[Feature Request]: Auto-Create Timestamps for Time Series Prediction

[aditisingh02]

Is your feature request related to a problem? Please describe.

When using the TimeSeriesDataset.prettify_prediction() method in flaml/automl/time_series/ts_data.py, if test_data is None, the function throws a NotImplementedError instead of automatically generating timestamps for the prediction output:

# ts_data.py, line 251-252
# TODO auto-create the timestamps for the time column instead of throwing
raise NotImplementedError("Need a non-None test_data for this to work, for now")

This is frustrating for users who want to make predictions without providing explicit test data timestamps, which is a common use case in time series forecasting.

Describe the solution you'd like

Implement automatic timestamp generation based on:

1. The training data's frequency - Infer the frequency from self.frequency (already available in the dataset)
2. The last timestamp in training data - Use self.end_date as the starting point
3. The prediction horizon - Generate the appropriate number of future timestamps

Example implementation approach:

if self.test_data is None:
    # Auto-generate timestamps based on frequency and end_date
    pred_timestamps = pd.date_range(
        start=self.end_date + pd.Timedelta(self.frequency),
        periods=len(y_pred),
        freq=self.frequency
    )
    y_pred[self.time_col] = pred_timestamps

Additional context

• The TimeSeriesDataset class already has self.frequency and self.end_date attributes that can be used for this purpose
• This would align with how other time series libraries (e.g., statsmodels, Prophet) handle predictions
• The existing create_forward_frame() function in the same file already implements similar logic for creating future timestamps, which could be reused

[eskayML]

looks like the TimeSeriesDataset.prettify_prediction() method should auto-generate timestamps when test_data is None. we can do this by inferring the frequency from self.frequency and using self.end_date as the starting point. should i open a PR for this?

[thinkall]

looks like the TimeSeriesDataset.prettify_prediction() method should auto-generate timestamps when test_data is None. we can do this by inferring the frequency from self.frequency and using self.end_date as the starting point. should i open a PR for this?

Hi, @eskayML, please go ahead to raise the PR. Thanks.