Sensitivity of Hurst parameter estimation to periodic signals in time series and filtering approaches

Sensitivity of Hurst parameter estimation to periodic signals in time series and filtering approaches D. Marković, M. Koch Department of Geohydraulics and Engineering Hydrology, University of Kassel, Kurt-Woltersstr. 3, 3409 Kassel, Germany (danijela.markovic@uni-kassel.de, kochm@uni-kassel.de) Accepted for publication in Geophysical Research Letters Copyright 2005 American Geophysical Union Further reproduction or electronic distribution is not permitted

Abstract: The influence of the periodic signals in time series on the Hurst parameter estimate is investigated with temporal, spectral and time-scale methods. The Hurst parameter estimates of the simulated periodic time series with a white noise background show a high sensitivity on the signal to noise ratio and for some methods, also on the data length used. The analysis is then carried on to the investigation of extreme monthly river flows of the Elbe River (Dresden) and of the Rhine River (Kaub). Effects of removing the periodic components employing different filtering approaches are discussed and it is shown that such procedures are a prerequisite for an unbiased estimation of H. In summary, our results imply that the first step in a time series long-correlation study should be the separation of the deterministic components from the stochastic ones. Otherwise wrong conclusions concerning possible memory effects may be drawn. KEYWORDS: Hurst parameter, Periodic signals in time-series, Wavelet-Filter, Singular Spectrum Analysis, Rhine river, Elbe river 2

Introduction The history of analyses of long-range-memory time series started with Hurst s (95) fundamental discovery of non-random walk correlations of the ile river discharges. umerous attempts have since then been made to explain the Hurst phenomenon from both the physical and mathematical point of view. One widely used concept is based on a class of models called fractional Gausian noises (fgn) invented by Mandelbrot (965) (cf. Mandelbrot, 999). Unlike that of a short-memory process which possesses an exponential decay, an fgn has an autocorrelation function (ACF ) with a slow, hyperbolic decay of the form ACF (k) H(2H ) k (2 2H) as k, with the Hurst exponent H /2. Thence, H = /2 corresponds to a uncorrelated (white noise) process; for 0 < H < /2 and /2 < H < the process has short- and long-range (long memory or persistence) correlations, respectively (Beran, 994). The Hurst phenomenon can also be evaluated by looking at the power-law dependence of the power spectra S(f) on the frequency of the form S(f) c f f β as f 0, with β related to H as β = 2H (Pelletier, 997). Since Hurst s pioneering work, his phenomenon has been recognized in diverse geophysical time series. Applying spectral analysis to worldwide monthly mean temperature series, monthly mean river discharges from regions of the U.S., tree ring width series in the western U.S., and the precipitation time series from the Vostok ice core database, Pelletier (997) found that all investigated variables, except the precipitation, result in values of H 0.75. Recently, Lohre et al. (2003) detected long-range dependence for the Rhine River, contradicting earlier findings of Mandelbrot and Wallis (969). One of the first physical explanations of the Hurst phenomenon in hydrology has been provided by Eltahir (996) who showed that the change of convergence of atmospheric moisture in the Ethiopian source region of the river ile due to the El iño Southern Oscillation (ESO) is causing a nonstationarity of the mean annual flow of the ile river, confirming also Klemeš s (974) hypothesis on the statistical cause of H > 0.5. Srikanthan et al. (2002) found an ESO-induced two state (dry and wet) persistence in the Australian rainfall time series. Koutsoyiannis (2005) showed that, under the assumption that all time scales are equally import, H may be obtained by the maximum entropy principle. As the idea of persistence of geophysical processes influenced by large scale circulation patterns like that of ESO or of the AO seems to be quite reasonable, the assumption of stationarity of the hydro-meteorological time series should rather be replaced by that of a time-dependent, low frequency mean. In fact, Coels (200) showed that for annual sea levels at Fremantle, that model where the scale parameter of the GEVD is a linear combination of both time and the Southern Oscillation Index leads to a maximized log-likelihood. Moreover, since an inherent property of large scale circulation patterns are signals at low-frequency scales, an important issue is to investigate the influence of the signal-to-noise ratio and of the length of the time series on the Hurst parameter estimate (Ĥ). Hydrologically, the latter explains a scaling structure of a process with statistically significant periodic signals with high signal-to-noise ratio. An example of this is river discharge. amely, the groundwater aquifer acts as a low pass filter for the base flow, removing high frequency components from the input recharge signal. Because of this skin-effect, random forcing and nonlinear groundwater storage are a mechanism for low-frequency amplification of runoff (Shun, 999). In the present paper we investigate how different, recent procedures for the estimation of the Hurst parameter will fare when the time series exhibits an inherent seasonal to in- 3

terdecadal periodicity. A study of the effects of the periodicity in older, mostly heuristic methods and the maximum likelihood Whittle estimator has been done by Montanari et al. (999), while a systematic study of the trend effects on the Detrended Fluctuation Analysis is provided by Hu et al. (200). However, in both studies only synthetic time series with lengths incomparable with the length of real hydrological time series are used (e.g. = 2 7, whereas shorter synthetic (up to 2 0 ) as well as real (discharge) time series are used within this study. It will be shown that there is a significant sensitivity of Ĥ to the sample size and to the signal-to-noise (s/n) ratio for some of the Hurst parameter estimation methods used. We will also investigate the effects of removing periodic components from the synthetic periodic time series for different filtering approaches. Finally, all filters will be applied to the discharge time series and conclusions will be drawn concerning their long-range memory, i.e. persistence. 2 Results 2. Ĥ for simulated fgn and periodic time series For the estimation of the Hurst parameter six methods are used that can be classified as temporal, spectral and time-scale methods, respectively. The temporal methods selected are: () Rescaled Range Analysis (R/S) (Beran, 994), (2) Level of Zero Crossings (ZC) (Coeurjolly, 2000) and (3) Detrended Fluctuation Analysis (DFA) (Hu et al., 200). From the group of spectral methods, the (4) log-periodogram method (LP) (Beran, 994) is used while the (5) Wavelet Transform Modulus Maxima method (WTMM) (Arneodo et al., 995) and the (6) Abry-Veitch (VA) estimator (Veitch and Abry, 999) are chosen from the group of the time-scale methods. For the lower and upper scales of the DFA fluctuation function we used yr. and and /4, respectively. The WTMM estimator is calculated using the continuous wavelet transform with the Morlet wavelet, and the AV estimate with the discrete wavelet transform including the Daubechies 2 wavelet. In an initial exploratory analysis we computed the 95% confidence band (upper and lower 95th percentiles) for each of the named methods using 000 simulated fgn time series sequences with H {, 5,..., 0.9}. The maximal data length of the time series is set to = 2 0 = 024, in order to have a realistic base for a comparison with results of observed geophysical records. The computed mean Ĥ indicated that the R/S and the LP methods are biased due to the under- or overestimation of the simulated H, while the best performance ( Ĥ H 0.05) can be attributed to the ZC and the VA method. The Ĥ obtained with the DFA and the WTMM methods are quite similar, yet with an estimation error of ±0.. For the simulation of the periodic time series we assume that a realization of a timedependent geophysical variable Y (t) (dt = /2 yr.) that is influenced by a large scale circulation pattern (whose low frequency variability can not be neglected) can be written as a superposition of a zero mean stochastic variable X(t) and a sum of k periodic components S i (t), i.e. Y (t) = X(t) + k i= S i(t). The variance σy 2 of Y (t) is equal to σy 2 = σε 2 + 0.5 k i= A2 i, where σ2 ε denotes the variance of the stochastic process X(t) i.e. of the noise and A i are the amplitudes of the periodic components of Y (t). Furthermore, it is assumed that X(t) is a normally distributed ℵ(0, σε) 2 (H = 0.5) and S i (t) is assumed to consist of one or both of the following two components: ) the annual signal with an amplitude A and a 4

period T = yr. and 2) the interdecadal, long-memory signal with an amplitude A 2 and period T 2 = 4 yr. The 4 yr. oscillation has been chosen because its presence was found recently by Marković and Koch (2005) in most of the German precipitation and discharge records. To estimate the concurrent effects of the different frequency components on Ĥ, the periodic time series Y (t) are simulated for two sets of constraints: () the percentage of the variance explained by the annual component (R 2 ) increases from 5 to 40, while the variance explained by the interdecadal component (R 2 2 ) is constant and equal to 5. Therefore, the signal-to-noise ratio defined as s/n = (A + A 2 )/σ ε lies within the range 7.63; (2) the percentage of the variance explained by the interdecadal component is varied from 5 to 40, while that of the annual component is set to zero, resulting in s/n = 0.32.5. For the simulation set () it turns out that, unlike all other applied methods, the DFA and the R/S Ĥ decrease with the increase of the amplitude of the seasonal (annual) component. In Figure the estimations of H for the second simulation set is shown. One observes that a common characteristic of all applied estimators is their sensitivity to the s/n -ratio, namely, Ĥ increases with s/n, i.e. increasing strength of the interdecadal periodic signal. Interestingly, the results of the commonly applied estimators W T MM, R/S and DF A are much more influenced by the time series length than those of AV and ZC which, provide stable, though still upwards biased Ĥ. Consequently, although the simulated time series are practically white noise, i.e. cannot not be characterized by a power low spectra, for s/n > a longmemory behavior would be wrongly concluded by the LP, the AV, and the ZC methods. Whereas for the DF A Ĥ increases with, an opposite behavior is observed for the R/S method. With the W T MM method the situation is even more confusing, since Ĥ increases up to 3 T 2 and then decreases, i.e. the conclusion whether or not a time series has H > 0.5 depends obviously on the time series length used. ote, however, that within the studied range of only for the R/S method the asymptotic value of Ĥ is 0.5, implying that for the samples large enough the Ĥ(R/S) is not influenced by the periodicity of the time series. This is due to the properties of the cumulative range R and of the statistics used in the R/S method. R is defined as the difference between the maximum and the minimum of cumulative departures from the mean (Beran, 994). Because the running mean of a periodic function is also a periodic function of the same period T, but with an amplitude that decays as / the cumulative range R is independent of the mean, i.e. the periodic trend, whenever dt >> T. It follows from the definition of the time series Y (t) above that its theoretical spectrum is a mixed one with a base proportional to the noise variance (2 σε) 2 and discrete peaks at periods T and T 2, equal to the squared amplitude of the corresponding oscillation times /2. As the signal contained in the simulated time series is periodic, so is the decay of the autocorrelation function and, thus, never dies. Therefore, an estimated H > 0.5 is not always a consequence of a power law behavior but may be just an artefact of the bias of the estimation methods when periodic signals are present in the time series. Montanari et al. (999) showed that the likelihood-type Whittle estimator provides a good estimate of H in the presence of periodicity, but information of the spectral density function is needed prior to the application of these methods, implying a turn back to the analysis of the spectrum and dealing with the periodicity. 5

Hest. Hest. Hest. a) LP 200 400 600 800 000 c) AV 200 400 600 800 000 d) ZC 200 400 600 800 000 Hest. Hest. Hest. b) WTMM 200 400 600 800 000 d) RS 200 400 600 800 000 f) DFA 200 400 600 800 000 Figure : H estimates as a function of the time series length for different s/n-ratios, s/n = 0.32 (dotted line), s/n = 0.7 (solid line) and s/n =.5 (dashed line), corresponding to an increased percentage of the interdecadal component (simulation set (2), see text). 2.2 Effect of the seasonal adjustment and low frequency filtering on the estimation of H Periodic signals in geophysical time series usually have strongly time-dependent amplitudes and lags between the underlying forcing signals. Choosing then a suitable filtering approach is not straightforward. To demonstrate the effects of the seasonal adjustment and of the low frequency filtering approach on Ĥ we employ again the simulated periodic time series of the previous section and, additionally, long-range (900-2002), extreme monthly discharge time series at one Rhine River gauge (Kaub) and one Elbe River gauge (Dresden). As a matter of fact, Lohre et al. (2003) set up a seasonal long-memory model for Kaub and showed that the persistence remains present after removing seasonal means from the data. Our waveletreanalysis of this data set shows that the long memory of the de-seasonalized data is not surprising since, additionally to the annual cycle, significant low frequency oscillations at 4.3 yr. and 3.9 yr. scales are present in the spectrum. Similar results are obtained for the Elbe River flow at Dresden, with low frequency oscillations at 6.9 yr. and 3.9 yr. scales. It is conjectured that these two long-term fluctuations are due to a teleconnective influence of ESO and AO, respectively. The time series filtering is performed by means of () a least squares linear regression (LSR), (2) moving average seasonal adjustment (MA) (Priestley, 98) (3) the wavelet filter (Torrence and Compo, 998) and (4) Singular Spectrum Analysis (SSA) (Golyandina et al., 200). As the regressors in LSR () we used the statistically significant periodic signals from the global wavelet spectrum. For the MA filter (2) the seasonal length s is defined as s = 2r with r = 6 months. The wavelet filter () is set to react only if the real part of the wavelet coefficients W nr (s) at the particular scale s and time t is greater or less than : X t = 0.08333/2 0.776 π /4 j=j 2 sgn[w nr(s j )][ W nr(s j ) ] j=j s /2 j. () 6

Table : Ĥ of the simulated periodic time series for the different estimators used applying three different filtering approaches a) Data LP W T MM DFA RS AV ZC R 0.56 3 8 9 0.73 0.77 SA 0.54 7 7 0.55 0.56 0.56 SA,LSRF 0.53 0.5 0.52 0.55 0.50 0.53 SA,WF 0.53 0.54 0.58 0.5 0.50 0.53 a) Abbreviations and specifications for the methods (see text) used: R- raw data (σε 2 = 0.55, = 2 0, R 2 =, R 2 2 = 0.05); SA, seasonally adjusted (MA); SA, LSRF- seasonally adjusted (MA) and low frequency detrended using the LSR; SA, WF- seasonally adjusted (MA) and the low frequency detrended using the wavelet filter (eq. ). which has the effect of removing the signal within the selected scale band (j j 2 ) and reducing the bias effects caused by removing the power which belong to the noise. The SSA method (4) consist of four steps for the time series decomposition, i.e. the increase of the s/n-ratio: ) embedding lagged vectors, 2) singular value decomposition (SVD), 3) grouping and 4) diagonal averaging. In the grouping stage 3) the periodic components are detected by analyzing pairwise relationships between the eigenvectors obtained by the SVD. Since the SSA-method does not enable automated separation of periodic components it will be applied only on the real data. 2.2. Results for simulated periodic time series In Table we list the mean Ĥ from 000 simulated periodic time series obtained with the six estimation methods defined earlier and applying the three filter techniques. The results of Table show that, whereas the R/S method provides Ĥ = 9 (dt >> T ), the DFA, AV and the ZC method clearly point to the long-range correlations in the raw data. Further, one notes that Ĥ(DFA) significantly increases after seasonal adjustment to 7. This can be explained by the dominating influence of the low frequency signal on the scaling behavior of the DFA-fluctuation function (Hu et al. 200). After removal of both periodicities Ĥ(DFA) indicates uncorrelated noise, too. We also note that the regression filter LSRF works better than the wavelet filter WF when the signal amplitude is constant, as manifested by a smaller Ĥ(DF A) = 0.52 for LSR instead of Ĥ(DF A) = 0.58 for WF. 2.2.2 Results for Rhine and Elbe river flow time series Unlike the extreme monthly river flows at the Elbe River gauge Dresden, those at the Rhine River gauge Kaub exhibit a statistically significant increasing linear trend with a slope coefficient equal to 0.33 (p = 5.5 0 6 ). Both data sets possess a pronounced annual cycle and the interannual to interdecadal quasiperiodic cycles. In Table 2 we list Ĥ obtained with the various estimation methods for the two discharge series employing () the raw data, (2) detrended, (3) detrended, seasonally adjusted and wavelet filtered data (4) detrended, seasonally adjusted, wavelet- and AR() filtered data 7

400 300 200 Q [m 3 /s] 00 0 00 200 300 900 920 940 960 980 2000 t [year] Figure 2: Low frequency component of the extreme monthly Elbe River flows at Dresden applying the wavelet filter (solid line) and the SSA filter (dashed line). and (5) SSA and AR() filtered data. Except for the R/S and the LP method (see Table 2), all other methods are insensitive to removals of trends from the data. For the scale methods as well as the DFA-method, this is because of their ability to filter out polynomial trends. For the ZC method the reason is obvious since the specified level of crossings is either the time series mean or the trend. The results based on the wavelet and the SSA filter do not vary significantly as illustrated in Figure 2 for Elbe River flows. Except for R/S at gauge Kaub, the filtering procedure reduces Ĥ for all other methods. However, some of them indicate long-range correlations in the residuals and some others noise or short-range correlations. Additional check of the periodogram of the time series residuals upon substraction of the linear and periodic trends has confirmed the absence of the significant spectral peaks while the partial ACF of the residuals points to a red-noise time series background. Thus, in the next step we fitted the AR() model to the time series stochastic component and then estimated the H value of the residuals. Except for ZC and R/S at station Kaub, all other methods do not support the hypothesis of long memory in the residuals. Further, since ZC turned out to have the largest Ĥ for a simulated periodic time series with a red noise background (not shown) and the R/S estimate is also sensitive to the AR() correlations, we conclude that the idea of a long memory structure has to be given up in favor of a short memory one. 3 Conclusions The influence of the periodic signals in time series on the Hurst parameter estimate (Ĥ) is investigated for the temporal (R/S, ZC and DFA), spectral (LP) and time-scale methods (WTMM and VA). Using simulated periodic time series with a white noise background we show that Ĥ depends clearly on the s/n-ratio and, for some methods, also on the data length used. Unlike all other methods, the DFA- and the RS- estimates of H decrease with an increase of the variance of the annual signal. On the other hand, with periodic signals at the interdecadal scale (4 yr.) present Ĥ indicate persistence and increases with the variance of that component for all methods. Ĥ of the monthly extreme discharge time series of the Elbe River gauge Dresden and of the Rhine River gauge Kaub indicate power law decay of the ACF (Ĥ > 0.5) when the raw 8

Table 2: Ĥ of the monthly extreme flow data a) Gauge LP W T MM DFA RS AV ZC 0.72 0.73 3 7 3 9 0.73 3 0.76 3 0.59 0.50 0.74 0.73 0.77 0.54 5 5 0.7 0.50 4 Kaub 0.5 6 3 7 0.52 3 9 0.75 9 0.57 0.78 3 9 0.75 9 0.52 0.78 3 0.58 0.58 0.52 9 4 0.75 0.50 0.5 0.50 3 9 5 Dresden 0.54 5 4 0.50 9 8 a) First row of each gauge set denotes Ĥ of the raw data; second raw the Ĥ after substraction of the linear trend; third raw the Ĥ after substraction of the linear trend, seasonal (MA) and low frequency cycles (eq. ); fourth raw Ĥ after substraction of the linear trend, seasonal cycle (MA), low frequency cycles (eq. ) and the fitted red noise; fifth row Ĥ after SSA and AR() filtering. data is processed. However, when applying detrending, seasonal adjustment (MA method) and the low frequency filtering (wavelet and SSA filter) smaller Ĥ are obtained. This shows that the correlations depend only on the lag, i.e. the theoretical spectrum of the residuals is of red noise, further confirmed by fitting an AR() model. We remark that for the filtering of the periodic signals with time-varying amplitudes like those in the discharge time series analyzed, the wavelet filter is advantageous over the conventional regression filter. Our results imply that prior to drawing a conclusion on the time series long-correlation structure, one should attempt to distinguish between the time series deterministic memory (e.g. trends, periodicities), nondeterministic periodicity and the correlation properties of its stochastic components. Acknowledgments: We would like to thank the German Federal Agency for Hydrology (BAfG) for providing the discharge time series. We are grateful to Alberto Montanari and the anonymous reviewer for constructive comments and suggestions. References [] Arneodo A., Bacry E., Graves P.V., Muzy J.F., Characterizing long-range correlations in DA sequences from wavelet analysis, Phys. Rev. Lett., 74(6), 3293-3296., 995. [2] Beran J., Statistics for long-memory processes, Chapman & Hall, 994. [3] Coeurjolly J.-F., Simulation and identification of the Fractional brownian motion: a bibliographical and comparative study, J. Stat. Soft. 5(7), 2000. [4] Coles S., An Introduction to Statistical Modeling of Extreme Values, Springer, 200. 9

[5] Eltahir E. A. B.,El iño and the natural variability in the flow of the ile River, Water Resour. Res., 32(), 3-37, 996. [6] Feder J., Fractals, Plenum Press, ew York, 988. [7] Golyandina., ekrutkin V., Zhigljavsky., Analysis of Time Series Structure: SSA and related techniques, Chapman & Hall/CRC, London, ew York, 200. [8] Hu K., Ivanov P. Chen Z., Carpena P., Stanley H. E., Effect of trends on detrended fluctuation analysis, Phys. Rev. E 64, 04, 200. [9] Klemeš V., The Hurst Phenomena: A puzzle, Water Resour. Res., 0(4), 675-688, 974. [0] Koutsoyiannis, D., Uncertainty, entropy, scaling and hydrological stochastics, 2, Time dependence of hydrological processes and time scaling, Hydrological Sciences Journal, 50(3), 405-426, 2005. [] Lohre M., Sibbertsen P., Könning T, Modelling water flow of the Rhine River using seasonal long memory, Water Resour. Res, 39(5), 32, 2003. [2] Mandelbrot B. B., Multifractals and /f noise, Springer, 999. [3] Mandelbrot, B. B., Wallis J. R., Some long-run properties of geophysical records, Water Resour. Res., 5(2), 32340, 969. [4] Marković, D., Koch M., Wavelet and scaling analysis of monthly precipitation extremes in Germany in the 20 th century: Interannual to interdecadal oscillations and the orth Atlantic Oscillation influence, Water Resour. Res., 2005 (in press). [5] Montanari A., Taqqu S., Teverovsky V., Estimating Long-Range Dependence in the Presence of Periodicity: An empirical Study, Math. and Comp. Mod., 29, 27-228, 999. [6] Pelletier J. D., Fractal models in Geology, PhD dissertation, Cornel University, 997. [7] Priestley M. B., Spectral analysis and time series, Academic Press, 98. [8] Shun T., Duffy C. J., Low-frequency oscillations in precipitation, temperature, and runoff on a west facing mountain front: A hydrogeologic interpretation, Water Resour. Res., 35(), 9-20, 999. [9] Srikanthan R., Thyer M. A., Kuczera G., McMahon T. A., Modelling Annual Rainfall Using a Hidden-State Markov Model, iemss proceedings, 2002. [20] Torrence C., Compo G. P., A practical guide to wavelet analysis, Bull. Amer. Met. Soc., Vol. 79, o., 62-78, 998. [2] Veitch D., Abry P., A wavelet-based joint estimator of the parameters of long-range dependence, IEEE Trans. Inf. Theory 45, 878897, 999. 0