Custom Feature Lags

The hardware and bandwidth for this mirror is donated by dogado GmbH, the Webhosting and Full Service-Cloud Provider. Check out our Wordpress Tutorial.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]dogado.de.

Custom Feature Lags

Nickalus Redell

2020-05-05

Purpose

The purpose of this vignette is to demonstrate how custom predictor or feature lags can be created for forecast model inputs in forecastML with the forecastML::create_lagged_df() function. The rationale behind creating custom feature lags is to improve model accuracy by removing noisy or redundant features in high dimensional training data. Keeping only those feature lags that show high autocorrelation or cross-correlation with the modeled outcome–e.g., 3 and 12 months for monthly data–is a good place to start.

Load packages and data

library(forecastML)
library(DT)

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

data <- data[, c("DriversKilled", "kms", "PetrolPrice", "law")]
DT::datatable(head(data, 5))

Create a Date Column

Dates are optional for forecasting with non-grouped data, but we’ll add a date column here to illustrate the functionality.
The dataset does not come with a date column, but the data was collected monthly from 1969 through 1984. This actually works out nicely because dates are passed in a separate argument, dates, in create_lagged_df().

date_frequency <- "1 month"

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

Custom Feature Lags

We’ll use custom lags for 3 of our 4 modeled features using the lookback_control argument in create_lagged_df().
- Column 1: 3 and 12 months in the past.
- Column 2: 1 through 3 months in the past.
- Column 3: 1 through 12 months in the past.
- Column 4: We’ll treat law as a dynamic feature which won’t be lagged.
Although it’s not required, our feature lags will be common across forecast models–3 models that forecast (1) 1 month out, (2) 1:6 months out, and (3) 1:12 months out. As we’ll see below, feature lags that don’t support direct forecasting to the given horizon are silently dropped from the lagged data.frames.

horizons <- c(1, 6, 12)  # forecasting 1, 1:6, and 1:12 months into the future.

# Create a list of length 3, one slot for each modeled forecast horizon.
lookback_control <- vector("list", length(horizons))

# Within each horizon-specific list, we'll identify the custom feature lags.
lookback_control <- lapply(lookback_control, function(x) {
  list(
    c(3, 12),  # column 1: DriversKilled
    1:3,       # column 2: kms
    1:12,      # column 3: PetrolPrice
    0          # column 4: law; this could be any value, dynamic features are set to '0' internally.
    )
  })

data_train <- forecastML::create_lagged_df(data, type = "train",
                                           outcome_col = 1,
                                           horizons = horizons, 
                                           lookback_control = lookback_control,
                                           dates = dates,
                                           frequency = date_frequency,
                                           dynamic_features = "law")

Below is a series of feature-level plots of the resulting lagged data.frame features for each forecast horizon in data_train.
Notice, for instance, how the 1:3 month lags for kms were dropped from the 6- and 12-month-out forecast modeling datasets as these lags don’t support direct forecasting at these time horizons.

plot(data_train)

Removing Features

Now, let’s say that a lag of 12 months for PetrolPrice is a poor predictor for our long-term, 12-month-out forecast model. We can remove it by assigning a NULL value in the appropriate slot in our lookback_control argument.
Notice that the NULL has to be placed in a list() to avoid removing the list slot altogether.

horizons <- c(1, 6, 12)  # forecasting 1, 1:6, and 1:12 months into the future.

# A list of length 3, one slot for each modeled forecast horizon.
lookback_control <- vector("list", length(horizons))
lookback_control <- lapply(lookback_control, function(x) {
  
  # 12 feature lags for each of our 4 modeled features. Dynamic features will be coerced to "0" internally.
  lapply(1:4, function(x) {1:12})
  })

# Find the column index of the feature that we're removing.
remove_col <- which(grepl("PetrolPrice", names(data)))

# Remove the feature from the 12-month-out lagged data.frame.
lookback_control[[which(horizons == 12)]][remove_col] <- list(NULL)

data_train <- forecastML::create_lagged_df(data, type = "train",
                                           outcome_col = 1, 
                                           lookback_control = lookback_control,
                                           horizons = horizons,
                                           dates = dates,
                                           frequency = date_frequency,
                                           dynamic_features = "law")

Inspecting the plot confirms that the 12-month-lagged feature for PetrolPrice is not a feature in our 12-month-out forecast model training data set.

plot(data_train)[[remove_col]]  # we're selecting 1 of our 3 lagged feature-level plots.

Inspecting the modeling data.frame confirms that the 12-month-lagged feature for PetrolPrice is not a feature in our 12-month-out forecast model training data set.

DT::datatable(head(data_train$horizon_12))

These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.
Health stats visible at Monitor.