Large Scale Neural Probabilistic Forecasting

Research Project by Dr. Nikolay Jetchev:

An e-commerce company like Zalando is a complex ecosystem of millions of customers and products, which makes efficient forecasting and risk assessment of future KPI values a challenge. Most existing methods have drawbacks: they either do not model the inherent probabilistic nature of the problem and provide only point estimates without regard of the uncertainty, or they do model it but cannot deal with the scale of big data. To address that, we will work on efficient large scale approaches to probabilistic time series modeling. In particular, we focus on a generative autoregressive approach, utilizing state-of-the-art recurrent neural networks and appropriate probabilistic modeling distributions.

Large Scale Neural Probablistic Forecasting: Event


Why is this of interest for Zalando? It is one of our research goals to create a general modeling toolbox capable of delivering large scale accurate forecasting for various Zalando KPIs. As a showcase of our capabilities (and potential product), we are working on a generative model of the customer evolution which can be used to model customer lifetime value (CLV), defined as the net cash flow. Our method is able not only to give point estimates of the CLV for 180 days ahead for millions of customers with improved accuracy, but also estimates the uncertainty in these estimated values and provides analysis into the dynamics of customer order and return behavior.

Nikolay Jetchev – Large Scale Neural Probabilistic Forecasting Scheme

Why is this of interest for the scientific community? Neural Autoregressive Denisty Estimation (NADE) and Long Short Term Memory (LSTM) neural networks have been recently a hot topic on major conferences. Even though they are often used for image generation tasks, their application to time series forecasting has not been fully explored. In our work we use various modifications to NADE-LSTM that will give new insights in that direction, e.g. probability distributions for sparse heavy tail data modeling and large scale approaches to Monte Carlo inference. Publishing a paper – in the time series, forecasting or deep learning fields – will boost the prestige of Zalando research in the academic community and showcase our excellence in the tech e-commerce field.