London, England, United Kingdom
You can find my blog on all things maths, stats, probability and modelling at: https://lewiscoleblog.com/ I am a highly motivated professional with strong mathematical and technical skills. I try and keep a wide variety of technical interests as I believe ideas from one discipline often cross-over to another. Some of my interests include: Monte Carlo methods, agent based modelling, machine learning ("AI"), Bayesian statistics, extreme value theory, information theory, mathematical biology and theoretical neuroscience.
As head of AI research and principal data scientist I am responsible for creating the AI roadmap, designing the “next-gen” of Sensa’s modelling capabilities, overseeing client engagements and leading a team of (currently) 11 data scientists. The team's research agenda includes: - Methods for Reliable and Trustworthy AI - how can we get predictive and generative models to tell us when we should rely on them and when we should ignore them? How can we bridge the gap between confidence based statistical methods and ML models? How can counterfactual analysis be used to assess model performance? - Agent Based Modelling (ABM) - how do we develop the industry leading ABM for the modelling of financial crime behaviours? - Behavioural Science - how can we present outputs of models and analytics to improve their efficacy? How can we leverage the way people make decisions to improve workflows for model users? - Hallucination Bounding - how can we estimate the rate of hallucination of a language model prior to deployment? - Computational Optimization - can we re-write crucial routines in C/C++/Cuda to enable more advanced analytic capabilities? - Operational Research - many problems can be reframed as resource allocation problems, can we apply established methods in operational research to the fin-crime problem?
As a senior data scientist at SymphonyAI Sensa (formally Ayasdi AI) I developed the 4 main models within the flagship SensaAML anti-money laundering product. Using these models on client POC projects I managed to secure market leading analytic results (in one such POC the model achieved 78% performance – while a competitor was unable to produce results, or effectively a 0% performance). These results led to the first sales of the product for the company.
As part of the corporate risk team I am responsible for the design, development and maintenance of quantitative risk modelling for all non-property non-natural catastrophe lines. During my time in this role I have designed a highly sophisticated one-of-a-kind modelling architecture. This allows the business to understand its risks in a “real world” context unlike other widely used methods and allows many disparate models to live within a unified framework. Within this architecture I have created models for credit, financial, surety, pandemic, aviation and marine risks – through time the coverage is expected to increase. Currently the entire model covers c.500k individual modelled risks in total and is expanding. The code is optimised in NumPy/Numba/Cython. As part of understanding the universe of risks facing the company, I have developed many prototype and proof of concept models these have ranged from fuzzy-logic, graphical models and MCMC type approaches to financial modelling, PDE based models and some agent-based modelling. I am also responsible for aggregating, analysing and presenting the modelled results to all audiences ranging from analyst through to board level. In addition, I am responsible for the production of detailed documentation for both internal and regulatory purposes, and where appropriate present key methodologies to the necessary stakeholders. Outside of my core modelling focus I have also worked on projects including: modelling a complicated reinsurance treaty resulting in the Lloyd’s platform saving $50m+ in capital requirement a year, developing a universal model uncertainty quantification technique, various machine learning projects including an “automated underwriting” tool for monitoring delegated authority, I have also worked on natural catastrophe modelling (including GIS data) and some data pipeline work as required. I also acted as “AI/ML advisor” to the company and assessed many fintech offerings in the specialty insurance space.
At Montpelier I was focussed heavily on capital modelling, being part of a very small team I was able to, very quickly, gain an understanding of all aspects of capital modelling. Within 6 months of joining I took complete ownership of the day to day running and calibration of the model, including the ESG calibration and investment portfolio modelling. It was at this time the company completely rebuilt the capital model from scratch, allowing me to gain an insight into the complete life cycle of product development, high level and low level testing. Following the rebuild of the calculation kernel I redesigned the model calibration process in order to reduce the time required to calibrate the model from months to days. Other key projects included: • Ownership of model coding/development for the London platform • The creation of a natural catastrophe simulation engine using RMS ELT output. This included: modelling of inwards reinsurance treaties, allowance for secondary uncertainty correlation, ability to model multiple accident years and hurricane seasonality • The creation of a natural catastrophe model for perils outside of the scope of a traditional vendor catastrophe model such as AIR or RMS • Development of a multi-year standard formula type capital forecasting tool aiming to bridge the gap between the EIOPA standard formula and the internal modelled view • A review of capital allocation methodologies and methods/metrics used for assessing performance and risk
• Within the actuarial support team, I was responsible for the production of annual reports and triennial valuations for trustees of various DB pension schemes. I also produced scheme correspondence and calculate strains put on schemes by early retirement of members.