Вы находитесь на странице: 1из 4

BSc Econometrics FEB23001-11 Time Series Analysis Assignment 1

Instructions This assignment should be made with groups of two students. It is not allowed to cooperate with other groups, or to copy their results. Possible founded fraud will be punished. The results of this assignment should be written down in a report consisting of max. 8 pages (not including a possible appendix). The report may be written in English or Dutch. The report should provide a detailed and careful description of your results and interpretation. Make the report self-contained, so that it can be read without reading these instructions rst. All relevant results should be included in the report (preferably in tables and graphs). Raw EViews output may be attached as an appendix, but should not be included in the main text of the report. Reports should be submitted by handing in a hard copy at the box outside the oce of Dick van Dijk at H11-05 before Friday May 18, 2012 before 12:00h. In addition you should send your report by e-mail to opschoor@ese.eur.nl. Reports should be submitted as a PDF le with the name FEB23001-11-A1-XXXXXX-YYYYYY.pdf, where XXXXXX and YYYYYY are the student numbers of the two group members. Note 1 : The report may be written in English or Dutch. Note 2 : Motivate your answers! For example, answers like Yes or No to question 1.1.b will not be rewarded any points. Note 3 : Some questions in the assignment are not specied up to the last detail and require you to make some choices yourself. Explain and motivate your choices in your report. Note 4 : Not all computations need to be done in EViews. You are free to use other programs such as Excel or Matlab. In cases you have to write a program, provide also the code in the appendix! Good luck!! For questions related to this assignment, please send an e-mail to Anne Opschoor at opschoor@ese.eur.nl

ASSIGNMENT 1.1 DATA The Excel le data ICI.xls contains 1259 daily observations on the Implied Correlation Index, which is an measure of the expected average correlation between stocks that are included in the S&P 500 Index1 over the period January 3, 2007 - December 30, 2011. The series is referred to as ICI below. 1.1.a) Create an EViews workle that contains daily observations for the period January 3, 2007 - December 30, 2011. Make a plot of the time series of the ICI. Do you observe any of the key features like a trend, aberrant observations and heteroskedasticity discussed in Chapter 2? You may use the simple regression tools discussed in this chapter. Provide also summary statistics of the ICI. Is this series normally distributed?
[EViews comment: This can be done in two ways. One way is to create an unbalanced workle with the required number of observations and import the data via . . . Import - Read Text-Lotus-Excel.... The second way is to immediately import the data as an workle via . . . Open - Foreign Data as Workfile. The dierence is that in the latter case, each observation is matched by the corresponding date, while the former way does not incorporate the corresponding date.]

1.1.b) Consider the sample January 3, 2007 - December 31, 2009 (755 observations). Compute the rst 20 empirical autocorrelations and partial autocorrelations for the ICI. Examine the signicance of the autocorrelations. What kind of AR, MA or ARMA model do the (partial) autocorrelations suggest?
[EViews hint: In the Series Toolbar (in the window for the series ICI), select View - Correlogram...]

1.1.c) Estimate AR(p) and MA(q) models for p = 0, . . . , 4 and q = 1, . . . , 4 for the ICI series on the same sample of the previous question (January 3, 2007 - December 31, 2009), and record the values of the Akaike Information Criterion and the Schwarz Information Criterion. (Include a constant in all models.) Which model do these criteria indicate as the preferred choice?
[EViews comment: To estimate an MA(2) model for a seried y, for example, you can give the commandy c MA(1) MA(2). This will estimate an MA(2) model with intercept (to allow for a non-zero mean). If you want to estimate an AR(2) model, you may enter y c AR(1) AR(2), but you could also write y c y(-1) y(-2). However, note that the estimated coecients have a dierent interpretation! See the EViews help for more information.]

1 This variable is constructed by the Chicago Board Options Exchange (CBOE) and is based on options (nancial asset), see http://www.cboe.com/micro/impliedcorrelation/. The S&P 500 Index is a famous equity Index in the United States.

1.1.d) Estimate an AR(1) model (with a constant included) for the ICI series based on the sample January 3, 2007 - March 18, 2007 (51 observations). According to the seven assumptions of the linear regression model, what are the standard errors of the estimated coecients? Is the asymptotic normal distribution a reasonable approximation to the nitesample distribution of the coecients of the AR(1) model? Perform a simulation study to answer this question.2 When you simulate from an AR(1) model, use 100 observations as an burn-in period. That is, simulate 150 values of yt and drop the rst 100 values. Provide histograms of the simulated distributions of both parameters. Do they look like a normal distribution? Provide your code in the appendix !!! 1.1.e) Estimate an AR(3) model for ICI series, based on the sample January 3, 2007 - December 31, 2009 (755 observations). What are the roots of the AR(3) polynomial? Derive the theoretical values of the rst 20 implied autocorrelations (that is, what are the values of the autocorrelations of an AR(3) process for the estimates 1 , 2 and 3 obtained here),3 and compare these with the empirical autocorrelations. How close do they match? 1.1.f) Create the time series of residuals from the estimated AR(3) model, and inspect its properties. In particular, consider the following questions: a. Are the residuals normally distributed? b. What are the autocorrelation properties of the residuals, and what does this imply for the adequacy of the estimated AR(3) model? c. What are the autocorrelation properties of the squared residuals? What does this imply for the assumption of homoskedasticity of the shocks in the model?
[EViews hints: The series of residuals may be created by selecting Proc - Make Residual Series... in the Equation Toolbar. Please note the following: in case you have estimated the model by entering y c AR(1) AR(2) AR(3), EViews will compute residuals for all observations in the sample period, and in doing so it sets pre-sample values of y equal to the unconditional mean . In this case, the rst residual is computed as e1 = y1 where y1 is the rst observation. 1 (y1 ) etc. As the properties of Similarly, the second residual is computed as e2 = y2 these rst three residuals are quite dierent from those for the remaining observations (which are obtained from the complete model), they should be set to NA before analyzing the properties of the residuals. Note that EViews does this automatically when you create the residual series for the model estimated in the specication with intercept (y c y(-1) y(-2) y(-3)) An easier way to compute the relevant statistics for each of the three questions above is to select View - Residual Tests in the Equation Toolbar. This works correctly both in case you have estimated the model in deviation-from-mean form or in the specication with intercept.] See the lecture slides how you can do this. You should express the autocorrelations for an AR(3) model in terms of the coecients 1 2 and 3 . See Chapter 3 and the lecture slides for examples showing how this can be done.
3 2

1.1.g) Re-estimate the AR(3) model for ICI series using the sample period January 3, 2007 - December 31, 2009 (755 observations). Construct one-, two-, three-, fourand ve-step ahead point forecasts for the sample period January 4, 2010 - December 20, 2011 (504 observations). Hence at each time T (T = 755, 756, . . . , 1254) you have to compute an forecast ICIT +h|T (h = 1, 2, 3, 4, 5). Re-estimate the AR(3) model repeatedly every 25 observations, using an moving window approach with a sample length of 755 observations. Evaluate the accuracy of these point forecasts in absolute terms, by considering their a) unbiasedness, b) accuracy, and c) eciency. Comment also on the possible dierences in the forecast horizon. Did you expect this result, why (not)? Also evaluate the relative accuracy of the point forecasts obtained from the AR(3) model by comparing them with so-called random walk forecasts, where ICI T +h|T = ICIT , that is, the forecast is simply the observed value on the current day. Give also a plot of the estimated coecients of the AR(3) model in the moving window approach. Interpret this gure. Do not forget to provide your written code in the appendix !!!

Вам также может понравиться