Many economic time series feature time-varying characteristics such as non-stationarity, volatility, seasonality, and structural discontinuities. Wavelet analysis is a natural framework for analyzing these phenomena without imposing simplifying assumptions such as stationarity. In particular, wavelet filters can decompose and reconstruct a time series, as well as its correlation structure, across time scales.

While wavelets have applications in diverse areas such as such as regression, unit root testing, bootstrapping (wavestrapping), and fractional integration order estimation, EViews focuses on four popular areas of application:

Additional discussion of wavelet analysis is provided in “Technical Background”.

To perform wavelet analysis in EViews, open the series of interest and click on View/Wavelet Analysis. There are four sub-menu entries corresponding to the four areas of application. We consider each in turn.

Typically, the first step in wavelet analysis is to perform a wavelet transform of some input series (discrete wavelet transform analysis is described in “Discrete Wavelet Transform”).

Click on View/Wavelet Analysis/Transforms... to display the wavelet transform analysis dialog:

The first step is to choose a discrete wavelet decomposition type.

Discussion is provided in “Discrete Wavelet Transform” and “Multiresolution Analysis”.

Next you may specify scale options:

Lastly, you may use the Filter group settings to specify properties of the filter

As a first example, consider the Canadian real exchange rate data extracted from the data set in Pesaran (2007) in “wavelets.WF1”. This is a quarterly series running from 1973q1 to 1998q4. We will use the default settings in the dialog. To display the transform, open the CANADA_RER series and select from the series window, click on OK.

Note that since the number of data points is 104, a dyadic adjustment is needed and the default setting pads the input series using the series mean is used to achieve dyadic length.

The output is a spool object with the spool tree listing the summary, original series, as well as wavelet and scaling coefficients for scales 1 through 7. The first of these is a summary of the wavelet transformation performed.

The second spool entry is a plot of the original series along with the padded values.

The remaining entries in the spool are wavelet and scaling coefficients plotted for each scale. Scroll down or click on an entry to bring it into view:

Here, we see the wavelet coefficients at scales 1 and 2. Notice that standard deviation of the coefficients at each plot has two dashed red lines. These represent the +/– standard deviation of coefficients at that scale. This is particularly useful in visualizing which wavelet coefficients should be shrunk to zero in wavelet shrinkage applications. Recall that coefficients exceeding some threshold bound (in this case the standard deviation) ought to be retained, whereas the remaining coefficients ought to be shrunk to zero.

Continuing to higher scales, we see much smaller relative values

Looking at the wavelet coefficients at each scale, it is clear that the majority of relevant information can be captured only in the first few scales and using very few coefficients. For instance, at scale 1 we see roughly 3 or 4 wavelet coefficients which exceed the threshold bounds. Those are the coefficients which will contribute most to explaining variation at that scale and will capture the most transitory (noise) features of the underlying series. At scale 2, there are 6 coefficients which exceed the threshold bounds. While these coefficients are in the lower half of the frequency range, they are not in the immediate vicinity of frequency 0. We can conclude that this is a series which has persistence, but is not necessarily non-stationary. A quick unit root test will also confirm this belief.

Finally, notice that the time-axis in the DWT is not actually aligned with time, but rather the number of observations. This offset occurs because the filters used are in fact not zero-phase filters and therefore their transformation will not be aligned with time.

For the second example, use the input series from the first example and perform a MODWT with the Daubechies filter of length 6. From the series window, click on View/Wavelet Analaysis/Transforms... to display the dialog.

Then in the Method dropdown select Overlap transform - MODWT. In the Class dropdown select Daubechies and in the Length dropdown select 6. Click on OK.

As before, the output is a spool object with wavelet and scaling coefficients as individual spool elements.

Note that since the MODWT is not an orthonormal transform and since it uses all of the available observations, wavelet and scaling coefficients are of input series length and do not require length adjustments.

Scrolling down or clicking on Wavelet 1 in the tree to bring the graph into view displays:

Next, notice that for each scale those coefficients affected by the boundary are displayed in red, and their count reported in the legends. Moreover, there is a dashed vertical line indicating the end of the boundary coefficients. Furthermore, since the MODWT is a redundant transform, the number of boundary coefficients standard deviation bounds will always be greater than those in the orthonormal DWT. As before the are available for reference. Finally, note that the time-axis in the MODWT is aligned with time. This is an advantage of the MODWT over the DWT at the expense of redundancy.

As in the DWT example, it's clear that the majority of relevant wavelet coefficients (those exceeding threshold bounds) are concentrated in the first few scales. We also observe that there are significantly more such coefficients in the MODWT transformation as opposed to the DWT since the MODWT is a redundant transformation and wavelet coefficients across scales are correlated. This behavior is also seen in the wave-like patterns in the output

Wavelet variance decomposition is described in “Wavelet Analysis of Variance”. To perform wavelet decomposition of variance in EViews, open a series window and proceed to View/Wavelet Analysis/Variance Decomposition... to display the dialog.

The Variance group settings control the computation of the variances and confidence intervals:

When the variance decomposition is unbiased the maximum filter scale is adjusted automatically to reflect dropping boundary coefficients.

If the decomposition is unbiased, one can also form confidence intervals for per-scale variance contributions.

Lastly, the Decomposition and Filter group elements were discussed in detail in “Wavelet Transforms”.

For this example, we use Japanese real exchange rate data from 1973Q1 to 1988Q4 extracted from the data set in Pesaran (2007). The objective is to produce a scale-by-scale decomposition of variance contributions using the MODWT with a Daubechies filter of length 4. Furthermore, we’ll produce a 95%confidence intervals using the asymptotic Chi-squared distribution using a band-pass estimate for the EDOF. The band-pass EDOF is preferred here since the sample size is less than 128 and the asymptotic approximation to the EDOF requires a sample size of at least 128 for decent results.

To proceed, open the series window and click on View/Wavelet Analysis/Variance Decomposition... First, set the CI type dropdown to Asymp. Band-Limited. Next, set the dropdown in the Decomposition group to Overlap transform - MODWT. Lastly, change the Filter group Class dropdown to Daubechies and click on OK to display the results.

The output is a spool object with the spool tree listing the procedure summary, energy spectrum, variance distribution across-scales, CIs across scales, and the cumulative variance.

The first spool entry is a summary of the wavelet variance decomposition performed including information about the decomposition and filter, sample, and the confidence intervals settings.

The second entry is a table of the variance spectrum listing the contribution to overall variance by wavelet coefficients at each scale. The column titled “Variance” shows the variance contributed to the total at a given scale. Columns titled “Relative Proportion” and “Cumulative Proportion” display the proportion of overall variance contributing to the total at a given scale and its cumulative total, respectively.

Here, since confidence intervals have been produced, the last two columns display the lower and upper confidence interval values at each scale.

We may scroll down or click on the Variance Distribution selector to view the next items in the spool:

The next object in the spool is the Variance Spectrum Distribution graph, which is a bar plot of variances at each given scale.

Following this is the Variance CI Distribution graph, which shows the variance values along with their 95% confidence intervals at each scale.

Scrolling down to the last graph or clicking on Cumulative Variance, we see a display showing the cumulative variances across scales and, if computed, the cumulative confidence interval.

As in the previous wavelet examples, wavelet thresholding in EViews is performed on a series. In particular, from an open series window, click on View/Wavelet Analysis/Thresholding (Denoising)... to display the dialog.

The Threshold group options control the specification of the threshold:

The Wavelet coefficient variance group specifies the method for computing the estimate of the variance (standard deviation) of the wavelet coefficients

The remaining Decomposition and Filter group elements were discussed in detail in “Wavelet Transforms”.

For our example, we use the French real exchange rate data from 1973q1 to 1988q4 extracted from the data set in Pesaran (2007). The objective is to de-noise the input function by extracting its signal using the DWT with a Daubechies filter of length 6. We will employ a hard thresholding rule using a false discovery rate threshold with significance level 05. The wavelet variances will be estimated using the median with a Guassian correction.

To proceed, open the series window and click on View/Wavelet Analysis/Denoising (Thresholding)...

Under the Threshold group, set the Method dropdown to Hard and set the Limit dropdown to False Discovery Rate. Next, Under the Wavelet coefficient variance group set the Method dropdown to Median (Gaussian). Lastly, change the Class dropdown to Daubechies and set the Length dropdown to 6. Click on OK.

EViews displays spool output with a tree listing the summary, denoised function, and noise.

Finally, the remaining plot is the noise process extracted from the original series.

At the top of the spool is a summary of the denoising procedure describing the decomposition method and filter, sample information, threshold method, and wavelet coefficient variance computation.

Below the summary information is a plot of the denoised function (signal) superimposed over the original series.

Scrolling down or clicking on the Noise entry on the left-hand side of the spool brings the graph of the noise (difference) into view.

Wavelet outlier detection is described in “Wavelet Outlier Detection”. To perform wavelet outlier detection in EViews, open a series window and proceed to View/Wavelet Analysis/Outlier Detection... to display the dialog.

The dialog is identical to the Wavelet Thresholding dialog described in

To demonstrate outlier detection, we employ data obtained from the U.S. Geological Survey website https://www.usgs.gov/.

As discussed in Bilen and Huzurbazar (2002), data collected in this database comes from many different sources and is generally notorious for input errors. Here we focus on a monthly dataset, collected at irregular intervals from Mary 19876 to June 2020, measuring water conductance at the Green River near Greendale, UT, identified by site number 09234500.

A quick glance at the series indicates that there is an unusually large drop from typical values (500 to 800 units) in September 1999 which drops to roughly 7.4 units. This is an unusually large drop and is almost certainly an outlying observation. In an attempt to identify the aforementioned outlier and perhaps uncover others, the example conducts the Bilen and Huzurbazar (2002) wavelet outlier detection using a DWT transform with a Haar filter, universal threshold, and a mean median absolute deviation estimator for wavelet coefficient variance. Furthermore, the maximum decomposition level is restricted to 3.

To proceed, open the series window and click on View/Wavelet Analysis/Outlier Detection.... Change the Max Scale to 4 and under the MAD group, set the Method to Mean Median Abs. Dev., and click on OK.

EViews will display output in the form of a spool object with the spool tree listing the summary, outlier table, and outlier graphs for each of the relevant scales.

The first spool entry is a summary of the outlier detection procedure describing the decomposition method and filter, sample information, threshold method, and wavelet coefficient variance computation.

Next is an outlier overview table listing the exact location of a detected outlier along with its value and absolute deviation from the series mean and median, respectively.

We may scroll down or click on one of the scale entries on the left-hand side to display graphs of the outliers.

These graphs show, for each scale, plots of the original series with red dots and dotted vertical lines identifying outlying observations.

If you wish to save the wavelet results to the workfile you may click on Proc/Make Wavelet Objects...

For discussion, see “Wavelet Objects”.

Many economic time series feature time-varying characteristics such as non-stationarity, volatility, seasonality, and structural discontinuities. Wavelet analysis is a natural framework for analyzing these phenomena without imposing simplifying assumptions such as stationarity. In particular, wavelet filters can decompose and reconstruct a time series, as well as its correlation structure, across time scales. Moreover, these decompositions can be orthogonalized so that a decomposition at one scale is uncorrelated with a decomposition at another. This property is clearly useful in isolating features associated with a given timescale, as well as achieving de facto decomposition and preservation of variation akin to eigenvalue diagonalization in principal component analysis.

Wavelet analysis is analogous, in many ways, to Fourier spectral analysis as both methods can represent a time series signal in a different space. This is achieved by expressing, to a desired degree of precision, the time series signal as a linear combination of basis functions.

For Fourier analysis, these basis functions are sines and cosines. These functions approximate global variation well, but are poorly adapted to modeling local time-variation, and to capturing discontinuous, non-linear, and non-stationary processes whose impulses persist and evolve with time.

Accordingly, Fourier basis functions approximate global variation well. Nevertheless, local variation, otherwise known as time-variation in time-series, is poorly adapted to cylical functions such as the sinus and cosinus. Accordingly, the trigonometric basis functions are ideally adapted to stationary processes characterized by impulses which wane with time, but are poorly adapted to discontinuous, non-linear, and non-stationary processes whose impulses persist and evolve with time. The Fourier transform is localized well in frequency, but lacks good time resolution. Furthermore, if components captured at a given frequency are non-stationary and time-dependent, then attempting to capture all such components with a single frequency (which is itself time independent), will clearly not work very well. To surmount this xed time-frequency relationship, a new set of basis

We say that wavelets have mean zero and unit energy. The term energy originates in the signal processing literature, and is formalized as

Since wavelet basis functions are de facto location and scale transformations of a single function, they are an ideal tool for multiresolution analysis (MRA)–the ability to analyze a signal at different frequencies with varying resolutions.

Moving along the time domain, MRA allows one to zoom to a desired level of detail such that high (low) frequencies yield good (poor) time resolutions and poor (good) frequency resolutions. Since economic time series often exhibit multiscale features, wavelet techniques can decompose these series into constituent processes associated with different time scales.

and the inverse transformation for reconstructing the original process is

See Percival and Walden (2000) for a detailed discussion.

In practice, the DWT proceeds as a filtering operation (both low and high pass) on a time series, following the pyramid algorithm of Mallat (1989).

Like traditional filters which are used to extract components of a time series such as trends, seasonalities, business cycles, and noise, wavelets filters perform a similar role. Like traditional filters, wavelet filters

are also designed to capture low and high frequencies, and have a particular length. This length governs how much of the original series information is used to extract low and high frequency phenomena. The role of the length is very similar to the role of the autoregressive (AR) order in traditional time series models where higher AR orders imply more historical observations influence the present.

In contrast, Daubechies (1992) introduced a family of filters of even length that are indexed by the polynomial degree they are able to capture – rather the number of vanishing moments. Thus, the Haar filter, which is of length 2, can only capture constants and linear functions. The Daubechies wavelet filter of length 4 can capture everything from a constant to a cubic function. Accordingly, higher filter lengths are associated with higher smoothness.

It should be noted that Daubechies (1992) filters, also known as daublets, are typically not symmetric. If a more symmetric version of the daublet filters is required, then the class known as least asymmetric, or symmlets, is used.

The MRA is typically used to derive approximations for the original series using its lower and upper frequency components. Since upper frequency components are associated with transient features and are captured by the wavelet coefficients, the detail series will in fact extract those features of the original series which are typically associated with “noise”. Alternatively, since lower frequency components are associated with perpetual features and are captured by the scaling coefficients, the smooth series will in fact extract those features of the original series which are typically associated with the “signal”.

It’s worth mentioning here that since wavelet filtering can result in boundary coefficients, the detail and smooth series will also have observations affected by the boundary.

This energy decomposition is known as the wavelet power spectrum (WPS) and is arguably the most important of the properties of the DWT, and bears resemblance to the spectral density function (SDF) used in Fourier analysis.

It is also possible to derive confidence intervals for the contribution to the overall variance at each scale.

An asymptotic normal confidence interval is given by

where

Percival and Waldon (2000) employ an asymptotic Chi-square confidence interval

A particularly important and useful application of wavelets is outlier detection. While the subject matter has received some attention over the years starting with Greenblatt (1996), here we focus on a rather simple and appealing contribution by Bilen and Huzurbazar (2002).

The appeal of the Bilen and Huzurbazar approach is that it doesn’t require model estimation, is not restricted to processes generated via ARIMA, and works in the presence of both additive and innovational outliers. The approach also assumes that wavelet coefficients are approximately independent and identically normal variates. While EViews offers the ability to perform this procedure using a MODWT, it’s generally better suited to the DWT. Moreover, Bilen and Huzurbazar suggest that Haar is the preferred filter since the latter yields coefficients large in magnitude in the presence of jumps or outliers. They also suggest that the transformation be carried out only at the first scale.

The overall procedure works on the principle of thresholding and the authors suggest the use of the universal threshold. The idea here is that extreme (outlying) values will register as noticeable spikes in the spectrum. These spike generating values are outlier candidates.

See Bilen and Huzurbazar (2002) for additional detail and discussion.

The principle behind wavelet-based signal extraction, otherwise known as wavelet shrinkage, is to shrink any wavelet coefficients not exceeding some threshold to zero and then exploit the MRA to synthesize the signal of interest using the modified wavelet coefficients. In other words, only those wavelet coefficients associated with very pronounced spectra are retained with the additional benefit of deriving a very sparse wavelet matrix.

The two most popular thresholding rules are

Given choice of the thresholding rules and optimal threshold value, all wavelet thresholding procedures involve the following:

The discussion above outlines the basic theory underlying wavelet analysis. Nevertheless, there are several practical (empirical) considerations which should be addressed. We focus here on three in particular: 1) selecting a wavelet filter, 2) handling non-dyadic series lengths, and 3) awareness of boundary conditions.

The type of wavelet filter is typically chosen to mimic the data to which it is applied. Shorter filters don’t approximate the ideal band pass filter well, but longer ones do. On the other hand, if the data are derived from piecewise constant functions, the Haar wavelet or other shorter wavelets may be more appropriate. Alternatively, if the underlying data is smooth, longer filters may be more appropriate. In this regard, it’s important to note that longer filters expose more coefficients to boundary condition effects than

shorter ones. Accordingly, the rule of thumb strategy is to use the filter with the smallest length that gives reasonable results. Furthermore, since the MODWT is not orthogonal and its wavelet coefficients are correlated, wavelet filter choice is not as vital as for orthogonal DWT. Nevertheless, if alignment to time is important (i.e. zero phase filters), the least asymmetric family of filters may be a good choice.

Whatever the approach, wavelet analysis should quantify how many wavelet coefficients are affected by boundary conditions.

Recall that the DWT requires an input series of dyadic length. Naturally, this condition is rarely satisfied in practice. Broadly speaking, there are two approaches: either shorten the input series to dyadic length at the expense of losing observations, or “pad” the input series with observations to achieve dyadic length.

Although the choice of padding values is ultimately arbitrary, there are three popular choices, none of which has proven superior: 1) pad with zeros, 2) pad with mean, 3) pad with median.