Forecasting Prices With Causality
Artificial Intelligence & Machine Learning
“Causal Forecasting for Pricing,” a team of researchers from Zalando SE, including Douglas Schultz, Johannes Stephan, Julian Sieber, Trudie Yeh, Manuel Kunz, Patrick Doupe, and Tim Januschowski, present an innovative approach to demand forecasting within the context of retail pricing. Published on January 24, 2024, this work impressively stands out for integrating Double Machine Learning (DML) with advanced transformer-based forecasting models.
Methodology?
The primary focus of this research is modeling the causal relationship between pricing and demand, a critical factor for retailers who aim to optimize pricing for profitability. The paper introduces a model-agnostic Full Graph Double Machine Learning training framework, which is instrumental in training Structural Causal Models. This framework is designed to be versatile across various causal AI modeling scenarios.
- Integration of DML and Transformer Models: The fusion of DML for causal inference and state-of-the-art transformer-based forecasting models is a significant advancement in the field.
- Realistic Synthetic Data and Real-world Application: The researchers demonstrate the effectiveness of their method through extensive experiments, both in controlled environments with synthetic data and in real-world settings.
Findings?
- Off-Policy and On-Policy Performance: The method showcased superior performance in off-policy settings, meaning it effectively adapts to changes in pricing policy. While it slightly trails in on-policy settings, where pricing policies remain constant, its overall efficacy is notable.
- Causal Effect Estimation: The approach provides a more accurate estimation of causal effects compared to traditional forecasting methods that do not incorporate causal understanding.
Data
The researchers adeptly manage time-series data, considering variables like demand, discount rates, and article-specific covariates, including past demand and catalog information. This comprehensive approach allows for nuanced demand forecasting.
Implications and Future Directions?
- Probabilistic Treatment and Inverse Propensity Scores: The paper suggests that future work should include probabilistic approaches and the incorporation of inverse propensity scores for enhanced flexibility and accuracy.
- Expansion of Forecasting Models: Further research is recommended to include more flexible outcome models and the integration of multivariate forecasting models.
“Causal Forecasting for Pricing” by the team at Zalando SE is a groundbreaking paper that significantly contributes to the fields of retail forecasting and causal inference. By marrying DML with transformer-based models, the researchers have paved the way for more sophisticated and accurate demand forecasting in retail, especially in dynamic pricing contexts. This paper not only addresses a crucial gap in current forecasting methodologies but also sets a foundation for future explorations in the domain.
Read the full paper: 2312.15282.pdf (arxiv.org)
For more information, the full paper and additional resources are available at: Zalando Research.
