Point-in-time for prediction models

Point-in-time support allows you to avoid look-ahead bias in your alternative data analysis workflows. We have now rolled out beta point-in-time support to Prediction Models, which use machine learning techniques to help forecast company- or macro-level KPIs based on a set of input signals.

Point-in-time evaluation is now available for the "Prediction Model (ML)" and "Classification Model" types.

To use this, simply turn on "Evaluate with point-in-time data" under Input signals, when configuring your prediction models.

As point-in-time evaluation can lead to significantly longer model run times, we recommend doing early testing and iteration without point-in-time, before enabling it for final model training.

Under the hood, when you enable point-in-time evaluation, for each data point in the target, we evaluate the input signals as of the date of that data point. For example, when making a prediction of a data point on 31 December 2021, the input signals are evaluated as they were known on 31 December 2021.