Podcast: Iabichino on finance-native neural networks

UBS quant explains how to incorporate financial laws into an AI framework

The typical approach to designing an artificial intelligence tool for finance is to take existing AI architecture and adapt it for the required task. The danger with this method is that the output can sometimes violate financial principles, such as the no-arbitrage condition.

Stefano Iabichino, a director in the quant investment strategy team at UBS in London, has turned the problem on its head by building a neural network in which the laws of finance are woven into the network’s fabric. 

“The objective of the paper is to rebuild the relationship between AI and finance on asset-pricing first principles,” says Iabichino. “The result is a neural network in which each of its components obeys the laws of finance.”

His approach is based on a classical neural network architecture, a set of neurons and synapses distributed over multiple layers. Its peculiarity is in the role played by those components. First, Iabichino interprets the number of layers, or the depth of the neural network, as time, which is therefore bounded to the duration of the financial problem under consideration. Neurons represent the market states, or the levels of the variables that are relevant to that problem – for instance, each neuron may represent a possible price of an asset or an index.

Quantcast – a Risk.net Cutting Edge podcast · Risk Quantcast Stefano Iabichino 06/11/25

 

The next element is the activation function, on which Iabichino worked from scratch and came up with a special family of functions designed to capture financial problems. He named them Markovian activation functions, and as such they need to adhere to precise probabilistic and mathematical rules.

“Once we have defined these Markovian activation functions, we can start augmenting the semantic of each neuron with additional information. For example, we can include a numeraire, which is used for discounting,” he explains. The term Markovian, which indicates the independence from past data, hints at the fundamental principles Iabichino intends to capture.

Related

AI as pricing law 

A neural network tailored to financial asset pricing principles is introduced

This setup allows the creation of a neural path that, instead of connecting layers to layers, connects neurons to neurons, so constraining the possible output.

Iabichino believes well-known approaches such as deep hedging or VolGan are based on a flawed idea: “They try to bend finance to obey the logic of AI.” Models originally designed for static problems such as image or text recognition shouldn’t be trusted on dynamic systems like financial markets, he argues.

Iabichino also discusses the more practical and applied aspects of his neural network: its interpretability and its uses in banks’ trading desks, whether in the QIS business or as part of derivatives valuation adjustments, or XVAs.

In addition, Iabichino believes the method described in his paper can be also applied to the buy-side business. In that context, where the no-arbitrage condition doesn’t apply, constraints may be represented by investors’ views – which can then be compared with market-implied prices to generate trading signals.

Index:

00:00 Introduction: what’s a finance-native AI architecture?

02:58 How to incorporate financial constraints into a neural network

08:22 Differences with other AI-based approaches

13:36 Interpretability

14:53 Uses for QIS, pricing and investment purposes

23:48 The genesis of the idea

24:55 Iabichino’s research on XVA

To hear the full interview, listen in the player above, or download. Future podcasts in our Quantcast series will be uploaded to Risk.net. You can also visit the main page here to access all tracks, or go to Spotify, Amazon Music or the iTunes store to listen and subscribe.