Buy-side quants of the year: Petter Kolm and Nicholas Westray

Risk Awards 2026: Study provides proof that simple is best when it comes to AI in finance

Petter Kolm and Nicholas Westray

Photo Nicholas Westray: Alex Towle

Petter Kolm and Nicholas Westray win this year’s award for research that shows how, when it comes to machine learning in finance, details matter, and simple is often best.  

In their paper – Deep learning alpha signals from limit order books: practical insights and lessons learned — Kolm and Westray tested the ability of four types of deep neural network to forecast high-frequency stock returns using limit order book data. 

As well as asking whether simpler or more complicated models did a better job, Kolm and Westray sought to determine how far different design choices might affect results.

Their intuition was that simple decisions, such as how to organise data and how far back in time a model should look, would matter as much or more than choosing the most cutting-edge neural network setup.

This work shows that details really matter…. There are a bunch of design issues. There’s more to machine learning than picking the right model

Petter Kolm

The hunch proved right. “Some of these fancy, very deep models, for this prediction task on order books, don’t really help,” Kolm says.

Instead, the two quants advise buy-siders not to overlook some basics: to keep model architectures simple, to time-stamp data and to pay attention to real-time order flow conditions to know how reliable a model’s forecasts are likely to be.

Massive machine learning models are making headlines in other domains. Large language models such as GPT-5 have trillions of parameters. But the size of models may matter less in finance because of the limited data available to train them, Westray says.

“If you take 10,000 images of cats, they all look look fairly similar. If you take ten thousand situations where a price went up, the order book slices preceding that could be really different.”

Kolm says the work shows the importance of looking past the clever tools now at a quant’s disposal. “People often think: ‘I have a black box, some kind of machine learning model. I stick data inside it, and I shake the box around, and something useful comes out,” he says. “If something useful comes out, I declare success. If it doesn’t work, I try a different machine learning model. 

“This work shows that details really matter…. There are a bunch of design issues. There’s more to machine learning than picking the right model.”

Keep it simple

Readers of the paper come away with a collection of practical tips.

Telling a model when updates to the order book occurred, or the duration between them, turned out to be “probably significantly more important than getting the latest and greatest deep neural network”, Kolm says. 

The quants found that transforming raw order book data themselves – from non-stationary to stationary – resulted in better outcomes than using sophisticated models that could learn the transformation autonomously along the way to making predictions.

Nicholas Westray

Photo: Alex Towle

When the models looked further back in history – at the 1,000 most recent order book updates, say, rather than 500 – performance deteriorated rather than improving.

The two quants also tested whether the characteristics of an individual stock’s order-book might make returns easier to predict. They found machine learning seems to perform best with stocks that have busy order books, measured by the number of updates.

The same holds for stocks with more stable buying and selling. “When you are around a price change in the order book, the order book starts to get reshuffled. People cancel orders. People post new limit orders, which means some of the structure you may have learned gets destroyed,” says Kolm.

This also helps explain why simple models may be better, he adds. “We think it’s partly the reason why some of the sophisticated encoder-decoder transformer-based models don’t add additional value. If the stocks have frequent changes, you need to reset your prediction. So it doesn’t help to look very far back in the history.”

Questions

Kolm studied mathematics at ETH Zurich and Yale before working for Goldman Sachs Asset Management in the early 2000s in the firm’s Quantitative Strategies Group. Since 2007 he has worked as a professor at NYU’s Courant Institute of Mathematical Sciences, where he also runs the Masters in Mathematical Finance programme. Much of his research has focused on machine learning, particularly natural language processing.

Westray completed his PHD at Imperial College in London and worked for Deutsche Bank in London and New York before joining Citadel for four years. In March 2021 he joined AB to lead execution research and work on deep learning. He joined Cubist, the quantitative investing arm of Point72, in March 2024 to work on high-frequency execution.

Westray won Risk.net’s buy-side quant of the year award in 2023 together with Kevin Webster of DE Shaw for work using machine learning to de-bias transaction cost analysis data. 

Related

Deep learning alpha signals from limit order books: practical insights and lessons learned 

An analysis on network architectures applied to limit order book data is presented

Kolm and Westray first met in New York in 2013 at a piano evening fundraiser, hosted by Baruch College’s Jim Gatheral, but only started to work on research together in 2020 during Westray’s gardening leave between working at Citadel and AB.

In 2021 they published a first paper on machine learning and stock return prediction – Deep order flow imbalance: extracting alpha at multiple horizons from the limit order book – that has since been downloaded more than 11,000 times.

That work made plain that practical questions about how to implement machine learning in finance remained unanswered. Research from other fields had tackled the queries. But how far such findings would read across to problems like predicting stock returns was far from clear, Westray says. The world outside finance he characterises as broadly a “high signal to noise domain”. Finance, he says, is “the polar opposite”. 

At the same time, most of the existing studies of machine learning in finance looked at small samples of stocks. “From a practitioner’s perspective, that’s frustrating,” Westray says. “People would typically have a model and test it on, say, four symbols. I think if you want to do this at scale in live trading, a typical universe in US equities might be a thousand or three thousand symbols.” 

Kolm and Westray tested the accuracy of four models in forecasting returns over multiple horizons, using order book data for 115 stocks.

The simplest model in the study was a long short-term memory network – a type of neural network used in many day-to-day applications such as language translation. The most complex model tested – a so-called Seq2Seq model, with attention – has been designed specifically to learn from and output sequential data, enhanced with a greater ability to focus on the most informative parts of the input data.

In November, Researchers from Scuola Normale Superiore, Pisa and the Universities of Bologna, Oxford and Toronto, published a paper that formulates trading strategies using the types of signals Kolm and Westray generated.