2024-07-11 –, Else (1.3)
A lot of industry-available Machine Learning solutions for causal forecasting have a very particular blind spot: unobserved confounders. We will present an approach that allows you to combine state-of-the-art Machine Learning approaches with advanced Econometrics techniques to get the better of both worlds: accurate causal inference and good forecasting accuracy.
Causal Forecasting is a very hot topic in the industry with many applications ranging from marketing spending to pricing. Disentangling causal effects from spurious correlations plays a key role when forecasts are used for decision making, such as in the case of pricing. Solutions available in the industry typically rely on Machine Learning methods that use techniques like DoubleML, Transformers, LSTM, and boosted tree algorithms. A common shortcoming of such solutions is that they do not account for the existence of unobserved confounders, such as world events, or other hard-to-measure effects that can bias the measurement of causal effects. We showcase a solution that was developed over the last 3 years that addresses these challenges by combining advanced Econometrics methods with ML techniques. The case-study will focus on the example of retail pricing, but the solution is broadly applicable and it has been tested in different settings, including airline pricing.