Optimize Loess Models for Time Series Resampling

Internal function used in zoo_resample(). It finds the span parameter of a univariate Loess (Locally Estimated Scatterplot Smoothing.) model y ~ x fitted with stats::loess() that minimizes the root mean squared error (rmse) between observations and predictions, and returns a model fitted with such span.

Usage

utils_optimize_loess(x = NULL, y = NULL, max_complexity = FALSE)

Arguments

x

(required, numeric vector) predictor, a time vector coerced to numeric. Default: NULL

y

(required, numeric vector) response, a column of a zoo object. Default: NULL

max_complexity

(required, logical). If TRUE, RMSE optimization is ignored, and the model of maximum complexity is returned. Default: FALSE

Value

Loess model.

See also

Other tsl_processing_internal: utils_drop_geometry(), utils_global_scaling_params(), utils_optimize_spline(), utils_rescale_vector()

Examples

#zoo time series
xy <- zoo_simulate(
  cols = 1,
  rows = 30
)

#optimize loess model
m <- utils_optimize_loess(
  x = as.numeric(zoo::index(xy)), #predictor
  y = xy[, 1] #response
)

print(m)
#> Call:
#> stats::loess(formula = y ~ x, data = model_df, enp.target = complexity_value, 
#>     degree = 1, surface = "direct")
#> 
#> Number of Observations: 30 
#> Equivalent Number of Parameters: 17.45 
#> Residual Standard Error: 0.1064 

#plot observation
plot(
  x = zoo::index(xy),
  y = xy[, 1],
  col = "forestgreen",
  type = "l",
  lwd = 2
  )

#plot prediction
points(
  x = zoo::index(xy),
  y = stats::predict(
    object = m,
    newdata = as.numeric(zoo::index(xy))
    ),
  col = "red4"
  )