Out-of-Sample Prediction
Evaluate a fitted Batch model at query points that were not in the training set.
Overview
Out-of-sample prediction is available in Batch mode only. Streaming and Online modes do not support it.
predict(model::PredictModel, new_x; kwargs...) evaluates the fit at arbitrary query points, like R's predict(model, newdata).
It reuses fit's own (possibly delta-interpolated) smoothed curve for its y output — so predicting at a training x always exactly reproduces that point's fit output, regardless of delta. A fresh local fit is only run when outputs includes "derivative" or "se" (or an interval level), or when max_neighbor_distance needs the actual regression slope or standard error.
model comes from result.predict_model, populated only when retain_model=true was passed to Lowess.
Options
| Keyword Argument | Type | Default | Description |
|---|---|---|---|
outputs | Vector{String} | String[] | Select "se" and/or "derivative" |
intervals | NamedTuple | nothing | Grouped confidence, prediction, and optional bootstrap options |
seed | Union{Integer, Nothing} | nothing | Reproducible prediction-time bootstrap draws; 0 is valid |
extrapolation | String | "clamp" | Behavior for query points outside the training x-range |
max_extrapolation_distance | Union{Float64, Nothing} | nothing | Under "linear" extrapolation, the max allowed distance beyond the training boundary before erroring |
max_neighbor_distance | Union{Float64, Nothing} | nothing | Max allowed distance to the farthest training point in a query's local window before erroring |
outputs
Request "se" for standard errors (from the retained model's residual scale and per-point leverage) and "derivative" for the local slope at each query point. Analytic intervals compute their required standard errors even without an explicit "se" output.
intervals
A NamedTuple such as (confidence=0.95, prediction=0.95, bootstrap=200). confidence bounds the mean response and prediction bounds a new observation at each query point. Without bootstrap, these use normal-theory standard errors and the retained residual scale.
Set bootstrap to at least 2 to resample the retained Batch residuals, refit the model, and calculate query-point percentile intervals and standard errors.
seed
Seeds prediction-time bootstrap draws, independently of the fit/CV seed on Lowess. It does not enable bootstrap by itself.
extrapolation
Behavior for query points outside [min(x_train), max(x_train)]:
| Policy | Behavior |
|---|---|
"clamp" (default) | Clamps the query to the nearest boundary window |
"linear" | Linearly extrapolates from the nearest boundary point's local fit and slope |
"error" | Fails the whole call with an error |
maxextrapolationdistance
Under "linear" extrapolation, the maximum allowed distance beyond the training boundary before predict errors, instead of returning an unbounded value. nothing (default, uncapped).
maxneighbordistance
Maximum allowed distance to the farthest training point in a query's local window before predict errors. Guards against an "empty range" blind spot: a query point can fall within [min(x_train), max(x_train)] yet still be far from any real training point (e.g. training x in [0,10] and [90,100], query at x=50). nothing (default, uncapped); applies regardless of extrapolation.
Example
Basic Usage
using FastLOWESS
x = [1.0, 2.0, 3.0, 4.0, 5.0]
y = [2.1, 4.0, 6.2, 8.0, 10.1]
model = Lowess(; fraction=0.7, retain_model=true)
result = fit(model, x, y)
prediction = predict(result.predict_model, [1.5, 4.5])
println("Predicted y: ", prediction.y)Predicted y: [3.05, 9.05]Standard Errors and Derivative
prediction = predict(result.predict_model, [2.5]; outputs=["se", "derivative"])
println("y: ", prediction.y)
println("SE: ", prediction.standard_errors)
println("Derivative: ", prediction.derivative)y: [5.1]
SE: [0.0]
Derivative: [2.2]Bootstrap Intervals
prediction = predict(
result.predict_model,
[2.5, 4.5];
intervals=(confidence=0.95, prediction=0.95, bootstrap=200),
seed=7,
)
println("CI lower: ", prediction.confidence_lower)CI lower: [5.1, 9.05]Linear Extrapolation
prediction = predict(result.predict_model, [10.0]; extrapolation="linear")
println("Extrapolated y: ", prediction.y)Extrapolated y: [10.1]