Using Machine-Learned Volume Forecasts to Reduce Portfolio Trading Costs
Summary
The document explains a portfolio construction approach that forecasts stock trading volume and uses the forecasts as a proxy for expected price impact. With trade size held fixed, lower volume implies a higher participation rate and greater expected impact, while higher volume allows more trading at lower expected cost. The portfolio optimizer balances these costs against tracking error and the opportunity cost of not trading.
The cited study trains recurrent neural network models using lagged returns and volume, company characteristics, and indicators for market or firm events such as earnings reports. It evaluates forecasts and portfolio applications on a US equity sample from 2018 to 2022, and reports that volume forecasts can improve net implementation outcomes, especially for portfolios with higher turnover. Forecast errors are asymmetric: overestimating volume can lead to costly excess trading, particularly when liquidity is low. The approach assumes fixed trade size and specified return moments; actual costs and benefits may differ with investor behavior, alternatives to a costly trade, assets under management, and execution choices. Reported results are study estimates, not guaranteed investor returns.
Key ideas
- Trading volume helps estimate price impact because trade size relative to volume determines market participation.
- The study uses recurrent neural networks with market, company, and event features to forecast volume.
- Portfolio optimization balances expected trading costs against tracking error and the cost of not trading.
- Overestimating volume can be more costly than underestimating it, especially for illiquid stocks.
- The reported portfolio benefits depend on assumptions about trade size, costs, and investor scale.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.