Quantcast – a Risk.net Cutting Edge podcast Risk Quantcast Stefano Iabichino 06/11/25
Nov 18, 2025
Stefano Iabichino, Director of the QS team at UBS, dives into the intersection of AI and finance. He discusses designing finance-first neural networks that adhere to no-arbitrage principles. Stefano critiques conventional AI for its pattern reliance, emphasizing the imperative of no-arbitrage across market regimes. He highlights the use of each neuron as a representation of future market states and explores practical applications in hedge funds, including learning market probabilities and innovative stress-based risk management. His research aims for coherence in financial AI models.
AI Snips
Chapters
Transcript
Episode notes
Network Depth As Time And Market-State Neurons
- Stefano reframes network depth as time and ties neurons to explicit market states to enforce financial semantics.
- Markovian activation functions replace ReLU/Sigmoid to ensure coherence with stochastic dynamics and Finetti principles.
Forward Pass As Monte Carlo Neural Paths
- The forward pass is recast as Monte Carlo-style neural paths that transition neuron-to-neuron using Markovian activations.
- Activating a neuron triggers its entire market semantic, including risk factors, numeraires and path-dependent asset state.
AI Patterns Clash With Finance Laws
- Stefano contrasts AI's pattern focus with finance's law-driven nature, calling naive outputs 'economic hallucination.'
- He argues models must embed the no-arbitrage (martingale) principle rather than enforce it after training.