Time Series Forecasting (Part II) Duong Tuan Anh Faculty of Computer Science and Engineering September 2011

Outline Stationary and nonstationary processes Autocorrelation function Autoregressive models AR Moving Average models MA ARMA models ARIMA models Estimating and checking ARIMA models(Box-Jenkins Methodology)

Stochastic Processes The time series in this part are all based on an important assumption – that the series to be forecasted has been generated by a stochastic process. We assume that X1, X2, …,XT in the series is drawn randomly from a probability distribution. In modeling such a process, we try to describe the characteristics of its randomness. We could assume that the observed series is drawn from a set of random variables. These random variables can be denoted by {Xt, t  T} , T is set of time indices.

Stationary and Nonstationary Processes We want to know whether or not the underlying stochastic process that generated the time series can be invariant with respect to time. If the characteristics of the stochastic process change over time, i.e., if the process is nonstationary, it will be difficult to represent the time series by a simple algebraic model. If the stochastic process is fixed in time, i.e., if it is stationary, then one can model the process via an equation with fixed coefficients that can be estimated from past data. The models described here represent stochatic processes that are assumed to be in equilibrium about a constant mean level. The probability of a given fluctuation in the process from that mean level is assumed to be the same at any point in time.

Stationary processes Mathematically, a stochastic process is called stationary if its first moment and second moment are fixed and do not change in time. The first moment is the mean, E[Xt], and the second moment is the covariance between Xt and Xt+k. The kind of covariance applied on the same random variable is called auto-covariance. Variance of a process, Var[Xt], is a special case of auto-covariance with the lag k = 0. Therefore, a process is called stationary if: Mean: E[Xt] =  <  ,  t (3.1) Variance: Var[Xt] = 2 <  ,  t (3.2) Auto-covariance is dependent only on the lag : (3.3) The set of all auto-covariance coefficients { k }, k = 0, 1, 2, ..creates the autocorrelation function(ACF) of the process. Note that 0 = 2

Autocorrelation Function ACF is often normalized to form the set of all autocorrelation coefficients { k }, k = 0, 1, 2, .. by formula: That means:

Sample autocorrelation function: If a process is stationary, its probability distribution p(Xt) is the same at any time point and its shape (or at least some its properties) can be inferred by looking at a histogram of the observed values X1, X2, …, Xt. And in practice, estimate of the mean  can be computed from the sample mean of the series: Estimate of variance can be computed from the formula of sample variance:

Sample autocorrelation function: And similarly, estimate of the autocorrelation function can be computed from sample autocorrelation function: It’s easy to see from their definitions that both the theoretical and estimated autocorrelation functions a symmetrical, i.e., that the correlation for a positive displacement is the same as that for a negative displacement, so that: k = -k

Figure 1 shows a time series and its autocorrelation function The plot of an auto-correlation function is also called correlogram. The readings from a chemical process with the autocorrelation function and the autocorrelation function.

How to check stationarity of a time series Stationarity of a time series can be identified intuitively through the plot of time series based on the two properties: The values fluctuate around a fixed mean in long-term Data variance does not change at any time in time Implicitly, in the condition of stationarity, we assume that the first two moments of Xt are finite. Another way to decide whether a series is stationary is looking at a plot of the autocorrelation function.

If the autocorrelation function falls off as k, the number of lags, increases, then the series is stationary. In this case, we have to solve two following problems: (a) determine whether a particular value of the sample auto-correlation function is close enough to zero to permit assuming that the true value of the autocorrelation function k is equal to 0. (b) test whether all the values of the autocorrelation function are equal to 0. (If they are, we know that we are dealing with white noise.).

Autoregressive Models (AR) In the autoregressive process of order p the current observation Xt is generated by a weighted average of past observations going back p periods, together with a random disturbance in the current period. A time series {Xt } is an autoregressive process with order p (AR(p)) if each observation Xt of AR(p) process can be denoted as the following equation: Yt = 1Yt-1 + 2Yt-2 +… + pYt-p +  + t (1) where: { t } is a white noise process with mean E[t] = 0, variance 2 and covariance k = 0 for k  0  (or 0) is a constant term which relates to the mean of the stochastic process. Note that  can be zero.

White noise A time series { t } is called a white noise if it is a sequence of independent and identically distributed random variables with finite mean and variance. In particular, if it is normally distributed with mean zero and variance 2, the series is called Gaussian white noise. For a white noise series, all the ACFs are zero. White noise processes may not occur very commonly, but weighted sums of a white noise process can provide a good representation of processes that are nonwhite.

First order Autoregressive Models – AR(1) Consider the first order autoregressive model AR(1): yt =  yt-1 + t (2) The constant term can be omitted without loss of generality. If the expected value of yt is denoted by t, then t =  t-1 From this equation, we get t = Nt-N if the autoregressive process starts at time –N. Since the time series is stationary – i.e. it has no trend, then || < 1. If the process begins in the infinite past, then t = 0 for all t. Squaring both sides of (2) and taking expected value yields the equation: E(yt2) = 2E(yt-12) + E(t2) + 2E(yt-1t) or y2 = 2y2 + 2 where y2 = var(yt) and 2 =var(t)

AR(1) If the series is stationary with zero mean, then To obtain the autocovariance coefficients, go back and multiply (2) by yt-1. That yields: E(ytyt-1) =  E(yt-12) + E(yt-1t) (3) Let denote y2 as 0 and the autocovariance coefficients as 1, 2,…, m. Equation (3) can be rewritten as 1= 0 and in general N= N0 , N  0. The larger the value of , assuming it is between 0 and 1, the smoother yt will be. If  < 0, the series will exhibit a jagged-edge pattern; technically, it will be less smooth than a pure white noise series.

AR(2) 1= 10 + 21 2= 11 + 20 2= 11 + 2 (6) AR(2) has the form: yt = 1 yt-1 + 2 yt-2 + t (4) If both sides of the equation (4) are multiplied by yt and expected values are taken, we get: 0= 11 + 22 + 2 Multiplying (4) by yt-1 and yt-2 and taking expected values yields 1= 10 + 21 2= 11 + 20 The above equations can be restated in terms of the auto-correlation coefficients 1 and 2 as 1= 1 + 2 1 (5) 2= 11 + 2 (6) These are called as the Yule-Walker equations.

Yule-Walker Equations Suppose we have the sample auto-correlation function for a time series which was generated by an AR(2) process. We could then measure 1 and 2 and substitute these numbers into the Yule-Walker equations. We have two algebraic equations which could be solved simultaneously for the two unknowns 1 and 2. Thus, we could use the Yule-Walker equations to obtain estimates of the AR parameters 1 and 2.

AR(p) Model Partial Autocorrelation Function. One problem in constructing autoregressive models is identifying the order p of the underlying process. For moving average models this is less of a problem, since if the process is of order q the sample autocorrelations should all be close to zero for lag greater than q. Although some information about the order of an AR process can be obtained from the oscillatory behavior of the sample ACF, much more information can be obtained from the partial autocorrelation function (PACF). To understand what the PACF is and how it can be used, let us first consider the covariances and autocorrelation function for the AR of order p.

Partial Autocorrelation Function First, notice that the covariance with displacement k is determined from: γk = E[yt-k (1 yt-1 + 2 yt-2 +…+ p yt-p + t )] (7) Now let k = 0, 1, .., p, we obtain the following p +1 difference equations that can be solved simultaneously for γ0, γ1, …, γk: 0= 11 + 22 +..+ pp +2 1= 10 + 21 +..+ pp-1 …………………. …….. ……….. (8) p= 1p-1 + 2p-2 +..+ p0 For displacement k > p the covariances are determined from: k= 1k-1 + 2k-2 +..+ pk-p (9) Now by dividing the left-hand and right-hand sides of the equations in (8) by 0, we can derive a set of p equations that together determine the first p values of ACF:

k= 1 k-1 + 2k-2 +..+ pk-p (11) 1= 1 + 21 +..+ pp-1 2= 11 + 2 +..+ pp-2 …. …. ….. …. … … (10) p= 1p-1 + 2p-2 +..+ p For displacement k > p we have, from Eq. (9): k= 1 k-1 + 2k-2 +..+ pk-p (11) The equations in Eq. (10) are the Yule-Walker equations. If 1,2,…,p are known, then the equations can be solved for 1, 2,…, p.and vice versa. Unfortunately, solution of the Yule-Walker equation requires knowledge of p , the order of the AR process. Therefore, we solve the Yule-Walker equations for successive values of p. In other words, suppose we begin by hypothesizing that p = 1, then Eqs. (10) reduce down to 1= 1 or, using the sample autocorrelations, ’1= ’1. Thus, if the calculated value is significantly different from 0, we know that the AR process is at least order 1. Let us denote this value ’1 by a1 .

Now let us consider the hypothesis that p = 2 and solve the Eqs Now let us consider the hypothesis that p = 2 and solve the Eqs. (10) for p = 2 . Doing this gives us a new set of estimates ’1and ’2. If ’2 is significantly different from 0 we can conclude that the AR process is at least order 2, while if ’2 is approximately 0, we can conclude that p = 1. Let denote the value ’2 by a2. We now repeat this process for successive values of p. For p = 3, we obtain an estimate of ’3, which we denote by a3, … We call this series a1, a2, a3,…the partial autocorrelation function (PACF) and note that we can infer the order of the autoregressive process from its behavior. In particular, if the true order of the process is p, we should observe that aj  0, j > p. In other words, for an AR(p) series, the sample PACF cuts off at lag p.

Let look at the second order autoregressive process AR(2): yt = 1.69 yt-1 – 0.8 yt-2 + t The graph of 120 observations on a series generated by the AR(2) process yt = 1.69 yt-1 – 0.8 yt-2 + t together with the theoretical and empirical ACFs (middle) and the theoretical and empirical PACFs (bottom). The theoretical values corresponds to the solid bars.

How to check whether aj is nonzero To test whether a particular aj is zero, we can use the fact that it is approximately normally distributed, with mean zero and variance 1/T. (T is the number of the data points in the time series). Hence, we can check whether it is statistically significant at, say, the 5 percent level by determining whether it exceeds 2/T in magnitude.

How to estimate parameters of AR(q) For AR(q) process, the difference equation for its autocorrelation function is given by: k= 1k-1 + 2k-2 +..+ pk-p We can rewrite this equation as a set of p simultaneous linear equations relating the parameters 1 ,…,p to 1,…,p: 1= 1 + 21 +..+ pp-1 2= 11 + 2 +..+ pp-2 …. …. ….. …. … … p= 1p-1 + 2p-2 +..+ p Using these Yule-Walker equations to solve for the parameters 1 ,…, p in terms of the estimated values of the autocorrelation function, we arrive at the estimates of the parameters 1, 2,…,p.

Moving Average (MA) Models Time series is called a moving average process of order q (MA(q)) if each observation can be written as the following equation: yt = t - 1t-1 - 2t-2 - qt-q (3.10) where random disturbance component {t } is white noise process with 0 mean, constant variance 2 and auto-covariance k = 0 for k  0. White noise processes may not occur very commonly, but weighted sums of a white noise process can provide a good representation of processes that are non-white noise. So, in the MA(q) each observation yt is generated by a weighted average of random disturbances going back q periods.

Lag operator We can rewrite equation (3.10) in the form: yt = (B)t where: (B)= 1 - 1B - 2B2 - …- qBq is a polynomial of order q in B, and B is the lag operator which is used to describe the lag in time. Bt = t-1 The mean of the moving average process is independent of time since E[yt] =  and  = 0. Each t is assumed to be generated by the same white noise process, so that E[t] = 0, variance 2 and covariance k = 0 for k  0

Stationary MA(q) = 2(1 + 12 + 22 +…+ q2) Let us now look at the variance, denoted by 0 of the moving average process of order q: Var[yt] = 0 = E[(yt - )2] =E(t2 +12t2 +… q2 t-q2 - 21tt-1-…) = 2(1 + 12 + 22 +…+ q2) From the above equation, we see that if MA(q) is the realization of a stationary random process, it must satisfy the following conditions: 1 + 12 + 22 +…+ q2 <  This result is trivial since we have only a finite number of i and thus their sum is finite.

ACF of a stationary MA(q) However, the assumption of a fixed number of i can be considered to be an approximation to a more general model. A complete model of most random process would require an infinite number of lagged disturbance terms (and their corresponding weights). Then, as q, the order of the MA process, become infinitely large, we must require that 1 + 12 + 22 +…+ q2 converge to ensure the stationarity of the MA process. Convergence will usually occur if the i become smaller as i become larger. We will see later that if the process is stationary, by a MA model of order q, we expect the autocorrelation function k will become smaller as k become larger. This is consistent with our result of the previous section that one indicator of stationarity is an ACF that approach to 0.

MA(1) We begin with the moving average process of order 1. The process is denoted by MA(1), and its equation is: yt =  + t - 1t-1 This process has mean  and variance 0 = 2(1 + 12). Now let us derive the covariance for a one-lag displacement 1: 1 = E[(yt - )(yt-1 - )] = E[(t - 1t-1)( t-1 - 1t-2)] = - 12 In general we can determine the covariance for a k-lag displacement to be k = E[(t - 1t-1)( t-k - 1t-k-1)] = 0 for k >1 Thus the MA(1) process has a covariance of 0 when the displacement is more than 1 period. We now can determine the autocorrelation function for the process MA(1): k = k/0 = - 1/(1+ 12) for k = 1 = 0 for k > 1

MA(2) Now let us examine the moving average process of order 2. The process is denoted by MA(2), and its equation is: yt =  + t - 1t-1 - 2t-2 This process has mean  and variance 0 = 2(1 + 12+ 22 ), and covariances given by 1 = E[(t - 1t -1- 2t -2 )( t -1 - 1t -2 - 2t -3)] = - 12 + 2 12 = - 1(1 - 2) 2 2 = E[(t - 1t -1- 2t -2 )( t -2 - 1t -3- 2t -4)] = - 12 and k = 0 for k > 2 The process MA(2) has a memory of exactly two periods, so that the value of yt is influenced only by events that took place in the current period, one period back, and two periods back.

MA(q) The MA process of order q has a memory of exactly q periods. Autocorrelation function of the MA process of order q is given by the following: k = 1 if k = 0 = (-k+ 1k+1+…+ q-k q)/(1 + 12 + 22 +…+ q2 ) if k= 1,2,..,q = 0 if k > q So, we can see why the sample ACF can be useful in specifying the order of a moving average process. An MA(q) series is only linearly related to its first lagged values and hence is a "finite memory" model.

An example of a second-order MA process might be: yt = t + 0 An example of a second-order MA process might be: yt = t + 0.9 t-1 + 0.8t-2 Figure 2. The graph of 120 observations on a series generated by the MA(2) process yt = t + 0.9 t-1 + 0.8t-2 together with the theoretical and empirical ACFs (bottom) and the theoretical and empirical PACFs (middle). The theoretical values corresponds to the solid bars.

Durbin’s method for estimating MA(q) Given the MA(q) with the equation: yt = t - 1t-1 - 2t-2 -… - qt-q The method for estimating MA(q) consists of two steps: Step 1: The first step consists of fitting an AR model of order m > q to {yt}. Once m has been specified, the estimated AR parameters {’k} (k = 1,…,m) can be obtained via Yule-Walker estimator. Hence estimated {’t} of the noise sequence {t} can be derived, using the equation: yt = ’1yt-1 + ’2yt-2 +… + ’pyt-p + t Step 2: Using {’t}, we can write: yt - ’t = 1’t-1 + 2’t-2 + …+ q’t-q for t = 0,…,N -1.

Durbin’s method (cont.) From which estimated values {’k} of {k} can be obtained through solving the equation system. The order m can be selected via the AIC or BIC. However, a more expedient rule for selecting m is m = 2q. The term Moving Average is historical and should not be confused with the moving average smoothing procedures. TThe term Moving Average is historical and should not be confused with the moving average smoothing procedures. he term Moving Average is historical and should not be confused with the moving average smoothing procedures.

Summary A brief summary of AR and MA models is in order. We have the following properties: - For MA models, the ACF is useful in specifying the order because the ACF cuts off at lag q for an MA(q) series. - For AR models, the PACF is useful in order determination because the PACF cuts off at lag p for an AR(p) process. - An MA series is always stationary, but for an AR series to be stationary, all of its characteristic roots must be less than 1 in modulus. - For a stationary series, the multistep ahead forecasts converge to the mean of the series and the variance of forecast errors converge to the variance of the series.

ARMA Models Many stationary random processes cannot be modeled as purely MA or as purely AR, since they have the qualities of both types of processes. The logical extension of the models MA and AR is the mixed autoregressive – moving average process of order (p, q). We denote this process as ARMA(p,q) and represent it by: yt = 1Yt-1 + 2Yt-2 +… + pYt-p +  + t - 1t-1 - 2t-2 -…- qt-q Why bother with the mixed model? The answer is parsimony: there are fewer parameters to be estimated. Note: In practice, the values of p and q each rarely exceed 2.

Summary on AR, MA and ARMA   ACF PACF AR(p) Die out Cut off after the order p of the process MA(q) Cut off after the order q of the process ARMA(p,q) In this context… “Die out” means “tend to zero gradually” “Cut off” means “disappear” or “is zero”

ARIMA Models In practice, many of the time series we will work with are nonstationary, so that the characteristics of the underlying stochastic process change over time. Now we construct models for those nonstationary series which can be transformed into stationary series by differencing one or more times. We say that yt is homogeneous nonstationary of order d if wt = Δd yt (3.32) is a stationary. Here Δ denotes differencing, i.e., Δyt = yt – yt-1 Δ2yt = Δyt – Δyt-1 After differencing time series to obtain a stationary series wt, we can model wt as an ARMA process.

ARIMA models (cont.) If wt = Δdyt and wt is an ARMA(p,q) process, then we say that yt is an integrated autoregressive-moving average process of order (p,d,q) or simply ARIMA(p, d, q). We can write the equation for the process ARIMA(p,d,q). We can restate the equation for the process ARIMA(p,d,q), using the lag operator (backward shift operator), as: (B)Δdyt = (B)t (3.33) with (B) = 1 - 1B - 2B2 - .. - pBp and (B) = 1 - 1B - 2B2 - … pBq We called (B) the autoregressive operator and (B) the moving average operator. ARIMA models are a class of linear models that is capable of representing stationary as well as nonstationary time series. Note that: ARIMA(p,0,q) = ARMA(p,q)

Estimating and checking ARIMA models We have seen that any homogeneous nonstationary time series can be modeled as an ARIMA process of order (p,d,q). The practical problem is to choose the most appropriate values for p, d, and q , that is, to specify the ARIMA model. This problem is partly resolved by examining both the autocorrelation function (ACF) and the partial auto-correlation function (PACF) for the time series of concern.

After d is determined, we have to find possible values for p and q The process for determining ARIMA model consists of the following steps: Check whether the time series is stationary. If it’s not stationary, determine d the number of times that the series must be differenced to produce a stationary series. After d is determined, we have to find possible values for p and q For MA(q) model, ACF will cut off after the order q of the process while PACF will die out very soon. For AR(p) model, ACF will die out very soon while PACF will cut off after the order q of the process. If both p and q are non-zero, it difficult to determine the exact the orders of AR and MA. Therefore, we apply an iterative approach called Box-Jenkins methodology (1972). This model-building methodology involves a cycle consisting the three stages of model selection (identification), model estimation and model checking.

Box-Jenkins methodology The cycle might have to be repeated several times and at the end, there might be more than one model of the same time series. The Box-Jenkins methodology uses an iterative approach as follows: An initial model is selected, from a general class of ARIMA models, based on an examination of the time series and an examination of its autocorrelations for several time lags The chosen model is then checked against the historical data to see whether it accurately describes the series: the model fits well if the residuals are generally small, randomly distributed, and contain no useful information. If the specified model is not satisfactory, the process is repeated using a new model designed to improve on the original one. Once a satisfactory model is found, it can be used for forecasting.

Notes on model selection (identification) The process of choosing the optimal (p, d, q) in an ARIMA model is known as model selection (or identification). Hannan & Rissanen have suggested the 3-step procedure: 1. determine the maximum length of lag for an AR model. 2. Use AIC (Akaike Information Criterion) to determine the maximum length of lag in an AR model. 3. Use SC (Schwarz Criterion) to determine the maximum length of lags for a mixed ARMA model.

AIC and SC With AIC, k is chosen to minimize where  t is the sum of the squared residuals, p is the maximum degree of ACF and T is the number of observations. With SC, k is chosen to minimize

Estimation of ARMA models Whenever an AR(1) or higher order process is used, a nonlinear estimation procedure is often utilized. This procedure is also an optimization algorithm that attempts to minimize the sum of squared residuals through an iterative procedure. S =  t2 The same situation applies to ARMA models. In using the common nonlinear algorithms, the answer that is obtained may differ depending on the initial guesses for the parameter values. Any nonlinear algorithm could produce an incorrect answer for 2 reasons: It could reach to a local optimum. It could fail to converge at all.

Estimation of ARMA models using software package. Parameter estimation of ARMA models can be automatically performed by sophisticated software packages. In some software packages, the user may have the choice of estimation method and can choose the most appropriate method based on the problem specifications The list of software packages for time series analysis and forecasting: SPSS SAS Minitab R EViews S-Plus

Model Checking After a time-series model has been specified and its parameters have been estimated, one must test whether the original specification was correct. The process of model checking involves two steps. 1. The autocorrelation function for the simulated series can be compared with the sample autocorrelation function of the original series. If the two autocorrelation functions seem very different, one needs to re-specify the model. 2. If the two are not markedly different, one can analyze the residuals of the model. Remember that we have assumed that the random error terms t in the actual process are normally distributed and independent. Then if the model has been specified correctly, the residuals t should resemble a white noise process.

Note: Before using the model for forecasting, it must be checked for adequacy. Basically, a model is adequate if the residuals cannot be used to improve the forecasts, i.e., - The residuals should be random and normally distributed The individual residual autocorrelations should be small. Significant residual autocorrelations at low lags or seasonal lags suggest the model is inadequate

References J. E. Hanke & D. W., Business Forecasting, 8th Edition, Pearson Prentice Hall, 2005. M.K. Evans, Practical Business Forecasting, Blackwell Publishers, 2001. F.X. Diebold, Elements of Forecasting, 4th Edition, Thomson-South-Western, 2007. R. S. Pindyck & D.L. Rubinfield, Econometric Models and Economic Forecasts, 3rd Edition, McGraw Hill, 1991. R. S. Tsay, Analysis of Financial Time Series, 2nd Edition, Willy, 2005.

Appendix: Parameter estimation of AR(p) by Least square method For AR(p) model, the least square method, which starts with the (p+1)th observation, is often used to estimate the parameters. Specifically, conditioning on the first p observations, we have: yt = 1yt-1 + 2yt-2 +…+ pyt-p + at, t = p+1,…,T. which is in the form of the multiple linear regression and can be estimated by the least square method. Denote the estimate of 1by ’1. The fitted model is y’t = ’1yt-1 + ’2yt-2 +…+ ’pyt-p And the associated residual is: a’t = yt – y’t. The series {a’t} is called the residual series, from which we obtain the variance of the series: