window_reg: General Interface for Window Forecast Models

In modeltime: The Tidymodels Extension for Time Series Modeling

General Interface for Window Forecast Models

Description

window_reg() is a way to generate a specification of a window model before fitting and allows the model to be created using different backends.

Usage

window_reg(mode = "regression", id = NULL, window_size = NULL)

Arguments

mode	A single character string for the type of model. The only possible value for this model is "regression".
id	An optional quoted column name (e.g. "id") for identifying multiple time series (i.e. panel data).
window_size	A window to apply the window function. By default, the window uses the full data set, which is rarely the best choice.

Details

A time series window regression is derived using window_reg(). The model can be created using the fit() function using the following engines:

• "window_function" (default) - Performs a Window Forecast applying a window_function (engine parameter) to a window of size defined by window_size

Engine Details

function (default engine)

The engine uses window_function_fit_impl(). A time series window function applies a window_function to a window of the data (last N observations).

• The function can return a scalar (single value) or multiple values that are repeated for each window

• Common use cases:

  • Moving Average Forecasts: Forecast forward a 20-day average

  • Weighted Average Forecasts: Exponentially weighting the most recent observations

  • Median Forecasts: Forecasting forward a 20-day median

  • Repeating Forecasts: Simulating a Seasonal Naive Forecast by broadcasting the last 12 observations of a monthly dataset into the future

The key engine parameter is the window_function. A function / formula:

• If a function, e.g. mean, the function is used with any additional arguments, ... in set_engine().

• If a formula, e.g. ~ mean(., na.rm = TRUE), it is converted to a function.

This syntax allows you to create very compact anonymous functions.

Fit Details

Date and Date-Time Variable

It's a requirement to have a date or date-time variable as a predictor. The fit() interface accepts date and date-time features and handles them internally.

• fit(y ~ date)

ID features (Multiple Time Series, Panel Data)

The id parameter is populated using the fit() or fit_xy() function:

ID Example: Suppose you have 3 features:

1. y (target)

2. date (time stamp),

3. series_id (a unique identifer that identifies each time series in your data).

The series_id can be passed to the window_reg() using fit():

• window_reg(id = "series_id") specifes that the series_id column should be used to identify each time series.

• fit(y ~ date + series_id) will pass series_id on to the underlying functions.

Window Function Specification (window_function)

You can specify a function / formula using purrr syntax.

• If a function, e.g. mean, the function is used with any additional arguments, ... in set_engine().

• If a formula, e.g. ~ mean(., na.rm = TRUE), it is converted to a function.

This syntax allows you to create very compact anonymous functions.

Window Size Specification (window_size)

The period can be non-seasonal (⁠window_size = 1 or "none"⁠) or yearly seasonal (e.g. For monthly time stamps, window_size = 12, window_size = "12 months", or window_size = "yearly"). There are 3 ways to specify:

1. window_size = "all": A seasonal period is selected based on the periodicity of the data (e.g. 12 if monthly)

2. window_size = 12: A numeric frequency. For example, 12 is common for monthly data

3. window_size = "1 year": A time-based phrase. For example, "1 year" would convert to 12 for monthly data.

External Regressors (Xregs)

These models are univariate. No xregs are used in the modeling process.

See Also

fit.model_spec(), set_engine()

Examples

library(dplyr)
library(parsnip)
library(rsample)
library(timetk)

# Data
m750 <- m4_monthly %>% filter(id == "M750")
m750

# Split Data 80/20
splits <- initial_time_split(m750, prop = 0.8)

# ---- WINDOW FUNCTION -----

# Used to make:
# - Mean/Median forecasts
# - Simple repeating forecasts

# Median Forecast ----

# Model Spec
model_spec <- window_reg(
        window_size     = 12
    ) %>%
    # Extra parameters passed as: set_engine(...)
    set_engine(
        engine          = "window_function",
        window_function = median,
        na.rm           = TRUE
    )

# Fit Spec
model_fit <- model_spec %>%
    fit(log(value) ~ date, data = training(splits))
model_fit

# Predict
# - The 12-month median repeats going forward
predict(model_fit, testing(splits))


# ---- PANEL FORECAST - WINDOW FUNCTION ----

# Weighted Average Forecast
model_spec <- window_reg(
        # Specify the ID column for Panel Data
        id          = "id",
        window_size = 12
    ) %>%
    set_engine(
        engine = "window_function",
        # Create a Weighted Average
        window_function = ~ sum(tail(.x, 3) * c(0.1, 0.3, 0.6)),
    )

# Fit Spec
model_fit <- model_spec %>%
    fit(log(value) ~ date + id, data = training(splits))
model_fit

# Predict: The weighted average (scalar) repeats going forward
predict(model_fit, testing(splits))

# ---- BROADCASTING PANELS (REPEATING) ----

# Simulating a Seasonal Naive Forecast by
# broadcasted model the last 12 observations into the future
model_spec <- window_reg(
        id          = "id",
        window_size = Inf
    ) %>%
    set_engine(
        engine          = "window_function",
        window_function = ~ tail(.x, 12),
    )

# Fit Spec
model_fit <- model_spec %>%
    fit(log(value) ~ date + id, data = training(splits))
model_fit

# Predict: The sequence is broadcasted (repeated) during prediction
predict(model_fit, testing(splits))

modeltime documentation built on Oct. 23, 2024, 1:07 a.m.