techniques and explain how they are used.

• Variance Reduction Techniques: These are methods used in simulations to increase the precision of the estimates without necessarily increasing the number of simulations. By reducing the variance of the estimates, these techniques allow for more accurate and reliable results from simulations.

• Antithetic Variates:

This technique involves generating pairs of negatively correlated variables in simulations. For example, if you are simulating stock prices, you could simulate one path using random numbers and another path using the negative of those random numbers. The idea is that any overestimation in one path will likely be offset by underestimation in the paired path, thereby reducing the overall variance of the estimate.

• Control Variates:

In this technique, you use additional information (a control variate) that is correlated with the variable of interest and has a known expected value. By adjusting the simulation results based on the difference between the observed and expected value of the control variate, you can reduce the variance of the estimate. For example, if you are estimating the mean return of a portfolio, you could use the known mean return of a benchmark index as a control variate. How They Are Used:

These techniques are applied during the simulation process to produce more accurate results with fewer simulations. By strategically controlling the sources of variability in the simulation, they improve the efficiency of the simulation and provide more reliable estimates with the same computational effort.

4. Why is it desirable to conduct simulations using as many replications of the experiment as possible?

Conducting simulations with a large number of replications is desirable because it increases the accuracy and reliability of the results. The law of large numbers suggests that as the number of replications increases, the average of the simulation outcomes converges to the true expected value. This reduces the effect of random fluctuations in the results, leading to more precise estimates and narrower confidence intervals. In practical terms, more replications reduce the risk of making incorrect inferences based on the simulation results.

5. How are random numbers generated by a computer?

Computer generated (‘pseudo-‘) random numbers are not random at all, but are entirely deterministic since their generation exactly follows a formula. In intuitive terms, the way that this is done is to start with a number (a ‘seed’, usually chosen based on a numerical representation of the computer’s clock time), and then this number is updated using modular arithmetic. Provided that the seed and the other required parameters that control how the updating occurs are set carefully, the pseudo-random numbers will behave almost exactly as true random numbers would.

6. What are the drawbacks of simulation methods relative to analytical approaches, assuming that the latter are available?

• Drawbacks of Simulation Methods:

• Computational Intensity:

Simulations, especially those requiring a large number of replications or involving complex models, can be computationally expensive and time-consuming. Approximation Errors:

Since simulations are based on random sampling, there is always some degree of approximation error in the results. This error diminishes with more replications but never completely disappears. Dependence on Model Assumptions:

The accuracy of simulation outcomes heavily depends on the assumptions and specifications of the model used. Incorrect or unrealistic assumptions can lead to misleading results. Interpretation Challenges:

The results of simulations can sometimes be difficult to interpret, especially when dealing with complex models or when the outcomes are sensitive to the assumptions made.

• Advantages of Analytical Approaches (When Available):

• Precision:

Analytical approaches, when applicable, provide exact solutions without approximation errors, offering more precise results. Speed:

Analytical methods, once derived, are typically much faster to compute compared to simulations.

• Insight:

Analytical solutions often provide deeper insight into the relationship between variables and the underlying structure of the problem, which may not be as apparent from simulation results. However, analytical approaches may not always be feasible for complex or non-linear problems, where simulations remain a necessary tool despite their drawbacks.

Question 5:

1. What is a news impact curve?

Using a spreadsheet or otherwise, construct the news impact curve for the following estimated EGARCH and GARCH models, setting the lagged conditional variance to the value of the unconditional variance (estimated from the sample data rather than the mode parameter estimates), which is 0.096:

\[\sigma_{t}^{2}= \alpha_{0}+\alpha_{1}u_{t-1}^{2}+\alpha_{2}\sigma_{t-1}^{2}\] \[ln(\sigma_{t}^{2})= \alpha_{0}+\alpha_{1}\frac{u_{t-1}}{\sqrt{\sigma_{t-1}^{2}}}+ \alpha_{2}ln(\sigma_{t-1}^{2})+\alpha_{3}[\frac{u_{t-1}}{\sqrt{\sigma_{t-1}^{2}}}-\frac{2}{\pi}]\]

2. In fact, the models in part (a) were estimated using daily foreign exchange returns. How can financial theory explain the patterns observed in the news impact curves?

• The financial theory explains that negative shocks often lead to larger increases in volatility than positive shocks (leverage effect). In EGARCH models, the asymmetry in the news impact curve captures this by allowing different reactions to positive and negative shocks.

• The foreign exchange returns exhibit this pattern, where bad news (large negative returns) tends to cause more volatility than good news (large positive returns), which aligns with the empirical findings in financial markets.

Question 2: (Page 721)

A researcher tells you that she thinks the properties of the Ljung Box test (i.e., the size and power) will be adversely affected by ARCH in the data. Design a simulations experiment to test this proposition.

1. Generate Data:

• Simulate a time series with no autocorrelation but with ARCH effects to observe the behavior of the Ljung-Box test under ARCH conditions.

• Simulate a time series with autocorrelation but without ARCH effects as a control group.

2. Apply Ljung-Box Test:

• Perform the Ljung-Box test on both simulated datasets.

• Measure the size (the probability of rejecting the null hypothesis of no autocorrelation when it is true) and the power (the probability of rejecting the null hypothesis when it is false) of the Ljung-Box test.

3. Evaluate Results:

• Compare the size and power of the Ljung-Box test between the datasets with and without ARCH effects. If the researcher’s claim is correct, you should observe that the Ljung-Box test’s properties are impacted by the presence of ARCH in the data.