--
This is an ongoing list of resources I've found helpful. Hopefully I'll organize it more later
Favorite Resources
- Zalando - Deep Learning Based Forecasting
- They use Multinomial imputation based on other SKUs. Didn't quite work for the dataset I had access to due to extreme censoring.
- Zalando's double ML pricing paper
- Discusses how to account for and model pricing in an unbiased way for demand models
- https://github.com/FlorianWilhelm/bhm-at-scale
- Implements a hierarchical forecaster at scale and offers one approach of accounting for censoring
- Modern forecasting in practice course
- Great course in general about forecasting. They discuss just not contributing to the log likelihood as a simple approach for accounting for stockouts, or imputing based on other SKUs
- Logit choice model
- Shows how to use discrete choice modeling in a way that could work for censored demand - basically adjusting the choice set and modeling utilities
- Discrete Choice Models with Simulation
- Provides a really nice framework for how to think about stockout substitution, product utilities, and cross-product effects
- The poisson trick a way of estimating Multinomial distributions as a series of poisson models
- Survival Convolutions (for modeling inventory returns and making simulations vectorized)
- Dan Marthaler's work on Wallenius Multivariate Hypergeometric distribution in Stan - a better alternative to a "censored multinomial" distribution, although tough to work with
- Evaluating Survival models
Other resources on censored demand
- Some papers
- https://par.nsf.gov/servlets/purl/10066022
- https://www.lancaster.ac.uk/stor-i-student-sites/tessa-wilkie/wp-content/uploads/sites/14/2020/03/Wilkie_RT1.pdf
- https://arxiv.org/pdf/1810.09166.pdf#:~:text=There%20is%20a%20wide%20range%20of%20econometric%20models%20that%20has,negative%2C%20leading%20to%20zero%20purchases
- Has some good resources referenced, and attempts to use ML
- Blog posts
- https://medium.com/afresh-engineering/accounting-for-censored-demand-in-sales-forecasting-bdc77192802d
- Uses imputation, which works when the censoring rate is low but probably doesnt work under extreme stockouts
- Nicholas Vandeput Blog: Talks about censoring at a high level
- https://medium.com/afresh-engineering/accounting-for-censored-demand-in-sales-forecasting-bdc77192802d