Abstract: 

Dynamic price optimisation represents an increasingly profitable yet challenging process, especially for large and established businesses with long-standing practices and legacy data retention systems. Machine learning models built on often large amounts of sales data provide opportunities to grow revenue both by increasing price and reducing lost sales. David, Alexey and their team have developed approaches to solving such problems, from initial data exploration to cloud-based deployment. Key insights from this experience will be covered.

Initial exploration often covers statistical analysis of available sales data, stakeholder engagement and reverse engineering of legacy systems, with the primary aim of understanding the key feature sets to be used in the models. Given the frequent dominance of large categorical features, methods of encoding have been developed to best fit with the optimal base algorithm for each solution. Choices of models based on such a feature set have been explored including the option of chaining clustering, regression and time series algorithms, with particular focus on choices between regression trees and neural networks. With a view to rapid re-training during production, efficiency modifications have been made to the use of standard python library approaches with C++ extensions, and the merits of this versus distributed computing will be discussed.

Finally how the solutions have been deployed in a cloud environment will be summarised with a focus on reliability, scalability and user experience.

Bio: 

David is a senior data scientist at Altius. He has a PhD in particle physics and has spent ten years working in accelerator and detector physics at laboratories in the UK, US and China, where his research focussed on simulation and analysis of rare events along with designing and constructing high-frequency, low-latency data acquisition systems. At Altius he focuses on price optimisation solutions from initial data exploration through to robust cloud-based solutions.