Abstract

As artificial intelligence becomes increasingly embedded in financial services—supporting tasks such as credit scoring, fraud detection, forecasting, and real-time decision-making—its environmental impact is driven not only by model complexity, but also by the computational burden induced by repeated training, backtesting procedures, and high-frequency inference.
This talk adopts a statistical perspective on Green AI, focusing on how model specification, estimation strategies, and data handling procedures affect computational efficiency and, consequently, energy consumption and emissions. In particular, attention is given to data-efficient estimation pipelines, regularization and model compression techniques (e.g., pruning and quantization), and the trade-off between predictive accuracy and computational cost.
From a measurement standpoint, the contribution discusses the formalization of computational efficiency metrics, including runtime, memory usage, and energy consumption (e.g., kWh per training iteration or inference batch), and their integration into model evaluation frameworks. These quantities can be mapped into emissions estimates by incorporating region-specific electricity carbon intensity, using tools such as Eco2AI or CodeCarbon.
The talk also emphasizes the need to extend standard model assessment criteria (e.g., predictive performance, goodness-of-fit) to include computational and environmental dimensions, thereby promoting a multi-objective evaluation framework. In this context, reproducible and comparable reporting of efficiency-related metrics becomes essential for enabling statistically grounded comparisons across models and applications.
Overall, the aim is to contribute to the development of a unified framework in which statistical modeling, computational efficiency, and environmental impact are jointly considered in the design and evaluation of AI systems in finance.