KYIV SCHOOL OF ECONOMICS Financial Econometrics (2nd part): Introduction to Financial Time Series May 2011 Instructor: Maksym Obrizan Lecture notes I # 3. This course assumes some basic knowledge of time series econometrics but the most important concepts will be (quickly) reviewed This lecture: very brief review of AR(p), and applications of linear time series models What is Financial Time Series? # 2. Purposes of the short course: (i) Review some theoretical models of financial time series (ii) Develop practical skills of applied financial time series Main Text: Analysis of Financial Time Series, by Ruey Tsay (selected chapters) Data: mostly US time series (the most liquid market) but methods are applicable to transition countries as well # 4. Financial Time Series Of course, financial time series analysis has to incorporate uncertainty about asset returns The most recent global crisis indicates that pricing bubbles and inadequate risk management are still present even in the most (financially) developed markets. # 5. Although, stock returns are often the focus of financial theory other important financial time series include: In addition, some of these methods can be used to study macroeconomic time series such as GDP or its components # 7. Basic concepts Let The sample mean is # 6. Most financial studies use returns, instead of prices of assets Campbell, Lo and MacKinlay (1997) give two reasons for this: (i) return is a scale-free summary of investment opportunity; (ii) returns have more attractive statistical properties than prices # 8. The sample variance is # 9. The third central moment measures the symmetry of X with respect to its mean (skewness) # 10. 0.4 Skewed to the left Standard normal 0.35 0.3 0.25 0.2 0.15 0.1 For figure on slide 10: Skewness is 0.007 for standard normal and -0.839 for skewed to the left 0.05 0 -25 -20 -15 -10 -5 0 5 0.4 4th # 11. The central moment measures the tail behavior of X # 12. t with 1 df Standard normal 0.35 0.3 0.25 0.2 Excess kurtosis K(x)-3 : 0.15 0.1 0.05 Kurtosis of Student t distrbution with 1 df is 3.43 on slide 12 0 -5 -4 -3 -2 -1 0 1 2 3 4 5 # 13. Skewness and Kurtosis in returns data # 14. Stationarity – Skewness A time series {rt} is strictly stationary if High excess kurtosis A time series {rt} is weakly stationary if In practice, high excess kurtosis means that the distribution of returns tends to contain more extreme values than the standard normal In this course: weakly stationary time series # 15. Linear time series # 16. Quick review of autoregressive models AR(p) model is The mean and the variance Meaning: the past p values of rt-I (i=1,…,p) jointly determine the conditional expectation of rt given the past data # 17. To identify the order p of AR(p) model in practice one can use: (i) PACF (ii) information criteria (AIC) # 18. The estimate of the second equation is called the lag-2 sample PACF of rt . Partial Autocorrelation Function (PACF): Intuitively, for an AR(p) model the lag-p sample PACF should not be zero but lag-j PACF should be close to zero for all j>p. # 19. Indeed, under certain regularity conditions the sample PACF of an AR(p) process has the following properties # 20. Alternatively, we can use Akaike Information Criterion for a Gaussian AR(k) model The second term is called the penalty function for adding additional parameters # 21. Monthly Value-Weighted Index Returns # 23. Parameter Estimation # 22. NOTES # 24. Model checking: ACF The sample autocorrelation of series {rt} Notice: is biased (but consistent) estimate of . However, if sample is large then bias is not serious Sample Autocorrelation Function (ACF) # 26. 0.8 Sample Autocorrelation # 25. After you fit the model obtain residual series to check for remaining autocorrelation Suggestion: Also plot 95% confidence intervals 0.6 0.4 0.2 0 Graph to the left? -0.2 Graph below? 0 2 4 6 8 10 Lag 12 14 16 Sample Autocorrelation Function (ACF) # 28. Ljung-Box (1978) statistics – # 27. Sample Autocorrelation 0.8 0.6 0.4 0.2 In practice, the choice of m may affect the performance of Q(m) statistics Simulations suggest to set m to approx. ln(T) 0 -0.2 0 2 4 6 8 10 Lag 12 14 16 18 20 18 20 # 29. For an AR(p) model, the Ljung-Box statistics Q(m) follows asymptotically a chisquared distribution with m-p degrees of freedom # 30. Implications # 31. Forecasting: we are at time h and are interested in forecasting {rh+b} where b>0 # 32. Multistep Ahead Forecast Forecast often employs the minimum squared error term loss function This forecast can be obtained recursively Important: for a stationary AR(p) model the long term forecast converges to unconditional mean (mean reversion) and the variance of forecast error approaches the unconditional variance # 33. NOTES # 34. NOTES # 35. Application: AR(2) model and business cycles Consider an AR(2) model # 36. This equation can be re-written as the second-order difference equation where B is called back-shift operator such that It can be shown that the ACF of a stationary AR(2) model satisfies Sometimes lag operator L is used instead of B # 37. Corresponding to the difference equation there is quadratic equation # 38. If characteristic roots are complex numbers (complex conjugate pair) then the ACF of this series shows damping sine and cosine waves which can be solved for characteristic roots For example, AR(2) model Interesting case when The graph is in the bottom left corner Sample Autocorrelation Function (ACF) 1 # 39. # 40. In business and economic applications complex characteristic roots give rise to business cycles 0.8 Sample Autocorrelation 0.6 For an AR(2) model on slide # 35 with a pair of complex characteristic roots the average length of the stochastic cycles is 0.4 0.2 0 -0.2 -0.4 -0.6 where the cosine inverse is stated in degrees 0 5 10 15 20 Lag 25 30 35 40 # 41. Illustration: US GNP seasonally adjusted from QII.1947 to Q1.1991 Fit AR(3) model Obtain a corresponding third-order difference equation # 43. Application: Seasonal Models Quarterly earnings per share of a company may exhibit cyclical or periodic behavior – seasonal time series # 42. Factor out as For the second-order factor is then 1-0.87B-(-0.27)B2 =0 we have 0.872+4(-0.27)<0 The average length of the stochastic cycles is # 44. # 45. Seasonal differencing # 46. In general, for a time series with periodicity s: # 47. Multiplicative Seasonal Models: The airline model # 48. Application to log series of Johnson and Johnson # 49. # 50. Regression models with time series errors Suppose, we are interested in term structure of interest rates # 51. If the error term is a white noise then the LS method results in consistent estimates # 52. Application to the US weekly interest rate series: r1t - 1-year Treasury constant maturity rate r2t - 3-year Treasury constant maturity rate Simple but inadequate model: # 53. Developing a more adequate model # 54. Cont’d # 55. Fitting a linear regression model with time series errors # 56. NOTES