Environmental & Analytical Toxicology

Automating the Smoothing of Time Series Data

Abstract

Shilpy Sharma, David A Swayne and Charlie Obimbo

Modelling requires comparison of model outputs to measurements, for calibration and verification. A key aspect data smoothing is to “filter out” noise. Often, data must be adjusted to a model’s time step (e.g. hourly to daily). For noisy data, LOWESS/LOESS (Locally Weighted Scatterplot Smoothing) is a popular piecewise regression technique. It produces a “smoothed” time series. LOWESS/LOESS is often used to visually assess the relationship between variables. The selection of LOWESS tuning parameters is usually performed on a visual trial and error basis. We investigate the so-called robust AIC (Akaike Information Criteria) for automatic selection of smoothing. Robust Pearson correlation coefficient and mean-squared error are employed to determine the polynomial degree of piecewise regression. The exclusion of outliers is attempted using a Hampel outlier identifier. We illustrate, assuming noisy linear data, how our proposed methods work for auto-tuning both the smoothing parameter and the degree of polynomial for LOWESS/ LOESS.