Research

A Harvard-led NBER working paper found a neural network could predict 71% of what active mutual-fund managers would do each quarter, buy, sell or hold

The model, trained on rolling five-year windows of fund and stock data from 1990 to 2023, missed 29% of trades, and those unpredicted trades were the ones associated with market-beating performance.

· Source: Bloomberg


A Harvard Business School-led working paper, circulated through the National Bureau of Economic Research, trained a neural network on rolling five-year windows of US mutual-fund data spanning 1990 to 2023, feeding it fund size, investor flows, stock characteristics and broader economic conditions. The task was narrow and specific: predict whether an active fund manager would buy, sell or hold a given stock in a given quarter. The model got it right about 71% of the time.

The more interesting number is the other 29%. Trades the model failed to predict, the ones falling outside the routine, learnable pattern the neural network had picked up, were on average associated with better performance than the trades it correctly called. In other words, the bulk of what active managers do each quarter is mechanically reconstructible from public and semi-public data, and the value a manager adds shows up disproportionately in the minority of decisions that don't fit that pattern.

That finding lands alongside a month of coverage, already reflected in this site's recent items on crowded AI trades and shrinking trading edges, arguing that widely shared models and data converge participants onto the same signals faster than before. A study showing 71% of active-management decision-making is predictable from public information is a structural version of the same argument: if a pattern is learnable by an outside neural network trained on public data, it is also learnable by every other quant desk doing the same exercise, which is not a description of a durable edge.

What the paper does not claim is that mimicking the predictable 71% would itself be a profitable strategy, only that it is a predictable one; the 29% is precisely the part resisting automation, and it's also the part a fund manager would point to as justifying their fee. Read that way, this is a measurement of how routine active management has become, not a blueprint for replacing it, and the honest caveat is that a working paper trained on data through 2023 is describing a pattern in past manager behavior, not verifying that following it would have made money.


Read the original: Bloomberg - A Harvard-led NBER working paper found a neural network could predict 71% of what active mutual-fund managers would do each quarter, buy, sell or hold. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.