Quants see promise in Bayesian machine learning

Risk USA: probability theory may hold key to creating ‘self-aware’ AI

A 250-year-old mathematical theory could be used to create ‘self-aware’ machine learning systems that understand when they are out of their depth, according to a panel of senior quants.

Bayes’ theorem, named after the 18^th century UK mathematician Thomas Bayes, is widely used to infer the probability of a hypothesis holding true as more information becomes available.  

“The Bayesian paradigm allows you to actually get a hint that maybe the data the model has been trained on is not relevant anymore,” said Andrei Modoran, a senior data scientist at Rotella Capital Management, which uses machine learning and quantitative models to detect trends and price signals in financial markets.  

He said the approach could be applied to machine learning to avoid overfitting and detect when models are going awry.

“A lot of statistical modelling is based on understanding the data in the best possible way, using the model that fits the data in the best way,” said Modoran, who was speaking on a panel at the Risk USA conference on November 11. “In contrast, under the Bayesian paradigm, a set of likely models is selected according to the data, and then the agreement among those models is an indication of the degree of confidence in the prediction that the system is making.”

The confidence readings could also be used to detect tail risks and regime shifts in the market, “because the level of agreement suggests the historical data used to train the models might not be relevant to the current conditions,” Modoran said. If the models do not agree with each other, “that will be an immediate hint that something is a bit off.”

“What is most interesting for us,” he added “is that [the model] knows when it doesn’t know … under this paradigm, you get a signal that you’re not very confident about what’s going on.”

What that means is that the model itself has a self-awareness of when its predictions should be used. The model can help communicate when it’s no longer fit for purpose

Ben Steiner, Columbia University

Other panellists were enthusiastic about the potential of combining Bayesian approaches with machine learning. 

“I love that idea,” said Ben Steiner, a lecturer at Columbia University and former senior quant at BNP Paribas Asset Management. “Because what that means is that the model itself has a self-awareness of when its predictions should be used. The model can help communicate when it’s no longer fit for purpose.”

Max Gokhman, chief investment officer at AlphaTrAI, said the approach could also be adapted to direct more capital to the best performing models. “I think that’s where some of the Bayesian concepts really apply. We can have certainty that there is a higher efficacy for a specific model or algorithm. Then we can give it more of the risk budget for a specific period.”

Modoran cautioned though that Bayesian techniques required an optimal number of models – with the exact number depending on the complexity of the problem – and that the models should be as uncorrelated as possible.

He also suggested Bayesian approaches could find wider application in machine learning. To date, they have been used for classification and regression, rather than training. Modoran said recent advances in reinforcement learning and variational auto encoders could change that. “This will actually allow the use of Bayesian concepts on the avenue of the variational auto encoders straight to reinforcement learning and training systems. This is something at Rotella we are exploring.”

• Tweet  
• Facebook  
• LinkedIn  
• Save this article
• Send to  
• Print this page  

• Topics
• Bayesian modelling
• Risk management
• Machine learning
• Artificial intelligence