Skip to contents

Fill Missing Target Values with Configurable Strategies

Source: R/gap_filling.R
fill_gaps.Rd

Apply gap-filling strategies to handle missing target values in time series data. This ensures train/forecast parity by making features deterministic and auditable.

Usage

fill_gaps(
  data,
  target,
  date,
  groups = NULL,
  strategy = "error",
  params = list()
)

Arguments

data

Data frame with time series data

target

Character, name of target column

date

Character, name of date column

groups

Character vector of group column names (NULL for ungrouped)

strategy

Character, gap-filling strategy. One of:

  • "error" - Fail if any NAs present (default, forces explicit choice)

  • "zero" - Replace NAs with 0 (appropriate for count data)

  • "locf" - Last observation carried forward

  • "nocb" - Next observation carried backward

  • "linear" - Linear interpolation (Phase 2)

  • "rolling_mean" - Centered rolling mean (Phase 2)

  • "stl" - Seasonal decomposition (Phase 3)

  • "borrow" - Cross-series borrowing for panel data (Phase 3)

  • "custom" - User-provided function (Phase 4)

params

List of strategy-specific parameters. Common parameters:

  • max_gap - Maximum gap length to fill (for locf, nocb)

  • extrapolate - Allow extrapolation (for linear)

  • window - Window size (for rolling_mean)

  • period - Seasonal period (for stl)

  • fn - User function (for custom)

Value

Data frame with:

  • Original columns

  • Filled target column

  • {target}_is_imputed - Logical flag indicating imputed values

Details

The is_imputed flag allows downstream models to:

  • Filter out imputed rows if desired

  • Use the flag as a predictor to learn different behavior

  • Weight imputed observations differently

Gap-filling respects group boundaries and never fills across groups. Each group's time series is filled independently.

Examples

if (FALSE) { # \dontrun{
# Retail sales: missing = no sale
sales_filled <- fill_gaps(sales_data, target = "revenue", date = "date",
                          groups = "store", strategy = "zero")

# Sensor data: carry forward up to 3 missing readings
sensor_filled <- fill_gaps(sensor_data, target = "temperature",
                           date = "timestamp", groups = "device",
                           strategy = "locf",
                           params = list(max_gap = 3))

# Check imputation summary
table(sensor_filled$temperature_is_imputed)

# Use filled data in forecasting
ts <- TimeSeries(sensor_filled, date = "timestamp", groups = "device")
m <- fit(temperature ~ p(7) + rollsum(7), data = ts, model = lm)
} # }