Time Series Forecasting Made Simple (Part 3.2): A Deep Dive into LOESS-Based Smoothing

LOESS (Locally Estimated Scatterplot Smoothing) is a technique used within STL decomposition to extract trend and seasonal components from time series data. It functions as a weighted simple linear regression, assigning greater importance to data points closer to the target point being smoothed. The process involves calculating distances from a target point, determining tricube weights, and then fitting a weighted regression line to find the smoothed value. This procedure is applied iteratively to the deseasonalized series to refine the trend and then to seasonal subseries to refine seasonality, ultimately separating the time series into its core components.