Skip to content

TimeSeriesInferenceFormatter

Specialized formatter for time series inference data.

This class helps bridge the gap between raw time series data and the format required by the preprocessor during inference. It handles the unique requirements of time series features such- as: 1. Historical context requirements (lags, windows, etc.) 2. Temporal ordering of data 3. Proper grouping of time series 4. Data validation and formatting

For non-time series data, this formatter falls back to basic data conversion.

Constructor

__init__(self, preprocessor)

Initialize the TimeSeriesInferenceFormatter.

Parameters- preprocessor: The trained preprocessor model to prepare data for

describe_requirements

describe_requirements(self) -> str

Generate a human-readable description of the requirements for time series inference.

Returns

String with requirements description

format_for_incremental_prediction

format_for_incremental_prediction(self, current_history: dict, new_row: dict, to_tensors: bool = False) -> dict | dict[str, tensorflow.python.framework.tensor.Tensor]

Format data for incremental time series prediction.

This is useful for forecasting scenarios where each new prediction becomes part of the history for the next prediction.

Parameters- current_history: Current historical data

  • new_row: New data row to predict
  • to_tensors: Whether to convert output to TensorFlow tensors

Returns

Properly formatted data for making the prediction

generate_multi_step_forecast

generate_multi_step_forecast(self, history: dict, future_dates: list, group_id: str | None = None, steps: int | None = None) -> pandas.core.frame.DataFrame

Generate a placeholder frame for multi-step forecasting.

The returned frame carries one row per forecast step, with the sort column filled from future_dates and every time series feature set to NaN. Callers fill each row in turn with their model's prediction, so the row becomes part of the history for the following step.

Parameters- history: Historical data dictionary or DataFrame. It is validated

    against the minimum history each configured feature needs.
  • future_dates: List of dates for future predictions.
  • group_id: Optional group identifier (e.g. store_id) if using grouped time series.
  • steps: Number of steps to forecast. Defaults to every date in future_dates.

Returns

DataFrame with placeholder rows for each future step.

Raises

  • ValueError: If the preprocessor has no time series features, if the feature has no sort column, if steps asks for more rows than future_dates provides, or if history is too short for the configured lookback.

is_time_series_preprocessor

is_time_series_preprocessor(self) -> bool

Check if the preprocessor has time series features.

Returns- bool: True if time series features are present, False otherwise

prepare_inference_data

prepare_inference_data(self, data: dict | pandas.core.frame.DataFrame, historical_data: dict | pandas.core.frame.DataFrame | None = None, fill_missing: bool = True, to_tensors: bool = False) -> dict | dict[str, tensorflow.python.framework.tensor.Tensor]

Prepare time series data for inference based on preprocessor requirements.

Parameters- data: The new data to make predictions on

  • historical_data: Optional historical data to provide context for time series
  • fill_missing: Accepted for backwards compatibility and currently inert. This formatter never fabricates history: when the data is too short for the configured lookback it raises so the caller can supply real context. Missing values inside otherwise sufficient history are handled in-graph by MissingValueHandlerLayer, via the feature's missing_value_config.
  • to_tensors: Whether to convert the output to TensorFlow tensors

Returns

Dict with properly formatted data for inference, either as Python types or as TensorFlow tensors

Raises

  • ValueError: If the data cannot be formatted to meet time series requirements