Description
This talk presents a systematic approach to improve model lifecycle management for machine learning models, using Python/DevOps tools and processes to enable self-served model development and deployment; a four-step workflow guides the developer/modeler to develop, automate, test and implement the model in a virtual environment in a production-ready manner.
Statistical/machine-learning modeling is often subject to regulatory mandates on model development, validation, deployment and monitoring. These mandates are liable to require that production models are fit-for-purpose, are accurately implemented and obey with local laws and regulations. However in many environment complying with such mandates is extremely onerous and time consuming, e.g. due to the use of legacy technologies and workflows or unstructured model development by data scientists.
This talk presents a systematic approach to this challenge, using Python/DevOps tools and processes to enable self-served model development and deployment; a four-step workflow guides the modeler to develop, automate, test and implement the model in a virtual environment in a production-ready manner. The approach is in use by the Quantitative Analytics team at Barclays bank, but we believe that the issues and potential solutions are relevant to the other regulated industries.