Making Explainable AI a Reality for Quant Investment Management
Even though forecasting methods driven by AI have improved, transparency has not kept up. Techniques such as SHAP values (also known as Shapley Additive exPlanations and involving the attribution of a model's prediction to each individual feature) offer useful insights into how models make their decisions; their complexity makes it hard for institutional clients, regulators, and compliance teams to understand them. This results in a continuing difficulty in that the predictive outputs stay hidden behind technical details, which non-specialists find hard to interpret.
Over the past year, our team has created predictive features that are both interpretable and high-quality for use by institutional investors. Yet the problem that has always remained is getting the technical outputs into forms which analysts can easily understand, and compliance officers can confidently approve.
In order to deal with this, we worked with the University of Modena and Reggio Emilia (UNIMORE) and AImageLab, both related to the FF+ project funded by EuroHPC.
We created a specialised large language model for use in the financial sector, which converts forecast signals, SHAP attributions, asset metadata, and market context into clear and consistent narratives that are easy for humans to understand. The team adopted a strict methodology: evaluating 15 available open-source and commercial models using a common evaluation process and LLM-as-a-Judge scoring to determine the ones that performed best. The most effective models were then fine-tuned using efficient parameter methods, including LoRA and mixed-precision training, leveraging 48,000 EuroHPC GPU-hours. This method allowed for quick improvements while at the same time keeping costs and its environmental impact low.
The solution has now been prototyped in Axyon IRIS and is producing measurable results. Compared to the baseline models, the quality of the explanations has improved by as much as 55%. The addressable market has increased from about 10% to 25% of relevant firms, which in turn has allowed for wider adoption of AI-driven strategies. In terms of operations, the efficiency gains amount to roughly one FTE-month saved per client each year, and the benefits grow as adoption increases.
The project has also strengthened something more difficult to measure, namely a working relationship between UNIMORE–AImageLab and NCC Italy, which will form the basis for future collaboration in the areas of modelling, evaluation and HPC infrastructure.
In an industry in which trust is the product, the capacity to make a decision clearly could be just as important as the decision itself.
Would you like to learn more? Download the Success Story flyer and watch the video interview for more details: