Intermittent Demand Forecasting in Scale Using Meta-Modelling (Deep Auto Regressive Linear Dynamic System)
- 데이터 사이언스, 머신 러닝 및 MLOps
- 소매 및 소비재
- 40 min
Demand forecasts of Items are integral to the health of Retail Operations, as accurate forecasts lead to improved decision-making and outcomes in replenishment, capacity, and resource planning. Granular Demand/Sales Forecasts at an Item-Store Level often yields more valuable insights and action plans but is often avoided due to the difficulty in predictions at this granularity, especially for slow moving items with temporal intermittencies. Such items are often categorised by long periods with no sales with seemingly random sales happening in between. Because of this behaviour, traditional statistical and time-series assumptions break; leading to prediction inaccuracies and over-forecasts. Granular Forecasts also increase the time complexity which is of primary importance when deploying in Scale. In this Presentation, we are going to speak about a novel Approach for Intermittent Demand Prediction using a Meta-Model Framework called Deep ARLDS (Deep Auto Regressive Linear Dynamic System), developed by our team in Walmart. The Approach entails the elicitation implicit features through Linear Dynamic Systems to map the underlying randomness, which along with explicit features; is fed into an Auto Regressive Recurrent Neural Network Architecture for obtaining the final forecasts. The Model is robust, scalable and yields far higher prediction accuracies than state-of-the-art; and uses shared memory transfer to provide forecasts for items with little or no history. The Model is currently scaled for around 35000 SKUs across ~250 Walmart Stores for making 14 weeks-ahead Demand Forecasts; using a CI/CD Pipeline with Parallel Processing in PySpark and NVDIA TESLA GPUs.