Skip to content

Podcast: Piterbarg and Antonov on alternate options to neural networks

The rise of neural networks in finance has been speedy and practically uncontested. The concept that such a synthetic intelligence is the most effective resolution for all method of issues is nearly casually accepted by quants. However there are some skeptics – one such being Vladimir Piterbarg, head of quantitative analytics and growth at NatWest Markets. Piterbarg jokes that his warnings about poorly thought-out functions of neural networks, which he has in comparison with a hammer in quest of nails, have given him a foul repute within the trade.

Regardless of many profitable functions in finance and particularly non-financial areas, Piterbarg argues quants too usually flip a blind eye to the drawbacks of neural networks, akin to coaching time, knowledge availability, and the predictability and interpretability of outputs.

He has spent the previous two years working with Alexandre Antonov, quantitative analysis and growth lead on the Abu Dhabi Funding Authority, on different approaches to fixing quantitative issues that don’t have the identical drawbacks.

On this episode of Quantcast, Piterbarg and Antonov talk about the 2 strategies they developed to approximate features which might be cumbersome to calculate – for example, as a result of they contain prolonged simulations – and to compute conditional expectations.

“We have provide you with a lot better and far sooner strategies for the standard issues that we see in finance,” says Piterbarg.

The 2 strategies are referred to as generalized stochastic sampling (GSS) and practical tensor practice (FTT). Antonov describes GSS as “a parametric illustration of the operate”. The strategy includes organising a grid of randomly distributed bell-shaped factors to characterize the multi-dimensional area of the operate. Randomly chosen factors are then used to rebuild the operate. The strategy mimics picture reconstruction and helps keep away from the curse of dimensionality that basic parametric fashions would encounter in a multi-dimensional area.

“If the operate is comparatively clean, then our methodology’s precision is greater [than that of neural networks],” Antonov explains. “All the issues we’ve got described for neural networks are roughly overcome.”

FTT reduces the dimensionality of a operate. Impressed by a 2011 paper by Ivan Oseledets, this methodology permits a 10-dimensional operate to be decomposed into the product of two-dimensional features. “That offers construction, explainability, and far sooner efficiency,” says Piterbarg

The 2 strategies may be mixed to “translate the issue from the operate being sampled on a stochastic right into a linear tensor practice downside”, says Piterbarg. That lightens the issue significantly, whereas nonetheless permitting for multi-dimensional features.

These strategies may be utilized to any downside described by features which might be sluggish to calculate, in addition to issues associated to the computation of conditional expectations, akin to payoffs of advanced monetary merchandise. The computation of derivatives valuation changes, which require the calculation of numerous current values ​​of simulated paths, is one other use case.

It is too early to understand how extensively this strategy might be adopted within the trade. “NatWest Markets is taking a look at it,” says Piterbarg, including the caveat that the analysis has solely lately been revealed and it takes time for banks to resolve whether or not to change to a brand new mannequin.

However he is optimistic that the trade will see the worth of GSS and FTT. “For a sure class of issues which might be vital in finance, this methodology beats neural networks arms down, there is not any doubt in my thoughts about it.”


00:00 Intro

01:48 The drawbacks of neural networks

08:00 The necessity for different approaches

10:38 Generalized stochastic sampling

17:40 Practical tensor practice

24:00 Combining the strategies in follow

30:10 Evaluating GSS and FTT to neural networks

31:40 Subsequent steps

To listen to the complete interview, pay attention within the participant above, or obtain. Future podcasts in our Quantcast sequence might be uploaded to danger.internet. It’s also possible to go to the primary web page right here to entry all tracks, or go to the iTunes retailer, Spotify or Google Podcasts to pay attention and subscribe.


Leave a Reply

Your email address will not be published. Required fields are marked *