Most quantitative research in finance obsesses over one question: can we predict returns? A recent working paper by Goyenko, Kelly, Moskowitz, Su, and Zhang (2025) argues that the field has been looking in the wrong place, or at least, not looking broadly enough. Their paper, "Trading Volume Alpha," shows that predicting trading volume and translating those forecasts into expected trading costs can generate performance improvements as large as those from return prediction itself.

The core idea

The insight is straightforward but underexplored. For any portfolio manager, the cost of executing a trade depends heavily on participation rate, the ratio of your trade size to total market volume in that stock. When volume is low, even a modest trade creates significant price impact. When volume is high, the same trade costs almost nothing to execute.

If you can predict which stocks will have high or low volume tomorrow, you can route your trades accordingly. Buy or sell more aggressively when volume is predicted to be high, and pull back when volume is expected to dry up. The authors formalize this intuition into a portfolio optimization framework that trades off tracking error (the cost of not trading) against price impact (the cost of trading).

An asymmetric cost structure

One of the more interesting findings is that the economic cost of volume prediction errors is asymmetric. Overestimating volume is far more dangerous than underestimating it. If you predict high volume and trade aggressively but actual volume turns out to be low, you get crushed by price impact. The reverse, predicting low volume and trading cautiously when volume is actually high, only costs you some missed opportunity.

This asymmetry has a practical consequence. The optimal strategy is naturally conservative. There are "inaction regions" where predicted volume is low enough that it is better not to trade at all, regardless of how attractive the target position looks. This provides a novel source of limits to arbitrage that goes beyond the usual explanations in the literature.

Machine learning meets market microstructure

The authors use 175 predictors, including lagged volume and returns, firm characteristics like book-to-market and market cap, calendar events like triple witching days, and earnings release schedules. They test linear models, feedforward neural networks, and recurrent neural networks (RNNs) that can capture time-series dynamics.

Each layer of complexity adds predictive power. Technical signals beat simple moving averages. Adding firm fundamentals improves further. Event-based indicators like upcoming earnings releases provide additional lift. Neural networks outperform linear models, and recurrent architectures do best of all.

But the most significant improvement comes not from a better model architecture, but from a better loss function. Standard machine learning models for volume prediction optimize mean squared error (MSE). The authors instead fine-tune their neural networks on an economic loss function derived directly from the portfolio optimization problem. This economic loss function recognizes the asymmetric cost structure described above: it penalizes overestimation of volume more heavily than underestimation, and it ignores prediction errors in regions where they do not matter much (very high or very low volume). This "economic learning" approach, a form of transfer learning between objective functions, delivers the largest marginal improvement of any single methodological choice.

The numbers

For a id="post-body" class="prose prose-lg max-w-none pb-16" style="font-family: 'Inter', sans-serif; line-height: 1.8; color: #374151;" data-astro-cid-xnoknld4> billion fund running a daily quantitative strategy, replacing a simple five-day moving average volume forecast with the full RNN model using economic loss optimization increases annual returns from 6.47% to 8.95% and more than doubles the Sharpe ratio from 2.21 to 4.53, all from volume prediction alone. No change in return signals, no change in the target portfolio, just better execution timing.

For a id="post-body" class="prose prose-lg max-w-none pb-16" style="font-family: 'Inter', sans-serif; line-height: 1.8; color: #374151;" data-astro-cid-xnoknld4>0 billion fund, where trading costs dominate, the improvement in annualized returns goes from 3.88% to 4.68%, worth roughly $80 million per year from better volume forecasts.

Applied to the "factor zoo" of 153 known anomaly portfolios, the improved volume prediction generates an average gain of 0.44% per year in net-of-cost returns across nearly all factors. High-turnover strategies like momentum and short-term reversals benefit the most, as expected, since these strategies face the largest execution costs.

What this means for practitioners

The paper makes a case that the research community has been too narrowly focused on return predictability while ignoring the economic value of predicting other variables. Trading volume is publicly available, highly predictable, and its economic benefits are substantial. The framework the authors provide is general enough to extend to other non-return predictors like volatility, spreads, or order flow metrics.

For systematic investors running capacity-constrained strategies, the message is clear: improving your execution model can be as impactful as improving your alpha model. The marginal dollar invested in better volume forecasting may yield higher returns than the marginal dollar spent on finding new return predictors, especially in an environment where return predictability continues to decay as more capital chases the same signals.

The full paper is available on SSRN (ID: 4802345).