AR-MDN for e-Retail Demand Forecasting
summary
AR-MDN combines feature embeddings, an MLP front end and LSTM recurrent layers to output probabilistic forecasts as a learned mixture density (Gaussian mixtures) so the model predicts a full sales distribution and supports Monte Carlo sampling. It trains end-to-end on a year of Flipkart data across tens of thousands of SKUs, models price and visibility and uses product embeddings to share signal across sparse SKUs, outperforming Cubist and ARIMA baselines with a reported significant accuracy gain.
AI-written summary · about
chapters (26)
- 0:00Introduction
- 0:34Background
- 0:57Motivation
- 3:11Data
- 4:27Goal Seek
- 5:07What happens amount
- 6:18Engineering
- 6:42Forecasting
- 7:18ARIMA Model
- 8:00Cubist Model
- 8:23Methods
- 9:18Main Topic
- 13:19Non Gaussian Distribution
- 16:59Notation
- 20:40Optimization
- 21:20Gaussian likelihood
- 22:18Multimodal distribution
- 24:39Product embedding
- 26:59Model training
- 28:54Cross entropy loss
- 29:10Experimental Evaluation
- 30:05Summary of Performance
- 30:31Category Based Performance
- 30:46Conclusion
- 33:49Conclusions
- 35:17Future Inputs
related talks
28mMastering Demand Forecasting: Lessons from Europe's Largest Retailer
30mAI in Retail - Advancing Promo Demand Forecasting
08mMachine Learning for Bangladeshi Retailers
22mFrontiers of Machine Learning from the Lens of Retail
23mMachine Learning with a Deep Dive into Automated Forecasting System
35mBuilding a successful inventory planning solution with Amazon Forecast
47mBuilding ML practices to address the top four use cases
37mForecasting using machine learning methods