Loess: local polynomial regression🔗ℹ

Loess is a library for calculating lines of best fit using local polynomial regression, which is a non-parametric regression method that calculates various individual regressions for each point. It is useful for when you have data that does not follow a clearly discernable pattern, or when you just want a smooth line for a data visualization.

xs and ys are assumed to be vectors of the same length, with each observation being at the same index in each vector.

The optional keyword argument #:span dictates the proportion of the data to factor into each local regression. For example, if the span is 0.1, the local regression will use the (* 0.1 (vector-length xs)) nearest neighbors to the input point. Generally, increasing #:span increases the influence of bias in the resulting model, and decreasing it increases the influence of variance. This number is assumed to be between (/ (add1 degree) (vector-length xs)) and 1 (as otherwise there are not enough points to create a fit).

The optional keyword argument #:degree dictates the degree of each local regression, and not the degree of the complete regression.

If the result is suboptimal or jittery, changing #:span and #:degree is generally the key. If the line appears too flat, generally you want to reduce #:span, and if it appears too chaotic, generally you want to increase #:span.