Marc Nientker - Causal Forecasting: How to disentangle causal effects, while controlling for ...

Learn how to disentangle causal effects in forecasting while controlling for confounders. Discover methods to estimate elasticities & incorporate causal relationships into predictions.

Key takeaways
  • When optimizing business decisions, pure forecasting is not enough - you need to understand causal relationships between variables to make effective decisions

  • Unobserved confounding variables (factors that affect both decisions and outcomes but aren’t tracked) can lead to biased predictions and suboptimal decisions

  • In pricing optimization, factors like competitor prices, inventory levels, marketing events, and seasonality all affect sales but are often not fully captured in data

  • To handle unobserved confounders:

    • First estimate causal elasticities/effects
    • Remove these effects from historical data
    • Create forecasts on the normalized data
    • Add back the causal effects for final predictions
  • Cross-sectional and panel data methods can help infer causal effects even when you don’t observe all combinations of variables

  • For products with visible competition, discount effectiveness depends heavily on competitor pricing - relative position matters more than absolute price

  • Large retailers need to consider thousands of products and complex interactions between pricing, inventory, marketing and external events

  • Pure machine learning approaches that don’t account for causality can lead to biased optimization, even if forecasts appear accurate

  • Proper causal forecasting requires both:

    • Methods to estimate causal effects
    • Ways to incorporate those effects into predictive models
  • The process requires iterating between causal inference and forecasting steps - most existing tools don’t handle this workflow end-to-end