Trading Costs and Machine-Learning Return Forecasts

Abstract

In a large panel of U.S. stocks, machine-learning return forecasts are most accurate where little capital can be invested. Among stocks that can absorb capital, predictability is weak and hard to estimate precisely. We ask whether trading costs should enter the loss on which the forecast is trained or the way the portfolio is held and traded. Across an elastic net, a boosted tree, and a neural network, reweighting the loss toward liquid stocks does not significantly improve the ordering of those stocks or the net Sharpe ratio, whereas restricting the training sample to liquid stocks lowers accuracy on the stocks it retains. Weighting holdings by liquidity and trading only when the forecast covers the half bid-ask spread raises the net Sharpe ratio of the standard forecast with $500 million of capital invested from -0.38 to 0.42, while the gross Sharpe ratio falls from 1.86 to 1.21; retraining under the reweighted loss adds a statistically negligible 0.12 to the net Sharpe ratio.

Download

SSRN Preprint

Citation

Bianchi, Daniele, and Teng Jiao. Trading Costs and Machine-Learning Return Forecasts. Working paper.

@article{bianchi2026trading,
  title={Trading Costs and Machine-Learning Return Forecasts},
  author={Bianchi, Daniele and Jiao, Teng},
  journal={Available at SSRN 7617338},
  year={2026}
}