This article discusses recent techniques and results in the area of forecasting intraday trading volume and intraday trading volume percentages. Why predict volume? A major reason is to improve the performance of trading algorithms, many of which depend upon the volume that will trade while the order is active. Traditionally, algorithms used historical averages when they needed to predict volume over the lifetime on an order. Improving upon this base case boosts the performance of the algorithm. We have shown that to be the case and present our results in this paper.
Volume predictions are also useful in situations where an algorithm may not be involved. One example is where a trader receives a large order in an unknown (to the trader) symbol ten minutes before the close with instructions to participate at all points to the close. In this situation, knowing average daily trading volume is of no value, while an accurate volume prediction for those final ten minutes of the trading day would be very helpful. Our example trader would like to know the number of shares that will trade during that ten-minute interval. We refer to this as a forecast of raw volume. Knowing upcoming raw volume is also of interest for a large group of algorithms, such as market participation models and portfolio trade scheduling tools, the latter of which generally use implicit volume forecasts within their cost models.
Improved volume forecasts aid alpha capture. Consider an alpha engine that continuously computes expected alpha trajectories for various stocks. Traders face an allocation problem in that they must maximize alpha capture without creating price impact. Applying sophisticated optimization tools to this problem while incorporating accurate volume forecasts simultaneously increases trading strategy capacity (and alpha), controls trading risk, and manages slippage.
Certain other algorithms, volume-weighted average price (VWAP) in particular, benefit from an accurate forecast of how much volume will trade in a given time interval as a percentage of the full day’s trading volume. We call these volume percentage forecasts, and they represent a different type of prediction problem than forecasting raw volume. This is easily seen in the context of the constraint imposed on forecasting volume percentages. Forecasts made early in the day constrain subsequent forecasts, as the day’s forecasts must total to 100% to be meaningful. Hence, a volume percentage forecast model couldn’t easily adapt to intraday news or other events.
Our focus is on making predictions intraday, forecasting the raw volume and volume percentages for fixed intervals of time from the present moment until market close. To measure and understand the performance of our models, we have developed a standard set of performance metrics for volume prediction and volume analysis studies.
When predicting volume, convention is to reference historical averages as the base case. Relative to that base case, we show improvements in the prediction of raw volume of 29%. In the case of volume percentages, we improved over the base case by 7%. More importantly, we show that using our predictive volume percentages improves the performance of a VWAP algorithm relative to using historical averages by 9%. These results are based on a larger set of symbols and a longer period of time than prior published works have analyzed.
The next section introduces terms we use throughout the paper. We then define the metrics we used to measure our forecasts and argue that these – or equivalent measures – represent the best way to measure performance. From there, we discuss our models and results. A section on prior published work then follows and precedes our summary of results.
Written by FlexTrade’s Venkatish Satish, Vice President, Abhay Saxena, Vice President and Max Palmer, Head of Trading Analytics and Algorithms