Machine learning teaches asset traders not to sweat the small stuff

Financial markets are governed by a combination of rational and irrational forces, statistical probabilities, and "animal spirits." It takes fluency in both to understand the market, let alone to beat it. Yet market actors, including asset traders, now frequently use machine-learning techniques to help generate predictions of future asset prices.

Scholars such as Bo Hu, assistant professor of finance at the Costello College of Business at George Mason University, are researching how these machine-learning tools are changing the decision-making processes that move the market, for better or worse.

The subject of Hu's recent paper in Management Science is a well-known machine-learning technique called LASSO (least absolute shrinkage and selection operator), which has been widely adopted by financial practitioners since its introduction in 1996 by statistician Robert Tibshirani.

"If you look at the original paper, it describes an approach created by adding a regularization penalty to the least-squares regression method," Hu says. Translation: "The power of LASSO is that it can screen out (i.e., penalize) weak signals while capturing stronger, potentially profitable ones. A LASSO-type trading strategy involves an 'inactive zone' for smaller-scale activity, in which the trading strategy is to do nothing."

The paper was co-authored by Wen Chen of Texas Tech University and Liyan Yang of the University of Toronto.

Despite LASSO's popularity and power, the soundness of its economic rationale remains unclear. Traders are presumably seeking any edge, however small, in the pursuit of outsized returns. How could it make sense for them to adopt a system designed to relegate signals of lesser magnitude to an ignored "inactive zone"?

To resolve this question, the researchers developed a theoretical framework to model a financial market in which multiple agents (read: traders) use an asset's price history to forecast its return and make trading decisions.

In the benchmark case, when traders know the trading environment and do not face model uncertainty, they act according to an alternative to LASSO known as MSE (mean squared error). "MSE is essentially a Bayesian learning approach grounded in economic rationality," Hu says. "It means that rational agents use Bayesian learning to update their beliefs and design their trading strategies. That stands in stark contrast to LASSO estimation, which filters out weak signals."

However, the researchers found that when traders faced substantial ambiguity about the distribution of asset values, the trading calculus shifted. Ambiguity-averse agents will adopt a robust, LASSO-like strategy, refraining from trading in response to weak or intermediate market signals. With linear constraints imposed on the allowable trading strategies, the equilibrium decisions exactly matched LASSO estimates.

As an equilibrium trading strategy, LASSO can improve aggregate profits relative to the edge-seeking Bayesian alternative in the benchmark case, because the more conservative positions dictated by the "inactive zone" soften competition among traders. In a large market, aggressive competition drives the aggregate profits of traders using the conventional MSE strategy toward zero. By trading less aggressively, LASSO traders may preserve positive aggregate profits-a mechanism the researchers describe as "implicit collusion," even though the traders do not communicate or explicitly coordinate.

However, Hu underscores that LASSO's usefulness as a hedge against ambiguity depends on how well traders' beliefs match the market's true distribution of risk. The profitability of a LASSO strategy therefore hangs in the balance between traders' biases and their enhanced market power (due to LASSO's conservatism).

This balance is especially important when traders must distinguish between temporary fads and persistent trends. "Recently, the semiconductor index had a record-breaking run of consecutive gains. That's very strong momentum, but if you are a contrarian, you might want to bet against that trend," Hu says. "A trader must decide whether to follow the trend or take a contrarian position. LASSO's inactive zone can help prevent overreaction to weak evidence, but it may also delay action when an emerging trend is genuine."

There is also the possibility that a LASSO strategy could increase market volatility when combined with the other objectives and constraints of market makers, including high-frequency traders. "These traders need to control their inventory," Hu says. "If they follow a LASSO-type strategy with an inactive zone, they will accommodate the liquidity demands of the market until their inventory reaches a certain threshold. At that point, they may start trading like momentum traders and cause the liquidity dry-up in financial markets. This dynamic is part of what fuels incidents like the 'flash crash' in 2010."

In This Story

Topics