Customizing Wrapper Functions"

In forecastML: Time Series Forecasting with Machine Learning Methods

knitr::opts_chunk$set(fig.width = 7.15, fig.height = 4)

Purpose

The purpose of this vignette is to provide a closer look at how the user-supplied model training and predict wrapper functions can be modified to give greater control over the model-building process. The goal is to present examples of how the wrapper functions could be flexibly written to keep a linear workflow in forecastML while modeling across multiple forecast horizons and validation datasets. The alternative would be to train models across a single forecast horizon and/or validation window and customize the wrapper functions for this specific setup.

Example 1 - Multiple forecast horizons and 1 model training function

• Load packages and data.

library(DT)
library(dplyr)
library(ggplot2)
library(forecastML)
library(randomForest)

data("data_seatbelts", package = "forecastML")
data <- data_seatbelts

data <- data[, c("DriversKilled", "kms", "PetrolPrice", "law")]

dates <- seq(as.Date("1969-01-01"), as.Date("1984-12-01"), by = "1 month")

• Create a lagged data frame for model training. We'll train 2 models: one across a 1:3-month horizon and the other across a 1:12-month horizon.

data_train <- forecastML::create_lagged_df(data,
                                           type = "train",
                                           outcome_col = 1, 
                                           lookback = 1:12,
                                           horizons = c(3, 12),
                                           dates = dates,
                                           frequency = "1 month")

# View the horizon 3 lagged dataset.
DT::datatable(head((data_train$horizon_3)), options = list("scrollX" = TRUE))

• We'll train our models to predict on 1 validation dataset. Setting window_length = 0 means that a single validation dataset will span from window_start to window_stop.

windows <- forecastML::create_windows(data_train, window_length = 0, 
                                      window_start = as.Date("1984-01-01"),
                                      window_stop = as.Date("1984-12-01"))

plot(windows, data_train)

User-defined model-training function

• The key to customizing training across forecast horizons--here we have 2--is to modify the model training wrapper function based on the horizon-specific dataset in our lagged_df object data_train.

• Each dataset's forecast horizon is stored as an attribute.

attributes(data_train$horizon_3)$horizon
attributes(data_train$horizon_12)$horizon

• We'll train a Random Forest model with different settings for the 3-month and 12-month datasets.

• The first argument to the user-defined model training function is always the horizon-specific dataset from create_lagged_df(type = "train") and is passed into the wrapper function internally in train_model(). Any number of additional parameters can be defined in this wrapper function by either (a) setting arguments here--like below--or (b) setting the arguments in train_model(...).

model_function <- function(data, my_outcome_col = 1, n_tree = c(200, 100)) {

  outcome_names <- names(data)[my_outcome_col]
  model_formula <- formula(paste0(outcome_names,  "~ ."))

  if (attributes(data)$horizon == 3) {  # Model 1

          model <- randomForest::randomForest(formula = model_formula, 
                                              data = data, 
                                              ntree = n_tree[1])

          return(list("my_trained_model" = model, "n_tree" = n_tree[1], 
                      "meta_data" = attributes(data)$horizon))

  } else if (attributes(data)$horizon == 12) {  # Model 2

          model <- randomForest::randomForest(formula = model_formula, 
                                              data = data, 
                                              ntree = n_tree[2])

          return(list("my_trained_model" = model, "n_tree" = n_tree[2],
                      "meta_data" = attributes(data)$horizon))
  }
}

• Train the models.

model_results <- forecastML::train_model(data_train, windows, model_name = "RF", model_function)

• View the return() values from the user-defined model_function(). The returned values are stored in my_training_results$horizon_h$window_w$model.

model_results$horizon_3$window_1$model
model_results$horizon_12$window_1$model

User-defined prediction function

• The user-defined prediction function only takes two positional parameters. The first parameter is the returned value of the user-defined modeling function; here, a list with 3 elements. The second parameter is the horizon-specific dataset with model features from create_lagged_df() (type = "train" or type = "forecast").

prediction_function <- function(model, data_features) {

    if (model$meta_data == 3) {  # Perform a transformation specific to model 1.

        data_pred <- data.frame("y_pred" = predict(model$my_trained_model, data_features))
    }

    if (model$meta_data == 12) {  # Perform a transformation specific to model 2.

        data_pred <- data.frame("y_pred" = predict(model$my_trained_model, data_features))
    }

  return(data_pred)
}

• Predict

data_results <- predict(model_results,
                        prediction_function = list(prediction_function),
                        data = data_train)

• Plot the performance on the validation dataset.

plot(data_results)

---

forecastML documentation built on July 8, 2020, 7:27 p.m.