A seasonal periodic long memory model for monthly river flows

https://doi.org/10.1016/S1364-8152(01)00025-1Get rights and content

Abstract

Based on simple time series plots and periodic sample autocorrelations, we document that monthly river flow data displays long memory, in addition to pronounced seasonality. In fact, it appears that the long memory characteristics vary with the season. To describe these two properties jointly, we propose a seasonal periodic long memory model and fit it to the well-known Fraser river data (to be obtained from Statlib at http://lib.stat.cm.edu/datasets/). We provide a statistical analysis and provide impulse response functions to show that shocks in certain months of the year have a longer lasting impact than those in other months.

Introduction

It is well known since the early work by Hurst on Nile data that river flows show long, persistent changes in mean which may be characterized by long memory. Additional to long memory, most river flow data display pronounced seasonality, both in mean and in variance.

Lawrance and Kottegada (1977) presented an overview of early results on the statistical modeling of river flows. One of the main objectives of these modeling efforts is to develop simulation models which can be used for the design and operation of reservoirs. Brochu (1978) noticed that even modest improvements in the operation of large reservoir systems can result in multi-million dollar savings per year. The time series analysis of river flow data has remained an innovating research area. Novel statistical models for simulation, forecasting and diagnostic analysis have been introduced for river flow data and other new methods have been tried on river flow data soon after their introduction, see e.g. McLeod and Hipel (1978) and Hipel and McLeod, 1978a, Hipel and McLeod, 1978b.

Noakes et al. (1988) compared the one-year-ahead forecasting ability of ARMA models, fractional Gaussian noise models, Fractional ARMA models, and Markov and nonparametric regression models for yearly river flow series, analysed earlier by Hipel and McLeod, 1978a, Hipel and McLeod, 1978b. They showed that it is hard to find significant differences between the models, but the simple fractional models seemed too restrictive for the four series they studied.

The evidence on the adequacy of statistical models for seasonality is much clearer. Not only the mean and variance of monthly river flows depend on the season. Other characteristics like skewness and autocorrelation do as well, as shown by Moss and Bryson (1974).

The skewness is usually taken care of by an a priori log transformation of the series and the seasonally dependent autocorrelations are successfully modeled using periodic autoregressive moving average (PARMA) models, see e.g. Vecchia and Ballerini (1991). Periodic autoregressive (PAR) models have definite computational advantages over PARMA models. PAR models are easy to identify using periodic partial autocorrelation functions, they are easy to estimate using least squares, from the Yule–Walker equations or by Maximum Likelihood, see e.g. McLeod (1994). PAR models for monthly riverflow modeling and simulation were originally introduced by Thomas and Fiering (1962).

Noakes et al. (1985) compared the short-term forecasting ability of seasonal ARIMA, deseasonalized ARMA and periodic autoregressive (PAR) models on 30 monthly river flow series. The results clearly suggest that periodic autoregressive models identified by the partial autocorrelation function, provided the most accurate forecasts. They also established the superiority of the natural log transformation over other Box–Cox transformations, Box and Cox (1964), in a classical likelihood framework.

Although the explicit modeling of long range dependence may not be too useful for point-forecasting, especially if the process is stationary, it can still be important for confidence interval forecasting, see e.g. Ray (1993). It also plays a deciding role in hypothesis testing and in the development of simulation models. It is, for example, important to take account of long range dependence if one wants to do inference about the (seasonal) long run mean of a process, as emphasized by Beran (1994) in the first chapter of his monograph on “Statistics for Long-Memory Processes”. Neglecting long-range dependence may result in gross downward biases in estimates of the uncertainty about the mean. This is especially relevant if one wants to test for structural stability of the correlation structure and the mean, where proper estimation of the variance of the (sub)sample means is crucial.

Beran and Terrin (1996) reanalyzed yearly minimum water levels for the river Nile (622-1281) using an ARFIMA(0,d,0) model and found significant changes in the correlation structure over time. The analysis of structural change in long geophysical time series is particularly interesting for climate research. Atkinson et al. (1997) used recent structural time series for the annual flow of the Nile (1871–1970) to illustrate new tests for structural breaks and found that the process could well be described by white noise allowing for a couple of additive outliers and a structural break due to the building of the Aswan dam in 1899. MacNeill et al. (1991) surveyed earlier analyses of those Nile data.

A class of models, which seems to have been overlooked in the literature on river flow modeling is the class of seasonal long memory models, introduced by Carlin et al. (1985) for economic time series. These models describe long range dependence in the seasonal pattern time series, and focus on the correlation structure at yearly intervals. Our model focuses on this aspect as well. We discuss the relationship between our model and other seasonal long memory models in more detail in Section 3.1 below. In our model we combine seasonal long memory allowing for the now well established periodic variation in the autocovariance function. The combination of these two features may explain both the long memory apparent in yearly series of minima of river flows, the so-called Joseph effect, and the absence of long memory in aggregate yearly river flow data. In this paper we specify and estimate such a seasonal periodic model for monthly river flows.

Droughts and floods are phenomena that are typical for special seasons of the year. If we look specifically at autocorrelations for data of a specific month we may notice the long nonperiodic cycles, whereas we overlook them if we aggregate flow data over the year. This is of course more likely to happen if the seasonal long range dependence occurs for months with relatively small flows.

There seems to be a misunderstanding among practitioners that seasonal (fractional) differencing as applied in seasonal AR(F)IMA modeling, and periodic modeling, as in PAR models, are substitutes for describing seasonal phenomena. Our application shows they are complements: seasonal parameters can be periodic as well!

The specification of our model does not involve new statistical problems. The models can easily be estimated using existing software for ARFIMA analysis. We basically extend the periodic AR(3) model of McLeod (1994), introducing error terms which display seasonal fractional integration which varies from month to month. For application we consider the monthly Fraser river flow data at Hope, B.C., made available by Ian McLeod at Statlib. The current URL is http://lib.stat.cmu.edu/datasets/.

We show how we can capture the interesting long memory characteristic, which appears evident from the periodic autocorrelation functions at long yearly lags. Statistical analysis shows seasonal long memory to be significant, especially for the spring and the autumn months. Our statistical analysis provides an additional test on model adequacy and can be used as a parametric complement to the residual serial correlation tests for periodic models developed by Vecchia and Ballerini (1991), McLeod (1994) and Franses (1996) and residual correlation tests for long memory models by Beran (1992) and Robinson (1994).

The outline of our paper is as follows. In Section 2 we present some characteristics of the monthly Fraser river flow data. In Section 3 we propose the novel seasonal periodic long memory model, compare it with related models and discuss estimation issues and available software. Section 4 provides the empirical analysis and Section 5 concludes.

Section snippets

Data and memory characteristics

Let yt denote a monthly time series, t=1,2,⋯, n, where yt concerns log-transformed data of the monthly mean river flows in cubic feet per second, following the analyses in Vecchia and Ballerini (1991) and McLeod (1994). Let Ym,T denote these observations, m=1,2,3,⋯,12 and T=1,2,⋯, N, so that m denotes the number of the month and T denotes the number of the year. We have N years of subsequent observations with monthly data. To simplify notation we only use complete years with observations

A seasonal periodic long memory model

The basic idea of our model is very simple. It is a standard Seasonal Autoregressive Integrated Moving Average (p,0,0)×(P,D,Q)12 model (SARIMA), where the integration parameter d and the MA order for monthly lags, q are zero and where the other parameters of the model, i.e. the seasonal AR and MA parameters, the seasonal integration parameter D and the AR parameters for the monthly lags are allowed to vary from month to month. Furthermore, we allow for fractional D, so that the seasonal pattern

Empirical results

We first check McLeod's periodic AR(3) model using the familiar diagnostic regression tests implemented in PcGive 9.0, Hendry and Doornik (1996). We simply regressed the April series on a constant and the series for March, February and January. The May series were also regressed on the three previous months and so forth. These regression showed no substantial problems with the model specification at yearly lags, with regard to nonnormality and short run serial correlation. McLeod (1994) already

Conclusion

We extended the periodic AR model for the monthly data for the logs of the Fraser river flow, developed by McLeod (1994) with a periodic seasonal long memory innovation proces. The statistical model detected the presence of long non-periodic cycles which are especially clear in sample correlations for the data for the month of March. We estimated the model using Gaussian Maximum Likelihood. If the long memory is not too pronounced and the MA-part of the model is clearly invertible one can use

References (46)

  • J. Beran et al.

    Testing for a change of the long-memory parameter

    Biometrika

    (1996)
  • G.E.P. Box et al.

    An analysis of transformations

    Journal of the Royal Statistical Society, Series B

    (1964)
  • M. Brochu

    Computerized rivers

    Science Dimension

    (1978)
  • R. Dahlhaus

    Efficient location and regression estimation for long range dependent regression models

    The Annals of Statistics

    (1995)
  • D.A. Dickey et al.

    Distribution of the estimators for autoregressive time series with a unit root

    Journal of the American Statistical Association

    (1979)
  • D.A. Dickey et al.

    Testing for unit roots in seasonal time series

    Journal of the American Statistical Association

    (1984)
  • J.A. Doornik et al.

    Give Win: An Interface to Empirical Modelling

    (1999)
  • J.A. Doornik

    Object-oriented Matrix Programming using Ox version 2.0

    (1998)
  • Doornik, J.A., Ooms, M., 1999. A Package for Estimating, Forecasting and Simulating Arfima Models, Arfima Package 1.0...
  • P.H. Franses

    Recent advances in modelling seasonality

    Journal of Economic Surveys

    (1996)
  • P.H. Franses et al.

    A periodic long memory models for quarterly UK inflation

    International Journal of Forecasting

    (1997)
  • C.W.J. Granger et al.

    An introduction to long-memory time series models and fractional differencing

    Journal of Time Series Analysis

    (1980)
  • W.H. Greene

    Econometric Analysis

    (1997)
  • Cited by (35)

    • Streamflow forecasting: Overview of advances in data-driven techniques

      2021, Advances in Streamflow Forecasting: From Traditional to Modern Approaches
    • Short-term air temperature forecasting using Nonparametric Functional Data Analysis and SARMA models

      2019, Environmental Modelling and Software
      Citation Excerpt :

      Tabari et al. (2015) used air temperature forecasts obtained by Artificial Neural Networks (ANN) to predict soil temperature. Seasonal long memory models have been commonly adopted for the modelling of climatic, hydrological and environmental time series (see for instance Ooms and Franses, 2001 and Reisen et al., 2014). Grimaldi (2004) showed that fractional ARMA (fARMA) models are more appropriate for time series that exhibit long term memory compared to the standard short memory models.

    • Forecasting daily river flows using nonlinear time series models

      2015, Journal of Hydrology
      Citation Excerpt :

      This approach is used for building mathematical models to generate synthetic, hydrological records and to forecast hydrological events (Salas, 1993). A variety of linear AR models are used to forecast streamflows (McLeod et al., 1977; Lu et al., 1996; Abrahart and See, 2000; Montanari et al., 2000; Ooms and Franses, 2001). Some nonlinear AR models are also applied to streamflow forecasting (Tong and Lim, 1980; Astatkie et al., 1997).

    • Improving estimation of the fractionally differencing parameter in the SARFIMA model using tapered periodogram

      2015, Economic Modelling
      Citation Excerpt :

      Then many works followed to discuss the applications, properties, estimation of the SARFIMA model. Previous empirical studies include Ray (1993) for the series of monthly IBM product revenues, Ooms and Franses (2001) for the monthly river flows, Koopman et al. (2007) for a combination of SARFIMA and GARCH for electricity price and so forth. In frequency domain, the spectrum poles of an SARFIMA process are displayed not only near zero frequency, but also at certain seasonal frequencies (see Reisen et al. (2006)).

    • A semiparametric approach to estimate two seasonal fractional parameters in the SARFIMA model

      2014, Mathematics and Computers in Simulation
      Citation Excerpt :

      Ray [25] used IBM product revenues to illustrate the usefulness of modeling seasonal fractionally differenced ARMA models by allowing two seasonal fractional differencing parameters in the model, one at lag 3 and the other at lag 12. Other papers related to this topic are, for example, Hassler [14], Gray et al. [11], Gray et al. [12], Giraitis and Leipus [9], Woodward et al. [35], Ooms and Franses [21], Arteche and Robinson [3], Palma and Chan [23], Reisen et al. [29,28], Gil-Alana [8], among others. Hassler [14] introduced the rigid and flexible seasonal filters and an application of this methodology to the economic activities in the Euro area is discussed in Guegan [5].

    • The application of a general time series model to floodplain fisheries in the Amazon

      2013, Environmental Modelling and Software
      Citation Excerpt :

      In the context of fisheries, ARIMA models are considered efficient tools for the analysis of the pilchard fishery (Koutroumanidis et al., 2006; Stergiou, 1989; Stergiou and Christou, 1996; Stergiou et al., 1997), and are most commonly used to make a small number (short-term) of future predictions (Stergiou, 1989; Lloret et al., 2000; Parsons and Colbourne, 2000). Extensions of ARIMA models for series whose correlation function decays slowly as a function of the lag distance (long memory time series) have been used to study monthly river flows (Ooms and Franses, 2001). More recently, recursive algorithms that produce optimal estimates (e.g. the Kalman filter) have been used when there is large natural variability and measurement error in fisheries data (Peterman et al., 2003).

    View all citing articles on Scopus
    View full text