August 26, 2024 — Easier Training of Forecasting Models from Real-World Data

We are pleased to announce that the Time-Series Forecasting ML Function now includes preprocessing features that allow you to successfully train a forecasting model even when your training data has missing, duplicate, or misaligned time steps. In the past, such issues, which are common in real-world data, typically prevented the model from being trained. These features are:

  • You can manually specify an event cadence in case the model fails to infer it or infers it incorrectly

  • The model can interpolate missing target values from nearby time steps.

  • The model can aggregate dimensional values from events occurring outside the canonical event cadence in a number of ways, and you can specify aggregation behaviors for the type of value or per column.

A relatively small number of such corrections does not noticeably affect prediction accuracy.

For more information, see Dealing with real-world data in Time-Series Forecasting.

Language: English