Robust exponential smoothing

The paper is devoted to robust modifications of exponential smoothing for time series with outliers or long-tailed distributions. Classical exponential smoothing applied to such time series is sensitive to the presence of outliers or long-tailed distributions and may give inadequate smoothing and forecasting results. First, simple and double exponential smoothing in the L₁ norm (i.e. based on the least absolute deviations) are discussed in detail. Then, general exponential smoothing is made robust, replacing the least squares approach by M-estimation in such a way that the recursive character of the final formulas is preserved. The paper gives simple algorithmic procedures which preserve advantageous features of classical exponential smoothing and, in addition, which are less sensitive to outliers. Robust versions are compared numerically with classical ones.     

Exponential smoothing: The state of the art

This paper is a critical review of exponential smoothing since the original work by Brown and Holt in the 1950s. Exponential smoothing is based on a pragmatic approach to forecasting which is shared in this review. The aim is to develop state-of-the-art guidelines for application of the exponential smoothing methodology. The first part of the paper discusses the class of relatively simple models which rely on the Holt-Winters procedure for seasonal adjustment of the data. Next, we review general exponential smoothing (GES), which uses Fourier functions of time to model seasonality. The research is reviewed according to the following questions. What are the useful properties of these models? What parameters should be used? How should the models be initialized? After the review of model-building, we turn to problems in the maintenance of forecasting systems based on exponential smoothing. Topics in the maintenance area include the use of quality control models to detect bias in the forecast errors, adaptive parameters to improve the response to structural changes in the time series, and two-stage forecasting, whereby we use a model of the errors or some other model of the data to improve our initial forecasts. Some of the major conclusions: the parameter ranges and starting values typically used in practice are arbitrary and may detract from accuracy. The empirical evidence favours Holt's model for trends over that of Brown. A linear trend should be damped at long horizons. The empirical evidence favours the Holt-Winters approach to seasonal data over GES. It is difficult to justify GES in standard form–the equivalent ARIMA model is simpler and more efficient. The cumulative sum of the errors appears to be the most practical forecast monitoring device. There is no evidence that adaptive parameters improve forecast accuracy. In fact, the reverse may be true.     

Testing for (common) stochastic trends in the presence of structural breaks

This paper considers the problem of testing for the presence of stochastic trends in multivariate time series with structural breaks. The breakpoints are assumed to be known. The testing framework is the multivariate locally best invariant test and the common trend test of Nyblom and Harvey (2000). The asymptotic distributions of the test statistics are derived under a specification of the deterministic component which allows for structural breaks. Asymptotic critical values are provided for the case of a single breakpoint. A modified statistic is then proposed, the asymptotic distribution of which is independent of the breakpoint location and belongs to the Cramér‐von Mises family. This modification is particularly advantageous in the case of multiple breakpoints. It is also shown that the asymptotic distributions of the test statistics are unchanged when seasonal dummy variables and/or weakly dependent exogenous regressors are included. Finally, as an example, the tests are applied to UK macroeconomic data and to data on road casualties in Great Britain. Copyright © 2002 John Wiley & Sons, Ltd.     

Some comments on the initialization of exponential smoothing

It is shown that the traditional choice for the initial smoothed statistics in general exponential smoothing leads to the same forecasts as the equivalent ARIMA model, provided that one uses zero starting values for the initial shocks. In addition, an initialization which uses ‘backforecasts’ as initial smoothed statistics is considered, and its relationship to unconditional least squares is explored.     

A new Bayesian formulation for Holt's exponential smoothing

In this paper we propose a Bayesian forecasting approach for Holt's additive exponential smoothing method. Starting from the state space formulation, a formula for the forecast is derived and reduced to a two‐dimensional integration that can be computed numerically in a straightforward way. In contrast to much of the work for exponential smoothing, this method produces the forecast density and, in addition, it considers the initial level and initial trend as part of the parameters to be evaluated. Another contribution of this paper is that we have derived a way to reduce the computation of the maximum likelihood parameter estimation procedure to that of evaluating a two‐dimensional grid, rather than applying a five‐variable optimization procedure. Simulation experiments confirm that both proposed methods give favorable performance compared to other approaches. Copyright © 2008 John Wiley & Sons, Ltd.     

Automatic monitoring of forecast errors

This paper evaluates a variety of automatic monitoring schemes to detect biased forecast errors. Backward cumulative sum (cusum) tracking signals have been recommended in previous research to monitor exponential smoothing models. This research shows that identical performance can be had with much simpler tracking signals. The smoothed-error signal is recommended for α = 0.1, although its performance deteriorates badly as α is increased. For higher α values, the simple cusum signal is recommended. A tracking signal based on the autocorrelation in errors is recommended for forecasting models other than exponential smoothing, with one exception. If the time series has a constant variance, the backward cusum should give better results.     

Prediction intervals for exponential smoothing using two new classes of state space models

Three general classes of state space models are presented, using the single source of error formulation. The first class is the standard linear model with homoscedastic errors, the second retains the linear structure but incorporates a dynamic form of heteroscedasticity, and the third allows for non‐linear structure in the observation equation as well as heteroscedasticity. These three classes provide stochastic models for a wide variety of exponential smoothing methods. We use these classes to provide exact analytic (matrix) expressions for forecast error variances that can be used to construct prediction intervals one or multiple steps ahead. These formulas are reduced to non‐matrix expressions for 15 state space models that underlie the most common exponential smoothing methods. We discuss relationships between our expressions and previous suggestions for finding forecast error variances and prediction intervals for exponential smoothing methods. Simpler approximations are developed for the more complex schemes and their validity examined. The paper concludes with a numerical example using a non‐linear model. Copyright © 2005 John Wiley & Sons, Ltd.     

Exponential smoothing: Estimation by maximum likelihood

In this paper several forecasting methods based on exponential smoothing with an underlying seasonal autoregressive-moving average (SARIMA) model are considered. The relations between the smoothing constants and the coefficients of the autoregressive and moving average polynomials are used. On that basis, a maximum likelihood procedure for parameter estimation is described. The approach rules out the need for initial smoothed values. Prediction intervals are also obtained as a by-product of the approach and a fast algorithm for implementing the method is outlined.     

基于模型突变检测的管道泄漏检测定位方法

为了改进负压波法在管道微弱泄漏的检测和定位中的效果,提出了基于模型突变检测的特征提取和泄漏定位算法。首先运用广义似然比检测理论,建立压力信号变化的统计检验模型,计算广义似然比比率r和下降幅度d作为压力信号突变发生的敏感特征,并通过求得的压力变化时间点进行泄漏定位。针对缓慢泄漏时压力信号变化的特点,提出了相应的改进模型,更精确的求出缓慢泄漏压力信号的下降拐点。实验证明,参数r和d对不同变化类型的压力信号有较好的区分度,利用管道内压力和流量信号的特征值作为神经网络的特征向量可以实现管道的微弱泄漏识别,且基于改进模型的定位算法在缓慢泄漏定位中有较好的定位结果和稳定度。     

Judgmental extrapolation and the salience of change

People may often forecast using cognitive procedures that resemble formal time-series extrapolation models. A model of judgmental extrapolation based on exponential smoothing is proposed in which the setting of the trend parameter is hypothesized to depend upon the relative salience of the successive changes. The salience hypothesis was first tested with exponential series by the use of a framing manipulation. As predicted, focusing the subjects' attention on the changes led to more accurate forecasts. In two investment simulation studies, the salience hypothesis was further examined by varying the statistical properties of the price changes. As predicted, subjects were more likely to sell as prices fell and to buy as prices rose (1) as the sample size of similar changes increased; (2) when the variance of the changes was low; and (3) when the absolute value of the mean change was high. Conditions that may influence judgmental forecasting processes are discussed.     

General exponential smoothing and the equivalent arma process

Although the basic principles of exponential smoothing and discounted least squares are easily understood, the full power of the technique is only rarely exploited. The reason for this failure lies in the complexity of the standard procedures. Often they require fairly complex mathematical models and use a variety of cumbersome algebraic manipulations. An alternative formulation for exponential smoothing is presented. It simplifies these procedures and allows an easier use of the full range of models. This new formulation is obtained by considering the relationship between general exponential smoothing (GES) and the well-known ARMA process of Box and Jenkins. The three commonest seasonal models have only recently been considered for GES systems. They are discussed in some detail here. The computational requirements of the GES and equivalent ARMA procedures are reviewed and some recommendations for their application are made. The initialization of GES forecasting systems and the important problem of model selection is also discussed. A brief illustrative example is given.     

Smooth transition exponential smoothing

Adaptive exponential smoothing methods allow a smoothing parameter to change over time, in order to adapt to changes in the characteristics of the time series. However, these methods have tended to produce unstable forecasts and have performed poorly in empirical studies. This paper presents a new adaptive method, which enables a smoothing parameter to be modelled as a logistic function of a user‐specified variable. The approach is analogous to that used to model the time‐varying parameter in smooth transition models. Using simulated data, we show that the new approach has the potential to outperform existing adaptive methods and constant parameter methods when the estimation and evaluation samples both contain a level shift or both contain an outlier. An empirical study, using the monthly time series from the M3‐Competition, gave encouraging results for the new approach. Copyright © 2004 John Wiley & Sons, Ltd.     

Robust forecasting with exponential and Holt–Winters smoothing

Robust versions of the exponential and Holt–Winters smoothing method for forecasting are presented. They are suitable for forecasting univariate time series in the presence of outliers. The robust exponential and Holt–Winters smoothing methods are presented as recursive updating schemes that apply the standard technique to pre‐cleaned data. Both the update equation and the selection of the smoothing parameters are robustified. A simulation study compares the robust and classical forecasts. The presented method is found to have good forecast performance for time series with and without outliers, as well as for fat‐tailed time series and under model misspecification. The method is illustrated using real data incorporating trend and seasonal effects. Copyright © 2009 John Wiley & Sons, Ltd.     

Towards a taxonomy of forecast error measures a factor-comparative investigation of forecast error dimensions

This study explores the nature of information conveyed by 14 error measures drawn from the literature, using real-life forecasting data from 691 individual product items over six quarterly periods. Principal components analysis is used to derive factor solutions that are subsequently compared for two forecasting methods, a version of Holt's exponential smoothing, and the random walk model (Naive 1). The results reveal four underlying forecast error dimensions that are stable across the two factor solutions. The potentially confounding influence of sales volume on the derived error dimensions is also explored via correlation analysis.     

Monitoring for outliers and level shifts in kalman filter implementations of exponential smoothing

This paper presents a new application of a Kalman filter implementation of exponential smoothing with monitoring for outliers and level shifts. The assumption is that each observation comes from one of three models: steady, outlier, or level shift. This concept was introduced as a multiprocess model by Harrison and Stevens (1976). However, their handling of the models is different. In this paper four different model-selection criteria are introduced and compared by applying them to data. The new features of the application include the four model-selection criteria and the estimation of the required parameters by maximum likelihood.     

Forecasting with leading indicators revisited

Transfer function or distributed lag models are commonly used in forecasting. The stability of a constant‐coefficient transfer function model, however, may become an issue for many economic variables due in part to the recent advance in technology and improvement in efficiency in data collection and processing. In this paper, we propose a simple functional‐coefficient transfer function model that can accommodate the changing environment. A likelihood ratio statistic is used to test the stability of a traditional transfer function model. We investigate the performance of the test statistic in the finite sample case via simulation. Using some well‐known examples, we demonstrate clearly that the proposed functional‐coefficient model can substantially improve the accuracy of out‐of‐sample forecasts. In particular, our simple modification results in a 25% reduction in the mean squared errors of out‐of‐sample one‐step‐ahead forecasts for the gas‐furnace data of Box and Jenkins. Copyright © 2003 John Wiley & Sons, Ltd.     

Bayesian forecasts following a major level change in exponential smoothing

Exponential smoothing techniques enjoy a wide range of applications to problems in signal detection, inventory and production control, financial planning, and many other areas of business and engineering. One of the most useful models used to explain the theoretical structure of the process is a changing levels model, in which the underlying level of the stochastic process is assumed to undergo random changes in each time period. The observation is modelled as a noisy disturbance of this level. Occasionally a major intervention or level change occurs that is much larger than the typical period-to-period fluctuation in the random level. Using a Bayesian approach it is the purpose of this paper to show how the distribution of the major level change can be detected, estimated and then incorporated in forecasts. Updating equations are obtained for the posterior mean and variance of the major level change as well as the new level.     

Forecasting non-seasonal time series with missing observations

Most forecasting methods are based on equally spaced data. In the case of missing observations the methods have to be modified. We have considered three smoothing methods: namely, simple exponential smoothing; double exponential smoothing; and Holt's method. We present a new, unified approach to handle missing data within the smoothing methods. This approach is compared with previously suggested modifications. The comparison is done on 12 real, non-seasonal time series, and shows that the smoothing methods, properly modified, usually perform well if the time series have a moderate number of missing observations.     

The wavelet scaling approach to forecasting: Verification on a large set of Noisy data

In the paper, we undertake a detailed empirical verification of wavelet scaling as a forecasting method through its application to a large set of noisy data. The method consists of two steps. In the first, the data are smoothed with the help of wavelet estimators of stochastic signals based on the idea of scaling, and, in the second, an AR(I)MA model is built on the estimated signal. This procedure is compared with some alternative approaches encompassing exponential smoothing, moving average, AR(I)MA and regularized AR models. Special attention is given to the ways of treating boundary regions in the wavelet signal estimation and to the use of biased, weakly biased and unbiased estimators of the wavelet variance. According to a collection of popular forecast accuracy measures, when applied to noisy time series with a high level of noise, wavelet scaling is able to outperform the other forecasting procedures, although this conclusion applies mainly to longer time series and not uniformly across all the examined accuracy measures.     

Testing cumulative prediction errors in event study methodology

This paper reconsiders event study methodology, a very popular technique in the applied finance literature, within the context of testing cumulative prediction errors. It extends the conventional test statistics in two directions. First, it accounts fully for the increased variance of prediction errors outside of the estimation period and for the cumulation of these errors across different event windows. Second, it also takes account of the fact that market model residuals are typically serially correlated, heteroscedastic and non-normal. The statistics are compared with the conventional approach by reassessing a previous application of the methodology to the impact of management buyouts.