Neural network learns ‘universal model’ for stock-price moves

Relationships between order flow and price “are stable through time and across stocks and sectors”

Imperial College’s Cont: “The relationships between order flow and price are universal and stationary”

Academics have used machine learning to create a “universal” model for predicting short-term stock-price changes – disproving common assumptions among dealers, hedge funds and high-frequency traders about how such models should be built.

In a recent study, Rama Cont, a professor at Imperial College London, and Justin Sirignano, assistant professor at the University of Illinois at Urbana-Champaign, used a neural network trained on two years of intraday data from Nasdaq’s limit order book to ‘learn’ how order flow affects movements in stock prices.

It disproves the assumption among market participants that intraday pricing and risk models should be based on sector-specific data and relatively short histories, Cont said, and could lead to more accurate models including for stocks with relatively sparse data.

“The relationships between order flow and price are universal and stationary,” Cont said. “The large-scale application of deep learning analysis [has] uncovered a universal model. The relationships are stable through time and across stocks and sectors.”

Cont was speaking at Risk’s Quant Summit Europe event in London on March 8.

The academics constructed a neural network – a type of machine-learning system designed to replicate networks of neurons in animal brains – with the aim of predicting whether stocks would move up or down from a given point in time.

The study comes at a point of growing interest from buy- and sell-side firms in monitoring and hedging intraday volatility, particularly for highly leveraged strategies that could be vulnerable to short-term asset price moves. 

The academics constructed two models: one trained on single-stock data, the other – a universal model – trained on data from all stocks. The single-stock model achieved an improvement in accuracy of more than 5% compared with the type of conventional linear models often used by dealers, hedge funds or high-frequency traders.

The universal model performed still better, with close to 70% accuracy on average, compared with about 60% on average for stock-specific linear models.

The universal model was also more effective than the single-stock neural network model in predicting the behaviour of stocks outside the dataset on which it was trained, and in predicting behaviour several months after the training data ended.

That means it could be applied even to newly listed stocks for which there is no price history, Cont said. 

The universal model benefits from learning from data across sectors because the dataset is richer, including more episodes of stress in individual stocks, for example, he said.

The study also showed that two-hour-old order-book information affected price formation, much older than is commonly assumed in models, Cont explained. Even for tick-by-tick prediction, there was a significant increase in accuracy when the model’s inputs were moved from a few minutes of data to two hours of data.

“There is some empirical evidence that shows there are common features [in price formation] across different assets but in empirical modelling and market practice, people don’t really buy that,” he said.

“They only use price history and data from stock A to model stock A. They don’t typically use data from some stocks to predict volatility for others. They’re not sure how to map one asset to another.”

“This model can be used for out-of-sample prediction and the prediction is stable even months out of sample and even for stocks that are not in the sample,” Cont concluded.