Takeaways from Quant Conferences – London 2025

I’ve recently assessed the current sentiment of quantitative professionals (quants) by attending Quant Strats 2025 and QuantMinds 2025 and shared some views in a podcast. While business cards are becoming out of fashion (sigh), in-person interactions can provide enjoyment and perspectives that are hard to find on Teams. From chatting with peers during networking breaks, to attending presentations of industry leaders, attending and speaking at such conferences has provided quite a fruitful experience. 

Nicole Konigstein (left), Federico Fontana (centre) and Alisa Rusanoff (right).

Context. XAI is a Commodity Trading Advisor (CTA). We trade futures systematically and directionally across multiple asset classes using an innovative investment framework. The takeaways below are presented from that perspective.  Some topics are timeless across quant conferences: portfolio construction, risk management, diversification, derivatives pricing and data. Most asset classes, instruments and investment time frequencies were covered at some point (you can name it). The more dynamic topics relate to niche use cases or discussions around the latest AI trends. 

Latest trends. It was evident that traditional machine learning techniques are now part of the arsenal of many firms. Large Language Models (LLMs) and generative AI were buzzwords at conferences last year, and this year has been no exception. While in previous editions the focus was on potential applications of LLMs, this year the discussion was more grounded in sharing experience from practical use cases and model biases. It’s clear that the rate of adoption has increased for auxiliary parts of the investment process, including: processing of internal documents and analyst reports, processing of text data to usable inputs for traditional investment pipelines, data augmentation and labelling. The overall effect is a speeding up of research and development cycles of quants. Most of the code will be written by LLMs and humans will become debuggers of AI-written code, an industry leader said at the conference. This dystopian future was also accompanied by a general scepticism about using non-linear models or end-to-end machine learning approaches in the investment process. We believe that it’s a fair position to hold for many, when tools evolve so dynamically and it is easier to overfit than underfit financial data. We also perceived frictions in larger organizations as architectural changes may be harder to deploy in production. At XAI it is quite different as we are a lean team and have adopted machine learning as a tool from the ground up. Models require data to be trained, so this brings us to the next hot topic. 

Data. While investing remains fundamentally a time series problem that often starts with price data, alternative datasets may help to complement the informational content. Many professionals shared the struggle to obtain reliable point-in-time data and metadata, necessary to perform a historical evaluation of alternative datasets. Data reconciliation and consolidation were other pain points that were mentioned at the conferences. I believe that we will see evaluation processes gradually shifting from back-testing to forward-testing as a way to bypass issues such as short time histories, or methodological discrepancies between historical and real-time data feeds. After all, if an alternative dataset has value, it is our job to build the technological infrastructure to support it, and not the other way around.

Model complexity and explainability. Once the investment problem is framed correctly and the data is clean, one has to decide between linear models (robust, fast and simple) and non-linear models (more flexible, harder to calibrate) to map input data to something more actionable. Non-linear models are also harder to interpret without dedicated tools. Model explainability can come from building narratives using bottom-up or top-down approaches, dashboards, SHAP-like analysis and surrogate models. I do not recall hearing the buzzword “black box” this year. Because giving up model explainability is not an option, this could indicate that quants have moved on and now use a variety of tools to address model explainability. 

I have recently discussed some of the themes above in a podcast hosted by Harrington Starr’s FinTech Focus TV, recorded live at Quant Strats 2025.

Conclusion. We experienced firsthand that it takes the right culture, integrity and technical skills to innovate investment processes. Quant Strats and QuantMinds have revealed that the industry is in constant evolution.