Quantile and expectile smoothing based on \(L_1\)-norm and \(L_2\)-norm fuzzy transforms (Q2329594): Difference between revisions

Often, the original signal is smooth, while the noise is usually not. Then, smoothing the observed signal will drastically decrease the effect of the noise. A natural way to smooth a function \(f(x)\) is to select several localized smooth function \(A_k(x)\) -- e.g., for which \(\sum_k A_k(x)=1\) for all \(x\) -- and approximate the original signal by a linear combination \(\sum_k y_k\cdot A_k(x)\). One way to compute \(y_k\) (known as \textit{F-transform}) is to select \(y_k\) that minimize the weighted mean squared difference \(\int (f(x)-y_k)^2 A_k(x)\,dx\). A more robust approach is to minimize the \(L_1\)-difference \(\int |f(x)-y_k| A_k(x)\,dx\). (Instead of approximating \(f(x)\), for each \(k\), by a constant \(y_k\), we can also approximate it by a polynomial \(y_k(x)\) of a given degree.) In many practical situations, positive noise values \(f(x)-y\) have a different probability than negative ones. To take this into account, the authors propose to use \textit{expectile} and \textit{quantile} techniques, i.e., to find the values \(y_k\) and \(\omega\) that minimize \(\int w(f(x),y_k,\omega) (f(x)-y_k)^2\,dx\) or \(\int w(f(x),y_k,\omega) |f(x)-y_k|^2\,dx\), where \(w(a,b,\omega)=\omega\) if \(ab\) and \(w(a,b,\omega)=1-\omega\) otherwise. On the example of several financial time series, they show that the result of such smoothing is closer to the original signal than for previously proposed smoothing algorithm.