Denoising Non-Stationary Signals via Dynamic Multivariate Complex Wavelet Thresholding Authors: Kim C. Raath, Katherine B. Ensor, Alena Crivello, David W. Scott Published: 2023 ABSTRACT Over the past few years, we have seen an increased need to analyze the dynamically changing behaviors of economic and financial time series. These needs have led to significant demand for methods that denoise non-stationary time series across time and for specific investment horizons (scales) and localized windows (blocks) of time. Wavelets have long been known to decompose non-stationary time series into their different components or scale pieces. Recent methods satisfying this demand first decompose the non-stationary time series using wavelet techniques and then apply a thresholding method to separate and capture the signal and noise components of the series. Traditionally, wavelet thresholding methods rely on the discrete wavelet transform (DWT), which is a static thresholding technique that may not capture the time series of the estimated variance in the additive noise process. We introduce a novel continuous wavelet transform (CWT) dynamically optimized multivariate thresholding method (WaveL2E). Applying this method, we are simultaneously able to separate and capture the signal and noise components while estimating the dynamic noise variance. Our method shows improved results when compared to well-known methods, especially for high-frequency signal-rich time series, typically observed in finance. 1. Introduction Various time series in economics and finance are formed by non-stationary processes. For instance, Exchange Traded Funds (ETF’s) consist of combinations of different equities functioning at different periodicities. In order to understand and identify dynamic associations, it is necessary to explore the characteristics within each individual time series and the changing dynamics across a series. Over the past decade, there has been a significant increase in the use of time-varying spectral representations to analyze dynamic time series behavior in finance. Recently, emphasis has been placed on using wavelets. Rostan and Rostan [1] mention that quantitative analysts (quants) “nibbling” with wavelets are heavily being recruited by major hedge funds. They specifically point out that the interest in wavelet analysis has spiked in recent years because new methods and techniques have come about to analyze physical phenomena, such as seismic and electrical signals, that propagate through time in waveform. These methods are being adapted and applied to economic and financial time series as these signals also propagate through time in waveform. The application of time-varying techniques to augment traditional portfolio management tools by distinguishing across multiple investment horizons or scales is becoming a growing field of interest as well [2,3,4]. Chaudhuri and Lo [5], who coined the term spectral portfolio theory, suggest that the flexibility of the wavelet transform could be used to overcome various difficulties of the Fourier transform for spectral portfolio analysis. Wavelet transforms, when specified correctly, are simple to interpret and can add tremendous value to the quantitative finance models that financial engineers implement. Wavelets have also proven useful in characterizing two-dimensional fractional Brownian fields, often used in financial studies [6]. For an expanded discussion of wavelet theory and implementation examples in finance, see [7]. The decomposition of time series into their component pieces is frequently applied, but further inference such as optimal dynamic thresholding and forecasting has only increased in interest over the past few years. A recent paper by Reményi and Vidakovic [8] proposes adaptive wavelet denoising methodology using a fully Bayesian hierarchical model in the complex wavelet domain. They compare their results to two methods proposed by [9]. These authors also compare their methods to the well-known Empirical Bayesian Threshold (EBayesThresh) of [10] and various other superior universal hard and soft thresholding methods such as the SureShrink methods of [11]. Even though these two papers were written ten years apart and are over a decade old, they are still commonly referenced and can be seen as the gold standard in the domain of wavelet thresholding. For a survey of other wavelet denoising or thresholding techniques and their comparisons, see [12]. More recent thresholding (denoising) methods were developed by the authors of [9] who proposed two new complex-valued wavelet techniques, while the authors of [13] proposed a data-driven threshold called the SureBlock, which uses Stein’s unbiased risk estimate (SURE) criterion. He et al. [14] designed a new threshold, considering interscale correlation by improving the basic universal hard threshold in order to address the issue of removing too many wavelet coefficients. Each method improved characteristics that are lacking in the existing “gold standard” methods. In general, hard thresholding reduces all coefficients to zero that do not exceed the threshold. Soft thresholding pushes toward zero any coefficient whose magnitude exceeds the threshold, and zeros the coefficient otherwise. As a comparison to our proposed method, we specifically use the SureShrink methods for the hard and soft thresholds. We also compare our methods to the EBayesThresh. We choose these methods for our comparison as they show superior results to the more recent and well-cited methods pointed out earlier in this paper. The choice of wavelet family is very important. Most solutions are not based on the continuous wavelet transform (CWT) due to the difficulty of having a tractable solution or an analytical solution [1]. The CWT is calculated by shifting, continuously, a scalable function over a signal and by calculating the correlation between the function and signal, resulting in potential redundancies. However, we mitigate these redundancies as well as address both the tractability and analytical issues by uniquely choosing the Morlet wavelet with the exact specifications for concise interpretation, see Section 2 and [15] for more details. The novel wavelet thresholding method we propose in this paper builds upon work on minimum distance estimation, initiated by the authors of [16,17]. We propose decomposing each univariate time series into scale-specific wavelet coefficients using a CWT. After decomposition, we apply Scott’s multivariate minimum distance partial density estimation (L2E) to each time increment. In doing so, we separate each time point into a mixture model of signal and noise without having to explicitly assume a model for the signal. Upon analyzing these results, we are not only able to find the optimal signal, but we also estimate the variance of the noise component. In essence, we model the wavelet coefficients as a two-component mixture, where the signal component is unspecified and the noise component is Gaussian; see the Methods section for more details. The L2E (distance criterion) can be used to exclude groups of outliers; therefore, by estimating the partial density parameters of the noise component, we do not have to make assumptions on the signal component. Ultimately, we are able to distinguish between signal and noise, by taking our signal to be “outliers” while assuming that the noise has a known structure which we can model. Section 2 describes the WaveL2E method, starting with a mathematical exposition of wavelets in Section 2.1. We use [18,19,20] as examples to explain wavelet characteristics and introduce statistical methods that extend the capabilities of wavelets to visualize unexpected structures in multidimensional data in Section 2.1.1. We then introduce the WaveL2E in Section 2.2 and further identify how our method compares to other well-established wavelet thresholding methods, using simulated time series that have dynamically changing behaviors. In Section 3, we discuss and display the comparative analysis. In this section, we also note that our WaveL2E results yield improved outcomes with added information and interpretations. Before concluding in Section 5 with the strengths of the new methods and a list of the next potential steps and applications, we introduce an example in Section 4. The selected example addresses the investment relationships in water and energy ETF’s [19]. In this example, we highlight the post-thresholding improved coherence analysis of the water–energy nexus. 2. Methods 2.1. Wavelets Wavelets are mathematical functions that decompose temporally localized frequency components, providing information of each component with a resolution matched to its scale. Wavelet analysis in economics and finance has demonstrated promising results. This has led to a resurgent of the literature focused on wavelet transforms and denoising. 2.1.1. CWT A wavelet is usually derived by scaling and shifting a mother wavelet ψ(t)∈L2(R) into daughter wavelets ψa,b(t)∈L2(R):(1)ψa,b(t)=1aψt−ba, where a>0 defines the scale and b∈R defines the shift. Given a time series x(t)∈L2(R), the continuous wavelet transform (CWT) provides a decomposition of x(t) into time-scale components:(2)Wx(a,b)=〈x,ψa,b〉=∫Rx(t)ψa,b∗(t)dt, where ∗ denotes the complex conjugate. Usually the CWT coefficients are presented in a power spectrum, where wavelet power is defined as |Wx(a,b)|2 and visualized in the time-scale {b,a} half plane with logarithmic scale a-axis (vertical) increasing upward and a linear time-scale b-axis (horizontal) (see Figure 1). More generally, the CWT can be seen as a set of continuous band-pass filters applied to a time series. Formally, ψ(t) is a mother wavelet if the following three conditions are met: (1) ∫Rψ(t)dt=0, (2) ∫Rψ2(t)dt=1, and (3) it satisfies an admissibility condition. The admissibility condition ensures the reconstruction of a time series from its wavelet transform:(3)x(t)=1a2Cψ∫0∞∫RWx(a,b)ψa,b(t)dbda, where Cψ=∫Rψ^(ω)2ωdω and ψ^(ω) is the Fourier transform of ψ(t) in the CWT. By the convolution theorem, the CWT, Wx(a,b), of a discrete time series x(t) can be approximated by averaging n (the length of the time series) convolutions for each scale using the discrete Fourier transform (DFT) of x(t) and ψ^(ω) [21]. Traditionally, the constant (Cψ) is called the wavelet admissible constant and a wavelet whose admissible constant satisfies 0