book
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
Poster Communications<br />
SP2<br />
Thursday, September 5th<br />
19:45<br />
Resampling Time Series using the Wavelet<br />
Transform<br />
Aline Edlaine de Medeiros<br />
UEM<br />
Márcia L. A. dos Santos<br />
Eniuce M. de Souza<br />
Brani Vidakovic<br />
Among the techniques that allow the resampling in Time Series is the decomposition from the<br />
Wavelet Transform. This technique is able to rewrite time series in terms of smooth and detail<br />
wavelet coefficients with reduced serial autocorrelation. One of the classical methods of resampling<br />
is the Parallel Bootstrap that uses the Decimated Discrete Wavelet Transform and the parallel<br />
re-sampling of wavelet coefficients to generate replicas of a time series. The advantage of applying<br />
the bootstrap is due to the possibility of estimating parameters as well as confidence intervals even<br />
with small time series. In this paper, we propose to comparethe bootstrap method using a new<br />
resampling methodology based on the Non-decimated Discrete Wavelet Transform (NDWT). Unlike<br />
the decimated transform, the potential of the NDWT is to preserve the same number of wavelet<br />
coefficients at each level of multiresolution. To evaluate the performance of both techniques,<br />
confidence intervals were estimated for simulated time series and also for time series from the<br />
number of Meninginte cases collected from the Brazilian DATASUS system. In addition, both<br />
methods were evaluated for the preservation of the Hurst exponent, a measure used to evaluate<br />
the intensity of autocorrelation inherent to time series, especially Higuchi’s estimation method is<br />
being used due to its robustness. When assessing the Hurst exponent before and after resampling,<br />
it is possible to analyze whether the autocorrection present in the time series was preserved by the<br />
bootstrap. The obtained results indicate that both resampling methods are effective in preserving<br />
the Hurst exponent, as well as generating satisfactory confidence intervals.<br />
Keywords: Non-decimate Wavelet Transform; Bootstrap; Hurst Exponent<br />
118