qreg_gam: Multiple Quantile Regression Using Generalised Additive...
In jbrowell/ProbCast: Flexible Probabilistic Forecasting

qreg_gam

R Documentation

Multiple Quantile Regression Using Generalised Additive Models and Linear Quantile Regression

Description

This function fits multiple conditional linear quantile regression models to the residuals of a generalised additive model using mgcv and with facilities for cross-validation. Quantile regression may be performed using user-specified formula or the design matrix of the fitted GAM.

Usage

qreg_gam(
  data,
  formula,
  formula_qr = NULL,
  model_res2 = F,
  formula_res2 = NULL,
  quantiles = c(0.25, 0.5, 0.75),
  cv_folds = NULL,
  use_bam = T,
  exclude_train = NULL,
  sort = T,
  sort_limits = NULL,
  ...
)

Arguments

`data`	A `data.frame` containing target and explanatory variables. May optionally contain a column called "kfold" with numbered/labelled folds and "Test" for test data.
`formula_qr`	Formula for linear quantile regression model for GAM residuals. Term `gam_pred` is the prediction from the above GAM may be included in this formula. If null, the "terms" of the GAM model are used as features in linear quantile regression.
`model_res2`	If `TRUE` also model squared residuals of GAM using a GAM. Defaults to `FALSE`.
`formula_res2`	Formula for GAM to predict squared residuals.
`quantiles`	The quantiles to fit models for.
`cv_folds`	Control for cross-validation with various options, either: the column name of the fold index supplied in data. Observations and inputs in the index labelled "Test" will serve as test data and held out in model training. an integer giving the number of cross validation folds to generate. Folds are constructed as block chunks. Default behaviour is 5 folds. NULL indicates that no cross validation should be performed and the returned model is trained on all `data`.
`use_bam`	If `TRUE` (default) then GAM is fit using (`bam()`) in stead of `gam()`. `bam` is better suited to large datasets but not all `gam` model options are available with `bam`. Alternative smooths, such as cubic regression splines aid faster estimation than the default smooth `bs="tp"`. See `bam()` documentation for further details.
`exclude_train`	A column name in `data` indicating if a row should be excluded from model training, i.e. if it contains bad data (will be coerced to `logical`). Alterntively, an `integer` or `logical` vector with length equal to the number of rows in `data` indicating the same. Rows labeled `TRUE` are excluded from model training.
`sort`	`boolean` Sort quantiles using `SortQuantiles()`?
`sort_limits`	`Limits` argument to be passed to `SortQuantiles()`. Constrains quantiles to upper and lower limits given by `list(U=upperlim,L=lowerlim)`.
`...`	Additional arguments past to `gam()` (or `bam()`).
`formala`	A `formula` object with the response on the left of an ~ operator, and the terms, separated by + operators, on the right passed to `gam()` or `bam()` from `mgcv`.

Details

The returned predictive quantiles and GAM predictions are those produced out-of-sample for each cross-validation fold (using models trained on the remaining folds but not "Test" data). Predictive quantiles corresponding to "Test" data are produced using models trained on all non-test data.