Concept | What if? analysis for time series

Definition

What if? analysis is a technique used to evaluate how variations in input parameters impact model outputs. In the context of time series forecasting, it involves modifying one or more variables that influence the temporal process and assessing how forecast trajectories respond to these changes.

In classical ML, What if? analysis is static and immediate. In time series, What if? analysis is dynamic and temporal: changes affect not only the current value but also propagate into future periods due to the inherent dependencies of sequential data.

In Dataiku, this feature is available after training a model for a time series forecasting problem. When you change values, you can directly see the changes reflected for the forecasted values once you compute them.

If you are predicting multiple time series, you can:

• Switch between time series identifiers.

• Modify values across time steps on the fly. Dataiku will highlight the changed values for easy tracking.

Advanced scenario features

The What if? analysis tool offers advanced capabilities for comprehensive testing and historical validation:

• Save scenarios: Save a set of modifications as a scenario. This way, you can browse and revisit scenarios to observe the effect of potential outcomes over time.

• Compare multiple scenarios: Create several scenarios for the same time series and visualize them side by side to assess their impact on future forecasts.

• Use historical data: Set the analysis window to a historical date to apply What if? changes to past time steps within the model’s training data. This allows you to explore how different historical input conditions would have impacted the time series forecasts, providing a deeper understanding of the model behavior and sensitivities.

What’s next?

To learn more on time series in Dataiku, see Tutorial | Time series forecasting (visual ML).