Forecasting Financial Markets. Time Series Analysis

Transcription

1 Forecasting Financial Markets Time Series Analysis Copyright Investment Analytics Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 1

2 Overview Time series data & forecasts ARIMA models Model diagnosis & testing Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 2

3 Time Series Data & Forecasting YBEG Historical Data Y T YEND Data Y1 Sample Y t Yn Time Forecasts ˆ Yt m Back- Casting Yˆ 0 Yˆ 1 Within- Sample Forecasts Yˆ n Y ˆ n1 Ex-Post Forecasts Yˆ N ˆ N 1 Y Ex-Ante Forecasts ˆ YN k Forecasting Period Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 3

4 Univariate Time Series Models Autoregressive AR(1): y t = a 0 + a 1 y t-1 + e t Moving Average MA(1): y t = e t + b 1 e t-1 e t = sequence of independent random variables Independent Zero mean Constant variance s 2 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 4

5 White Noise Mean is constant (zero) E(e t ) = m (0) Variance is constant Var(e t ) = E(e t2 ) = s 2 Uncorrelated Cov(e t, e t-j ) = 0 for j < > 0 and t Gaussian White Noise If e t is also normally distributed Strict White Noise e t are independent Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 5

6 Lag Operator L m y t = y t-m So AR(1) process can be represented as: (1 - bl) y t = e t Invertibility An AR(1) process can be represented as MA(): If b < 1 y t = (1 - bl) -1 e t y t = [1 + bl + (bl) ] e t y t = e t + b e t-1 + b 2 e t Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 6

7 Stationarity Weak (covariance) stationarity Population moments are time-independent: E(y t ) = m Var(y t ) = s 2 Cov(y t, y t-j ) = g j Example: white noise e t Strong stationarity In addition, y t is normally distributed Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 7

8 Stationary Series Stationary Series ~ N(0,1) Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 8

9 Stationarity of AR(1) Process AR(1) Process: y t = a 0 + a 1 y t-1 + e t Expected value E(y t ) is time-dependent: t 1 0 i0 E( y ) a a a y t i 1 If a 1 < 1, then as t, process is stationary Lim E(y t ) = a 0 / (1 - a 1 ) Hence mean of y t is finite and time independent Also Var(y t ) = E[e t + a 1 e t-1 + a 12 e t ) 2 ] = s 2 [1 + (a 1 ) 2 + (a 1 ) ] = s 2 /[1 - (a 1 ) 2 ] And Cov(y t, y s ) = s 2 (a 1 ) s /[1 - (a 1 ) 2 ] t 1 0 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 9

10 Stationarity Considerations Sample drawn from recent process may not be stationary Hence many econometricians assume process has been continuing for infinite time Can be problematic E.G. FX rate changes post Bretton-woods Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 10

11 Random Walk Process Random Walk with drift y t = a 0 + a 1 y t -1 + e t With a 1 = 1 A non-stationary process Random Walk with Drift Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 11

12 Random Walk Process Random Walk without drift y t = a 0 + a 1 y t -1 + e t With a 1 = 1, a 0 = 0 Dy t = e t or y t = (1- L) -1 e t = e t + e t + e t-1 + e t-2... Also a non-stationary process Variance of y t gets larger over time Hence not independent of time. Var n 2 2 ( yt ) E et 2ete s ns 1 ts Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 12

13 Moving Average Process MA(1) process y t = e t + be t-1 = (1 + bl)e t Invertibility: b < 1 (1+bL) -1 y t = e t y t = S(-b) j y t-j + e t So MA(1) process with b < 1 is an infinite autoregressive process Similary an AR(1) process with b < 1 is invertible i.e. can be represented as an infinite MA process Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 13

14 MA(1) Process MA(1) Process Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 14

15 ARMA(1, 1) Process y t = a 1 y t-1 + e t + be t-1 ARMA(1, 1) Process y t = ay t-1 + e t + bet Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 15

16 General ARMA Process Any stationary time series can be approximated by a mixed autoregressive moving average model ARMA(p, q) y t = f 1 y t-1 + f 2 y t f p y t-p + e t + q 1 e t-1 + q 2 e t q q e t-q F(L) y t = q(l)e t F and Q are polynomials in the lag operator L F(L) = 1 - f 1 L - f 2 L f p L p Q(L) 1 q 1 L + q 2 L q q L q Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 16

17 Unit Roots Stationarity Condition Roots of f(l) must lie outside the unit circle x i > 1 for all roots x i Invertibility Condition Roots of q(l) must lie outside the unit circle z i > 1 for all roots z i Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 17

18 Autocorrelation Population autocorrelation between y t and y t-t r t g t /g 0 (t 1, 2,...) g t is the autocovariance function at lag t g t = Cov(y t, y t-t ) g 0 = Var (y t ) r 0 = 1, by definition Sample autocorrelation: r t c t /c 0 Where c t is the sample autocovariance c t 1 n t n tt 1 ( y t y)( y tt y) Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 18

19

20 ACF for AR(1) Process AR(1) Process: y t = a 0 + a 1 y t-1 + e t Correlation: r s = (a 1 ) s, s = 0, 1,... Since: g 0 s 2 /[1 - (a 1 ) 2 ] g s s s (a 1 ) s / [1 - (a 1 ) 2 ] Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 20

21 a 1 = 0.75 ACF for AR(1) Process ACF for AR(1) Process Estimated Theoretical Lag a 1 = ACF for AR(1) Process Theoretical Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 21 Lag Estimated

22 ACF for MA(1) Process MA(1) Process: y t = e t + be t-1 Yule-Walker Equations g 0 = Var(y t ) = E(y t y t ) = E[(e t + be t-1 ) (e t + be t-1 )] = (1 + b 2 )s 2 g 1 = E(y t y t-1 ) = E[(e t + be t-1 ) (e t-1 + be t-2 )] = bs 2 g s = E(y t y t-s ) = E[(e t + be t-1 ) (e t-s + be t-s-1 )] = 0, s >1 ACF r 0 = 1 r 1 = b / (1 + b 2 ) r s = 0, for s > 1 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 22

23 b = 0. 5 ACF for MA(1) Process ACF for MA(1) Process b = -0.5 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 23

24 Partial Autocorrelation Function (PACF) In AR(1) process y t and y t-2 are correlated Indirectly, through y t-1 r 2 = Corr(y t, y t-2 ) = Corr(y t, y t-1 ) * Corr(y t-1, y t-2 ) = r 1 2 Partial autocorrelation between y t and y t-s Eliminates effects of intervening values y t-1 to y t-s+1 Effectively doing autoregression of y t against y t-1 to y t-s y t = S b i y t-i + e t Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 24

25 Form series y* t = y t - m M is mean e{y t } Calculating PACF Form first-order autoregressive equation Y* t = f 11 y* t-1 + e t e t is error process which may not be white noise f 11 is both AC and PAC between y t and y t-1 Form second-order autoregressive equation Y* t = f 21 y* t-1 + f 22 y* t-2 + e t F 22 is PAC between y t and y t-2, i.e autocorrelation between y t and y t-2 controlling (netting out) effect of y t-1 Repeat for all additional lags to obtain PACF Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 25

26 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 26 PACF by Yule-Walker Form PACF from ACF f 11 = r 1, f 22 = (r 2 - r 12 ) / (1 - r 12 ) Formula for additional lags s = 3, 4, s j j s s j j s s s ss r f r f r f 1 2,... 1, 1,, 1 s j j s s ss j s sj f f f f

27 PACF for AR and MA Processes For AR(p) process No direct correlation between y t and y t-s for s > p Hence f ss = 0 for s > p Good means of indentifying AR(p) type process For MA(1) process y t = e t + be t-1 = (1 + bl)e t y t = S(-b) j y t-j + e t for b < 1 Hence y t is correlated with all its own lags PACF will decay geometrically Direct if b < 0 Alternating if b > 0 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 27

28 Lab: ARMA(1, 1) Process ARMA(1, 1): y t = a 1 y t-1 + e t + b 1 e t-1 Lab: Generate time series Compute theoretical ACF Yule-Walker equations Estimate sample ACF Autocorrel function How does pattern of ACF depend on parameters? Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 28

29 Solution: ARMA(1, 1) Process Yule-Walker Equations g 0 = E(y t y t ) = a 1 E(y t-1 y t )+ E(e t y t ) + b 1 E(e t-1 y t ) = a 1 g 1 + s 2 + b 1 E[e t-1 (a 1 y t-1 + e t + b 1 e t-1 )] = a 1 g 1 + s 2 + b 1 (a 1 + b 1 ) s 2 g 1 = E(y t y t-1 ) = a 1 E(y t-1 y t-1 )+ E(e t y t-1 ) + b 1 E(e t-1 y t-1 ) = a 1 g 0 + b 1 s 2 g s = E(y t y t-s ) = a 1 E(y t-1 y t-s )+ E(e t y t-s ) + b 1 E(e t-1 y t-s ) = a 1 g s-1 Solution g b1 2a1b 1 2 s (1 a ) 2 1 g 1 (1 a1b1 )( a b1) 2 s (1 a ) Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: (1 a1b1)( a1 b1) r1 2 (1 b 2a b ) 1 1 1

30 ACF for ARMA(1, 1) Process a 1 = b 1 = ACF for ARMA(1,1) Process Theoretical Estimated Lag a 1 = 0.6, b 1 = ACF for ARMA(1,1) Process Theoretical Estimated Lag Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 30

31 PACF for ARM(1,1) Process a 1 = b 1 = PACF for ARMA(1,1) Process Theoretical Estimated Lag PACF for ARMA(1,1) Process a 1 = 0.7, b 1 = Theoretical Estimated Lag Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 31

32 Properties of ACF and PACF Process ACF PACF White Noise All r s = 0 All f ss = 0 AR(1): a > 0 Geometric decay: r 1 = a s f 11 = r 1 ; f ss = 0; s>1 AR(1): a < 0 Oscillating decay: r 1 = a s f 11 = r 1 ; f ss = 0; s>1 MA(1): b > 0 +ve spike at lag 1. Oscillating decay r 0 = 0 for s > 1 f 11 > 0 MA(1): b < 0 -ve spike at lag 1. Decay r 0 = 0 for s > 1 f 11 > 0 ARMA(1,1): a < 0 Geometric decay at lag 1 Osc. decay at lag 1 Sign r 1 = sign(a+b) f 11 = r 1 ARMA(1,1): a > 0 Oscillating decay at lag 1 Geom. decay at lag 1 Sign r 1 = sign(a+b) f 11 = r 1 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 32

33 Box-Jenkins Methodology Phase I - identification Identify appropriate models Phase II - estimation & testing Estimate model parameters Check residuals Phase III application Use model to forecast Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 33

34 Phase I - Identification Data preparation Transform data to stabilise variance Difference data to obtain stationary series Model selection Examine data, ACF and PACF to identify potential models Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 34

35 Phase II - Estimation & Testing Estimation Estimate model parameters Select best model using suitable criterion Diagnostics Check ACF/PACF of residuals Do portmanteau test of residuals Are residuals white noise? If not, return to phase I (model selection) Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 35

36 Model Selection Criteria Two objectives Minimize sums of squares of residuals Can always reduce by adding more parameters Parsimony Avoid excess paramterization I.E. Loss of degrees of freedom Better forecasting performance Solution Penalize the likelihood for each additional term added to model Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 36

37 Likelihood Function Assume y t ~ No(m, s 2 ) Likelihood L = (-n/2)[ln(2) +Ln(s 2 )] - (1/2s 2 )S(y t - m) 2 Maximizing wrt m, s 2 : MLE Estimates m = Sy t / n (s) 2 = S(y t - m) 2 / n Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 37

38 Likelihood Function in Regression Simple Linear Regression: y t = bx t + e t e t ~ IID No(0, s 2 ) Likelihood L = (-n/2)[ln(2) +Ln(s 2 )] - (1/2s 2 )S(y t - bx t ) 2 MLE Estimates (s) 2 = S(e t ) 2 / n b S(x t y t ) / S(x t ) 2 Standard Error s b = s / {S(x t - x mean ) 2 } 1/2 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 38

39 Maximum Likelihood Estimation Akaike information criterion (AIC) AIC = -2Ln(Likelihood) + 2m nln(sse) + 2m Schwartz Bayesian information criterion (BIC) BIC = -2Ln(Likelihood) + mln(n) nln(sse) + mln(n) L is likelihood function SSE is error sums of squares n is number of observations m is number of model parameters Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 39

40 Using MLE Model Criteria When comparing models: N should be kept fixed E.G. With 100 data points estimate an AR(1) and AR(2) using only last 98 points. Use same time period for all models AIC and BIC should be as small as possible What matter is comparative value of AIC or BIC BIC has better large sample properties AIC will tend to prefer over-paramaterized models Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 40

41 Sums of Squares Sums of Squares Due to Model = SSM SSM ( yˆ y) t 2 Due to Error = SSE SSE y t yˆ ( t 2 ) Total Sums of Squares = SST = SSM + SSE SST ( y y) t 2 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 41

42 ANOVA and Goodness of Fit F test statistic = MSR/MSE With 1 and n-m-1 degrees of freedom n is #observations, m is # independent variables Large value indicates relationship is statistically significant Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 42

43 Coefficient of Determination R 2 = SSR/SST How much of total variation is explained by regression Adjusted R 2 Adjusted R 2 = 1 - (1 - R 2 ) (n - 1) / (n - m - 1) Idea: R 2 can always increase by adding more variables Penalize R 2 for loss of degrees of freedom Useful for comparing models with different # independent variables m Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 43

44 Diagnostic Checking You need to check residuals: e i = (y i - f i ) Residual = actual - forecast Residual plot: residual vs. actual Residual plot should be random scatter around zero If not, it implies poor fit, confidence intervals invalid However, estimates are still the best we can achieve, but we can t say how good they are likely to be. Test for: Bias: non-zero mean Heteroscedasticity (non-constant variance) Non-normality of residuals Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 44

45 Residual Residual Plot F t Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 45

46 Residual Residual Plot - Bias F t Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 46

47 Residual Residual Plot - Heteroscedasticity F t Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 47

48 Anderson, Bartlett & Quenoille ACF and PACF coefficients ~ No(0, 1/n) If data is white noise Hence 95% of coefficients lie in range 1.96/n Stationary Series ACF Lag Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 48

49 Durbin-Watson Test DW n t2 ( e t1 ) 2 t1 Check for serial autocorrelation in residuals Range: 0 to 4. DW 2 for white noise Small values indicate +ve autocorrelation Large values indicate -ve autocorrelation NB not valid when model contains lagged values of y t Use DW-h = (1 - DW/2){n/[1-ns a ]} ~ no(0,1) for large n S a is the standard error of the one-period lag coefficient a 1 t n e e 2 t Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 49

50 Portmanteau Tests: Box-Pierce Simultaneous tests of ACF coefficients to see if data (residuals) are white noise Box-Pierce Q n s1 Usually h 20 is selected Used to test autocorrelations of residuals h r 2 s If residuals are white noise the Q ~ c 2 (h-m) m is number of model parameters (0 for raw data) Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 50

51 Portmanteau Tests: Ljung-Box More accurate for small n Q* n( n 2) h s1 r 2 s n s If data is white noise then Q* ~ c 2 (h-m) Usual to conclude that data is not white noise if Q exceeds 5% of right hand tail of c 2 distn. Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 51

52 Tests for Normality Error Distribution Moments Skewness: should be ~ 0 Kurtosis: should be ~ 3 Jarque-Bera Test J-B = n[skewness / 6 + (Kurtosis 3) 2 / 24] J-B ~ c 2 (2) Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 52

53 Lab: Box-Jenkins Analysis Test Data Set 1 Fit ARMA model using Using box Jenkins methodology Compute & examine ACF and PACF Estimate MLE model parameters Check residuals using portmanteau tests How good is your model at forecasting? 5.0 Time Series and Forecast Actual Forecast Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 53

54 Solution: Box-Jenkins Analysis Test Data Set 1 ACT and PACF suggest AR(1) Model ACF and PACF - Time Series ACF PA C F U p p er 9 5 % Lo wer 9 5 % Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 54

55 Solution: Box-Jenkins Analysis Test Data Set 1 MLE Estimate: a = Residuals are white noise: ACF & PACF - Residuals ACF PA C F U p p er 9 5 % Lo wer 9 5 % Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 55

56 Forecast Function E.g. AR(1) process: y t+1 = a 0 + a 1 y t + e t+1 Forecast Function E t (y t+1 ) = a 0 + a 1 y t E t (y t+j ) = E t (y t+j y t, y t-1, y t-2,..., e t, e t-1,...) E t (y t+2 ) = a 0 + a 1 E t (y t+1 ) = a 0 + a 0 a 1 + a 12 y t E t (y t+j ) = a 0 (1 + a 1 + a a 1 j-1 ) + a 12 y t a 0 /(1- a 1 ) Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 56

57 Forecast Error J-step ahead forecast error: h t (j) = y t+j - E t (y t+j ) h t (1) = y t+1 - E t (y t+1 ) = e t+1 h t (2) = y t+2 - E t (y t+2 ) = e t+2 + a 1 e t+1 h t (j) = e t+j + a 1 e t+j-1 + a 12 e t+j a 1 j-1 e t+1 Forecasts are unbiased E t [h t (j)] = E[e t+j + a 1 e t+j-1 + a 12 e t+j a 1 j-1 e t+1 ] = 0 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 57

58 Forecast Variance Confidence Intervals Var[h t (j)] = s 2 [1 j + a a a 1 2(j-1) ] s 2 /(1- a 1 2 ) Forecast variance is an increasing function of j In limit, forecast variance converges to variance of {y t } Confidence Intervals Var[h t (1)] = s 2 Hence 95% confidence interval for y t+1 is a 0 + a 1 y t 1.96s Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 58

59 Non-Stationarity Non-stationarity in the mean Differencing often produces stationarity E.g. if y t is random walk with drift, Dy t is stationary Differencing Operator: D d Difference y t d times to yield stationary series D d (y t ) For most economic time series d = 1 or 2 is sufficient Non-stationarity in the variance Use power or logarithmic transformation E.g. Stock returns r t = Ln(P t+1 / P t ) Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 59

60 ARIMA(p,d,q) models ARIMA Models Autoregressive Integrated Moving Average d is the order of the differencing operator required to produce stationarity (in the mean) Many economic time series are modeled ARIMA(0,1,1) Dy t = a 0 + e t + b 1 e t-1 e.g. GDP, consumption, income Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 60

61 Seasonal Models Box-Jenkins technique for seasonal models No different than for non-seasonal Seasonal coefficients of the ACF and PACF appear at lags s, 2s, 3s,.. Examples of Seasonal Models Additive y t = a 1 y t-1 + a 4 (y t-4 ) + e t y t = e t + b 4 (e t-4 ) + e t Multiplicative (1 - a 1 L)(1 - a 4 L 4 )y t = (1 - b 1 L)e t Captures interactive effects in terms e.g. (a 1 a 4 y t -5 ) Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 61

62 How to Model Seasonal Data Explicitly in model With AR and/or MA terms at lag S Seasonal differencing Difference series at lag S to achieve (seasonal) stationarity E.g. for monthly seasonality form y* t = y t - y 12 Model resulting stationary series y* t in usual way Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 62

63 Lab: Modelling the US Wholesale Price Index 140 US Wholesale Prices Index (1985 = 100) Jan-60 Jan-62 Jan-64 Jan-66 Jan-68 Jan-70 Jan-72 Jan-74 Jan-76 Jan-78 Jan-80 Jan-82 Jan-84 Jan-86 Jan-88 Jan-90 Jan-92 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 63

64 Lab: US Wholesale Price Index Data preparation Clearly non-stationary in mean and variance Consider DLn(WPI) Identification for transformed series Examine transformed series, ACF and PACF Seasonal (quarterly) Model estimation & testing AR(2) ARMA(1,1) ARMA(2,1) with seasonal term at lag 4 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 64

65 Solution: US Wholesale Price Index Best model is seasonal ARMA[1, (1,4)] y t = y t-1 + e t e t e t-4 Model a 0 a 1 a 2 b 1 b 4 AIC BIC Adj. R 2 AR(1) % 0.04% 0.00% AR(2) % 0.52% 0.00% 0.46% ARMA(1,(1,4)) % 5.96% 0.03% 3.48% 0.07% ARMA(2,(1,4)) % 6.25% 0.02% 43.38% 2.98% 0.06% Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 65

66 Solution: US Wholesale Price Index Changes in Log(Wholesale Prices Index) Actual Forecast Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 66

67 Regression Models Linear models of form: Y t = b 0 + b 1 X 1t + b 2 X 2t b m X mt + e t {e t }is strict white noise process X i are independent, explanatory variables May or may not be causal Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 67

68 Example: Regression Model for Excess Equity Returns Pesaran & Timmermann (1974) Y t b0 b YSP b PI12 b DI11 b DIP12 1 t1 2 t2 3 t1 4 t2 e t Y t is excess return on S&P500 over the 1-month T-Bill rate. YSP is the dividend yield, defined as: 12-month average dividend / month-end S&P500 Index value PI12 is the rate of change of the 12-month moving average of the producer price index: PI12 = Ln{PPI12 / PPI12(-12)} DI11 is the change in the 1-month T-Bill rate DIP12 is the rate of change of the 12-month moving average of the index of industrial production DIPI12 = Ln{IP12 / IP12(-12)} Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 68

69 Regression Methods Standard Method Use all data Problem: data dependent; structural change Stepwise Forward: start with minimal model, add variables Backward: start with full model and eliminate variables Estimate contribution of individual variables Rolling/ Recursive Re-estimate regression over overlapping, successive fixedlength periods Re-estimate regression after adding each new period s data Useful for ex-ante estimation & out of sample forecasting Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 69

70 Lab: Recursive Regression Prediction of Excess Equity Returns Replicate part of Pesaran & Timmermann study Monthly SP500 excess returns Use recursive regression & ex-ante variables Examine forecasting performance Develop trading system Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 70

71 Recursive Parameter Estimates Recursive Parameter Estimation M2 1962M M6 1968M2 1970M M6 1976M2 1978M M6 1984M2 1986M M6 1992M2 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 71

72 Parameter Estimates & ANOVA PARAMETERS SE t-statistic Prob 1.497% 0.003% 0.002% 0.595% 0.009% ANOVA R 2 8.6% Correl 20.7% F DF Prob 0.000% Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 72

73 Residuals 20% 15% 10% 5% Errors 0% -6.00% -4.00% -2.00% 0.00% 2.00% 4.00% 6.00% 8.00% 10.00% 12.00% 14.00% -5% -10% -15% -20% -25% Forecast Excess Returns Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 73

74 Trading System Performance S&P500 Cumulative Trading Returns 350% 300% Regression Buy & Hold 250% 200% 150% 100% 50% 0% -50% 1960M2 1962M2 1964M2 1966M2 1968M2 1970M2 1972M2 1974M2 1976M2 1978M2 1980M2 1982M2 1984M2 1986M2 1988M2 1990M2 1992M2 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 74

75 Random Walk Model Special case of AR(1) with a 0 = 0 and a 1 = 1 y t = y t-1 + e t y t = y 0 + Se i Mean is Constant for i = 1,..., t E(y t ) = E(y 0 ) + E(Se i ) = y 0 Conditional Mean = y t y t+s = y t + Se t+i for i = 1,..., s E t (y t+s ) = y t + E t (Se t+i ) = y t Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 75

76 Shocks and Random Walks Series is permanently affected by shocks e t has non-decaying effect on {y t } Variance is time-dependant Var(y t ) = Var(Se t ) = ts 2 Hence non-stationary Covariance E[(y t - y 0 )(y t-s - y 0 ) = E[(Se i )(e t-s + e t-s e 1 )] = E[(e t-s ) (e 1 ) 2 ] g t-s = (t - s)s 2 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 76

77 Correlation of Random Walk Process Correlation: r s = [(t-s)/t] 1/2 For small s, (t-s)/t 1 As s increases, r s will decay very slightly Identification Problem Can t use ACF to distinguish between a unit root process (a 1 = 1) and one in which a 1 is close to 1 Will mimic an AR(1) process with a near unit root Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 77

78 Testing for Random Walk AR process y t = a 1 y t-1 + e t Hypothesis test a 1 = 0 Can use t-test OLS estimate of a 1 is efficient Because a 1 < 1 and {e t } is white noise Hypothesis test a 1 = 1; can t use t-test {y t } is non-stationary: y t = Se i Variance becomes infinitely large OLS estimate of a 1 will be biased below true value a 1 ~ r 1 = [(t-1)/t] 1/2 < 1 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 78

79 Random Walk Example Appears stationary ACF decays to zero Random Walk with Drift: y t = y t-1 + e t ACF for Random Walk Estimated Lag Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 79

80 Dickey-Fuller Methodology Use Monte-Carlo Generate 10,000 unit root processes {y t } Estimate parameter a 1 Estimate confidence levels: 90% of estimates are less than 2.58 SE from 1 95% of estimates are less than 2.89 SE from 1 99% of estimates are less than 3.51 SE from 1 Test Example Suppose we have series for which estimated value of parameter a 1 is 2.95 SE < 1 Reject hypothesis of unit root at 5% level Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 80

81 Dickey-Fuller Tests Unit Root Process: y t = a 1 y t-1 + e t Equivalent form Dy t = gy t-1 + e t g = 1 - a 1 Test: g = 0 Equivalent to testing a 1 = 1 Other unit root regression models Dy t = a 0 + gy t-1 + e t Dy t = a 0 + gy t-1 + a 2 t + e t Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 81

82 Dickey-Fuller Test Procedure Test Procedure Estimate g using OLS Compute t-statistic Divide OLS estimate by SE Compare t-statistic with appropriate critical value in Dickey-Fuller tables Critical value depends on sample size form of model confidence level Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 82

83 Critical Values Model Hypothesis Test Statistic 95% and 99% critical values Dy t = a 0 + gy t-1 + a 2 t + e t g = 0 t t & g = a 2 = 0 f & 8.73 a 0 g = a 2 = 0 f & 6.50 Dy t = a 0 + gy t-1 + e t g = 0 t m & a 0 g = 0 f & 6.70 Dy t = gy t-1 + e t g = 0 t & Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 83

84 Joint Tests Used to test joint hypotheses e.g. a 0 = g = 0 Constructed like ordinary F-test f i [ RSS ( restricted ) RSS ( unrestricted ) RSS ( unrestricted ) /( T k) RSS(restricted) = error sums of squares from restricted model RSS(unrestricted) = error sums of squares from unrestricted model r = # restrictions T = # observations k = # parameters in unrestricted model Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 84 / r

85 Extensions of Dickey-Fuller AR(p) Process y t = a 0 + a 1 y t a p-2 y t-p+2 + a p-1 y t-p+1 + a p y t-p + e t Add and subtract a p y t-p+1 y t = a 0 + a 1 y t a p-2 y t-p+2 + (a p-1 + a p )y t-p+1 - a p Dy t-p+1 + e t Add and subtract (a p-1 + a p )y t-p+2 y t = a 0 + a 1 y t (a p-1 + a p ) Dy t-p+2 - a p Dy t-p+1 + e t Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 85

86 Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 86 General Form of AR(p) Process If g = 0, equation has unit root (since all in differences) Hence can use same Dickey-Fuller statistic No intercept or trend: t Intercept, no trend: t m Intercept and Trend: t t D D p i t i t i t t y y a y e b g p i j i i p i i a b a 1 1 g

87 Problems With Dickey-Fuller How to handle MA terms Invertibility: MA model AR() model Said & Dickey: ARIMA(p,1,q) ARIMA(n, 1, 0) N T 1/3 Require order of AR(p) process to estimate g Start with long lag and pare down model using standard t-tests Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 87

88 Tests for Multiple Unit Roots Dickey & Pantula Perform DF tests on successive differences E.g. 2 unit roots suspected Form D 2 y t = a 0 + b 1 Dy t-1 + e t Use DF t statistic to test b 1 = 0 If b 1 differs from zero then test for single unit root Form D 2 y t = a 0 + b 1 Dy t-1 + b 2 y t-2 + e t Test null hypothesis: b 1 = 0 using DF If rejected, conclude {y t } is stationary Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 88

89 Phillips-Perron Tests Phillips-Perron generalizes DF to cover: Serially correlated errors and non-constant variance Models: y t = a 0 + a 1 y t-1 + a 2 t + m t Test a 1 = 0 using standard DF critical values and statistic: 2 X 1/ Tw 2 2 ) ( D X = det(x T X), the determinant of the regressor matrix X S is the standard error of the regression w is the # of estimated correlations ~ s Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: s~ s~ S ~ 4 T 1 ~ s S ~ t D T T T Tw ui T i1 T s1 ts1 a T w u t u ts T w

90 Problems in Testing for Unit Roots Low power of unit root tests Can t distinguish between unit root and near unit root process Too often indicate that process contains unit root Tests are conditional on model form Tests for unit roots depend on presence of deterministic regressors Test for deterministic regressors depend on presence of unit roots Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 90

91 Unit Roots In FX Markets Purchasing power parity Currency depreciates by difference between domestic & foreign inflation rates PPP model E t = p t - p* t + d t E t is log of dollar price of foreign exchange p t is log of US price levels p* t is log of foreign price levels d t represents deviation from PPP in period t Testing PPP Reject if series {d t } is non-stationary Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 91

92 Real Exchange Rates Real exchange rates Define r t e t + p* t - p t PPP holds if {r t } is stationary Create series using: r t = Ln(S t x WPI JP t / WPI US t) S t is the spot yen fx rate at time t WPI JP t is the Japanese whole price index at time t (Feb 1973 = 100) WPI US t is the US whole price index at time t Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 92

93 Lab: Testing Purchasing Power Parity Worksheet: PPP Series of real Yen FX rates Dickey Fuller Test Form series Dr t = a 0 + gr t-1 + e t Estimate parameters using max. likelihood Do T-Test D-F test with critical value of Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 93

94 Solution: Purchasing Power Parity MLE SE t p Max Likelihood a % AIC g % BIC DW 2.03 m 1 R 2 1.6% n 202 Adj. R 2 1.1% ANOVA DF SS MS F p Portmanteau Tests Model % Q(24) p Error Box-Pierce % Total Ljung-Box % T-Test: H 0 : g = 0 Could reject at the 93% confidence level Conclude series is stationary and PPP holds Dickey-Fuller Can t reject unit root hypothesis at 95% level Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 94

95 Summary: Time Series Analysis Simple methods Exponential smoothing, etc. Simple, low cost, often effective Limitations Query out of sample performance Underlying model not articulated ARIMA models Staple of econometricians Models articulated and testable Limitations Estimation is non-trivial Problems with (near) random processes Copyright Investment Analytics Forecasting Financial Markets Time Series Analysis Slide: 95