Machine learning is ubiquitous in financial markets, and we tend to associate it with the sexy parts of hedge fund investing: what to buy, what to sell and how to pick a bottom or a top. But if you run a multibillion-dollar quantitative investment strategy with high turnover, the operative question is frequently, “How do I avoid shooting myself in the foot with poor trade execution?” In that sense, predicting volume can be just as important as predicting prices. If you place a buy order on a low-volume trading day and it causes the share price to skyrocket, it might not matter how clever your investment thesis was.
That’s the subject of the paper “Trading Volume Alpha,” published this month as a National Bureau of Economic Research working paper. Here’s an excerpt from the paper (emphasis mine):
...we find that trading volume alpha is substantial. The marginal improvement on a portfolio from trading volume alpha is as large as finding return alpha. For example, for a $1 billion fund, the after-cost improvement in portfolio performance due solely to trading volume prediction beyond using lagged volume measures, can be as much as double in terms of expected returns or Sharpe ratio after trading costs.
It isn’t much of a secret that trading costs matter and that you can use data and math to improve them. But many otherwise sophisticated firms are still pretty old-school in the ways that they think about volume and the optimal time to step into the market, Tobias Moskowitz, one of the co-authors, told me Monday. “A lot of people are using machine learning to predict things like returns and prices, but actually there are lots of other economic variables,” said Moskowitz, a chaired professor at Yale School of Management and a principal at AQR Capital Management.
In their paper, researchers Ruslan Goyenko, Bryan T. Kelly, Moskowitz, Yinan Su and Chao Zhang compiled data on daily trading volume for thousands of stocks across 1,258 days from 2018 to 2022. On one hand, the researchers find that simple moving averages do quite well at predicting what volume will be in the future. But there are still plenty of curve balls that come along in the market, and mistakes can be costly on low-volume days. Their machine-learning exercise brought in variables including basic measures of company fundamentals; holidays; “triple-witching days” with high futures and option contract expirations; and earnings release schedules, among others. The authors concluded that “volume is highly predictable” (unlike returns) and that “adding more predictors improves accuracy.”
