NBER WORKING PAPER SERIES

TRADING VOLUME ALPHA

Ruslan Goyenko
Bryan T. Kelly
Tobias J. Moskowitz
Yinan Su
Chao Zhang

Working Paper 33037

[Link]

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
October 2024

We are grateful to the Columbia & RFS AI in Finance Conference participants and discussant
Dmitriy Muravyev; seminar participants at Cornell, Syracuse, CityU Hong Kong, and George
Mason; as well as Martin Lettau, Lu Lu, Andrew Patton, and Annette Vissing-Jorgensen for
valuable comments and suggestions. We thank Zhongji Wei and Andy Yang for their excellent
research assistance. AQR Capital Management is a global investment management firm, which
may or may not apply similar investment techniques or methods of analysis as described herein.
Moskowitz is a member of the NBER, has an academic consulting relationship with AQR Capital,
and sits on the board of Commonfund. The views expressed here are those of the authors and not
necessarily those of AQR or the National Bureau of Economic Research. Send correspondence to
Bryan Kelly, [Link]@[Link].

NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies
official NBER publications.

© 2024 by Ruslan Goyenko, Bryan T. Kelly, Tobias J. Moskowitz, Yinan Su, and Chao Zhang. All
rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit
permission provided that full credit, including © notice, is given to the source.
Trading Volume Alpha
Ruslan Goyenko, Bryan T. Kelly, Tobias J. Moskowitz, Yinan Su, and Chao Zhang
NBER Working Paper No. 33037
October 2024
JEL No. C45, C53, C55, G00, G11, G12, G17

ABSTRACT

Portfolio optimization focuses on risk and return prediction, yet implementation costs critically
matter. Predicting trading costs is challenging because costs depend on trade size and trader
identity, thus impeding a generic solution. We focus on a component of trading costs that applies
universally – trading volume. Individual stock trading volume is highly predictable, especially with
machine learning. We model the economic benefits of predicting volume through a portfolio
framework that trades off tracking error versus net-of-cost performance – translating volume
prediction into net-of-cost alpha. The economic benefits of predicting individual stock volume are
as large as those from stock return predictability.

Ruslan Goyenko Yinan Su

McGill University Johns Hopkins University
Faculty of Management Carey Business School
1001 Sherbrooke St West 100 International Drive
Montreal, Quebec Canada Baltimore Maryland 21202
[Link]@[Link] ys@[Link]

Bryan T. Kelly Chao Zhang

Yale School of Management Hong Kong University of Science
165 Whitney Ave. and Technology (HKUST)
New Haven, CT 06511 chao.zhang94@[Link]
and NBER
[Link]@[Link]

Tobias J. Moskowitz
Yale School of Management
165 Whitney Avenue
P.O. Box 208200
New Haven, CT 06520-8200
and NBER
[Link]@[Link]
1 Introduction

Research on portfolio optimization chiefly focuses on mean return prediction, and to a lesser extent,

variance-covariance prediction.1 However, real-world implementation costs also play a critical role

in the efficacy of portfolios. While the benefits and pitfalls of mean return and volatility forecasts are

well-studied in the literature, trading costs have received relatively little attention, and forecasting

trading costs has received no attention.2

Predicting trading costs, particularly at the individual stock level, is challenging since the

biggest component of these costs for a large investor is price impact, which depends on the size of

the trade and the amount traded by other traders in that security, as well as the identity of the

trader (Frazzini, Israel, and Moskowitz, 2018).3 Since each trader may face their own cost function,

finding a generic solution to this piece of the portfolio problem is challenging. Moreover, the size

of the trade is endogenously a function of expected trading costs, finding an estimate of expected

costs is critical to the portfolio decision.

To circumvent these issues, we take a unique approach to predicting stock-specific expected

trading costs by focusing on the component of costs that is neither trader-specific nor endogenous

– the level of total trading volume in the stock. Trading volume is driven by other traders in the

market for the same security, which is generic to all traders. As Frazzini, Israel, and Moskowitz

(2018) show, trade size divided by daily trading volume – termed the market participation rate in a

stock – is the key driver of price impact costs. Following Kyle (1985), price impact is an increasing

function of participation rate. Holding trade size constant, the less trading volume in the stock the

greater the trader’s price impact will be.4 Our empirical strategy is to forecast trading volume for
1
A very long literature in asset pricing focuses on return prediction, with summaries on the state of this literature,
including some of its criticisms, found in Harvey, Liu, and Zhu (2016); McLean and Pontiff (2016); Jensen, Kelly, and
Pedersen (2022). A good summary of the literature on volatility prediction can be found in Engle (2004).
2
Several papers focus on trading costs from a theoretical perspective: Kyle (1985); Gârleanu and Pedersen (2016).
Even less work has provided empirical estimates of trading costs for use in a portfolio optimization context: Frazzini,
Israel, and Moskowitz (2012, 2018).
3
While many different trading cost models exist, they universally contain these three elements: trade size, market
size, and trader specifics (identify, information, motive, patience, etc.).
4
Theoretical foundations of trading costs in the classic market microstructure literature include asymmetric infor-
mation (Kyle, 1985; Glosten and Milgrom, 1985) and inventory costs (Stoll, 1978; Ho and Stoll, 1983; Grossman and
Miller, 1988). In either case, volume is the major determinant of liquidity (Benston and Hagerman, 1974; Glosten
and Harris, 1988; Brennan and Subrahmanyam, 1995) and is used to proxy for trading costs or liquidity (Campbell,
Grossman, and Wang, 1993; Datar, Naik, and Radcliffe, 1998; Amihud, 2002).

2
each security as a proxy for expected trading costs, and then use this forecast to optimize portfolios

net of those costs and quantify its benefits.

Importantly, our aim is to provide a general forecast of costs for an individual stock that

abstracts from the need to specify trade size and applies universally across all traders. Our goal

is not to provide the “best” or most reliable trading cost model or forecast. Rather, our simple

objective is to provide a forecast of trading costs that any trader could use. This aim necessitates

a simple, rather than sophisticated, trading cost model. The goal is to focus on one component of

a trading cost model that is generic and see how valuable that can be.

To quantify the economic magnitude of volume prediction, we incorporate volume prediction

into a portfolio theory problem. We model a portfolio framework that seeks to maximize the

net-of-cost performance of the portfolio using a mean-variance utility function, where the cost of

transacting scales linearly with participation rate (motivated by theory and empirical work from

the literature). The optimization trades off the cost of trading versus the (opportunity) cost of

not trading – minimizing trading costs versus minimizing tracking error to the before-cost optimal

portfolio – where trading costs and tracking error are endogenously negatively related. In the

model, we take the first and second moments of security returns as given and focus solely on the

tradeoff between trading costs and tracking error to the pre-cost optimal portfolio.

Using this framework, we can couch the volume prediction problem into a portfolio problem

and quantify the economic value of volume prediction in terms of its benefit to after-cost returns

or Sharpe ratio.5 In essence, we translate trading volume predictability into portfolio alpha, which

we term “trading volume alpha.” This translation opens up after-cost portfolio modeling, which

has been restricted due to limited access to trading cost data, to the widely available volume data.

This paper demonstrates the economic value of volume prediction with one set of predictors and

standard neural network implementation. Yet the framework laid out in the paper illustrates the

importance of and enables more work on constructing volume prediction signals and models.6

We impose a functional form for trading costs due to price impact motivated by theory (Kyle,
5
Related, Balduzzi and Lynch (1999) and Çetin, Jarrow, and Protter (2004) model transaction costs in portfolio
settings and study the economic value in various transaction cost models.
6
This research answers the call to integrate market microstructure frictions into asset pricing studies using machine
learning tools (Goldstein, Spatt, and Ye, 2021).

3
1985) and empirical evidence (Frazzini, Israel, and Moskowitz, 2018), where price impact (Kyle’s

Lambda) is an increasing linear function of the trader’s participation rate. All else equal, a higher

predicted volume allows the trader to trade more aggressively (larger size) because price impact per

dollar traded will be lower. Conversely, a lower predicted volume causes the trader to scale back

the trade (even perhaps to zero and not trade) because the price impact per dollar will be higher.

By modeling trading costs and benefits as a tracking error problem, we abstract from return or

variance prediction and focus exclusively on volume prediction and its economic impact through

expected trading costs.

An interesting feature emerges from the model that generates an asymmetry in the economic

value of volume prediction. Price impact costs are linear in participation rate, but non-linear

in trading volume. Very low trading volume implies exponentially high impact costs, whereas

very high volume implies negligible costs. As volume tends to zero, price impact costs approach

infinity, whereas when volume becomes large, impact costs are bounded by zero. Hence, predicted

changes in volume have much more economic impact when volume is low versus high, thus creating

asymmetric costs of volume forecast errors. Conversely, tracking error, or the opportunity cost of

not trading, is independent of volume. The combination of these two effects implies the optimization

will penalize overestimating volume more than underestimating volume. Trading too much when

you overestimate volume is more costly than trading too little when you underestimate it. At lower

volume, the cost of trading with respect to volume is very steep, and at high volume the cost-volume

relationship is flat. Intuitively, an illiquid stock’s price impact is very sensitive to small changes

in volume and in a highly non-linear way – participation rates move by orders of magnitude for

small changes in volume – resulting in trading costs increasing at a much faster rate when trading

too aggressively. Conversely, a very liquid security’s price is fairly inelastic to changes in volume,

because having more or less liquidity at that point has little impact on costs. As a result, the

optimization seeks to be conservative, rather than aggressive.7

Since participation rate drives trading costs, it is not only trading volume that matters, but also
7
Our framework, and its implication that portfolio optimization will seek to trade conservatively rather than
aggressively because of the asymmetric cost of volume forecast errors, implies that arbitrage activity may be limited
as a consequence. This implication provides a novel and additional source of limits to arbitrage activity in the spirit
of Shleifer and Vishny (1997).

4
the size of the trade. Trade size is determined endogenously and is a function of expected trading

volume and aversion to tracking error. In the model, the size (assets under management, AUM) and

volatility of the fund (and risk aversion of the investor) also affect the costs and benefits of trading.

Because price impact is an increasing function of participation rate, trading costs increase with

AUM endogenously, and the relative penalty for tracking error decreases with AUM. The optimal

tradeoff between trading costs and tracking error will therefore vary with the size of the portfolio,

and so will the economic impact of volume prediction. For small AUM, tracking error considerations

likely dominate trading cost considerations, hence the economic benefit to predicting volume may

be relatively less valuable. For large AUM, trading cost considerations likely dominate.8

Applying the model framework to data, we run a series of trading experiments for optimally

designed portfolios that take into account trading costs using stock-level volume prediction as the

sole input. Since liquidity is an unknown quantity to the portfolio manager, she uses volume

prediction as an input to alter the expected cost and benefit of trading, endogenously responding

to her forecast of volume by altering her portfolio. We assess the out-of-sample performance of this

portfolio, net of trading costs, to evaluate the economic impact of predicting volume. More accurate

volume predictions provide more efficient implementation and hence more efficient portfolios net of

costs.

We experiment with various sets of target positions to mimic realistic trading tasks. We start by

simulating an extremely profitable (before-cost) daily quantitative trading strategy, which represents

an unachievable target. This portfolio, which requires high turnover and aggressive trading, is much

less profitable after accounting for trading costs. We then consider expected trading costs in the

optimization to maximize the after-cost performance of the fund (at various fund sizes). We also

target a host of factor portfolios from Jensen, Kelly, and Pedersen (2022) that are based on ex-ante

return-predicting characteristics to examine the effectiveness of trading volume alpha across the

spectrum of factors. We examine the net of trading cost performance of an investment strategy

that seeks to target each of these factor portfolios while taking expected trading costs into account,
8
From this analysis, it is also possible to assess the optimal dollar size of the portfolio, including the break-even
fund size where trading costs exactly offset portfolio returns or the fund size that maximizes after-cost dollars rather
than returns (Frazzini, Israel, and Moskowitz, 2018).

5
using predicted market trading volume as the only input into forecasting expected trading costs.

Expected trading costs dictate how frequently to trade, which stocks to trade, and how much to

trade, with each of these choices simultaneously affecting tracking error to the portfolio.

We find that volume prediction has a measurable and significant effect on after-cost portfolio

performance. To predict volume, we use technical signals, such as lagged returns and lagged trading

volume, as well as firm characteristics that the literature finds capture return anomalies, but not

necessarily trading volume. We then add indicators for various market-wide or firm-level events

associated with volume fluctuation, including upcoming and past earnings releases. We analyze

both linear and non-linear prediction methods using various neural networks designed to maximize

out-of-sample predictability. Finally, we alter the objective/loss function of the neural network to

take into account the portfolio problem’s economic objective when predicting volume. Our aim,

again, is not to find the best prediction model for stock trading volume, but rather to couch the

prediction problem into economic outcomes and measure its costs and benefits accordingly. To

that end, we assess a number of different models and variables for predicting volume in order to

highlight how different prediction methods lead to different economic consequences.9

We find that volume prediction improves significantly over moving averages of lagged trading

volume when using technical signals. Adding firm characteristics (such as BE/ME) further im-

proves volume predictability, even though these variables are primarily used for return forecasting

and are not necessarily related to trading volume. Information on events such as earnings releases

further enhances volume predictability. Non-linear functions from neural network searches provide

better predictability over simple linear models, controlling for the same set of variables/information.

Recurrent neural networks, which learn dynamic predictive relationships, yield additional improve-

ments.

Finally, we find that imposing an economic loss function consistent with the portfolio prob-

lem greatly improves the value of volume predictability from a portfolio performance perspective.

Specifically, we fine-tune the neural networks on the economic objective (derived from the portfolio

problem) rather than on a statistical objective, such as mean squared errors used to pre-train the
9
For example, we leave out many potential variables that relate to trading volume, such as other microstructure
variables, cross-security lead-lag effects, etc.

6
volume prediction model. Fine-tuning yields significant gains in out-of-sample portfolio perfor-

mance compared to the statistically pre-trained model. This result obtains because the portfolio

problem recognizes that trading costs are not a linear function of trading volume. The neural

network places greater weight on observations that impact trading costs more and worries less

about predicting volume where trading costs are less affected, such as recognizing the asymmetric

costs of over- versus underestimating volume. While an MSE objective criterion may maximize

the out-of-sample R2 of volume prediction, an economic loss function directly tied to the portfolio

problem provides more valuable volume predictability that results in better out-of-sample after-cost

portfolio performance.

Since portfolio size affects the tradeoff between the cost and benefit of trading, the resulting

portfolio solution also changes with different levels of AUM. In addition, while volume prediction

has benefits across all factors, some factors have a greater trading volume alpha than others due to

the varying tradeoff between the cost of trading and the opportunity cost of not trading. Intuitively,

we find that factors with a higher turnover (e.g., momentum, short-term reversals) benefit more

from portfolio optimization that accounts for expected trading costs based on volume forecasts.

In general, we find that trading volume alpha is substantial. The marginal improvement on a

portfolio from trading volume alpha is as large as finding return alpha. For example, for a $1 billion

fund, the after-cost improvement in portfolio performance due solely to trading volume prediction

beyond using lagged volume measures, can be as much as double in terms of expected returns or

Sharpe ratio after trading costs. Among popular asset pricing factors, the improvement in after-cost

returns ranges from 20 bps to 100 bps above using a moving average of lagged volume to predict

future volume. Refining the prediction methods and deepening the prediction models could add

substantially to these improvements, generating even larger trading volume alpha.

The rest of the paper is organized as follows. Section 2 covers some preliminaries to the analysis:

a motivation for predicting trading volume and a description of our data. Section 3 examines volume

prediction from a purely statistical perspective to maximize out-of-sample predictability. Section

4 presents a theoretical optimal portfolio framework for quantifying the economic value of volume

prediction in terms of portfolio performance, which we call “trading volume alpha.” Section 5

7
discusses the empirical results of predicting volume through the lens of our theoretical framework

using a variety of machine learning methods. Section 6 applies these insights and methods to

trading experiments that characterize the net of cost performance improvement of simulated real-

world portfolios. Section 7 concludes.

2 Preliminaries: Motivation and Data

We first motivate why predicting volume is interesting and useful and then describe the data.

2.1 Motivation

Trading costs are critical to realizing investment performance, yet insufficient empirical attention

has been paid to them regarding optimal portfolio construction. This lack of research is largely

due to the challenges imposed by modeling trading costs and finding a generic portfolio solution.

As a consequence, and despite its importance, portfolio theory mainly focuses on modeling and

forecasting the first and second moment of returns gross of costs.

One of the main challenges in modeling and forecasting trading costs is that the largest com-

ponent of these costs for a large investor is price impact, which depends on the size of the trade,

the amount traded by other traders (in the same and in opposite directions), and the identity of

the trader – different traders may face different price impacts. These features frustrate a generic

portfolio solution that incorporates trading costs. In particular, since optimal portfolio weights

should reflect the expected net of cost return, a solution requires an expected cost function, which

varies by investor, and hence so will the portfolio solution.

Following Kyle (1985) and subsequent empirical work (Frazzini, Israel, and Moskowitz, 2018),

price impact depends on the trader’s participation rate, defined as the dollar amount traded by

investor n in security i relative to total dollar volume in stock i (the amount traded in the market

by everybody in security i) at the same time t,

$T radedn,i,t
P articipationRaten,i,t = .
$V olumei,t

8
 Price impact is an increasing function of participation rate (modeled linearly in Kyle, 1985 and

the positive relation empirically verified in Frazzini, Israel, and Moskowitz, 2018), whose elasticity

varies by investor n. The numerator is also endogenous to expected price impact, which itself is

a function of the participation rate. The circular nature of trade size, and its variation across

investors, makes modeling trading costs particularly challenging. However, the denominator of the

participation rate is independent of n and exogenous to the trader’s desired trade size (assuming a

single investor’s trade is too small to materially affect total dollar volume). Thus, trading volume is

generic and exogenous to each trader and provides a variable universal to all investors for modeling

costs.

From a prediction standpoint, total dollar volume is also easier to forecast than a specific

investor’s trades. Predicting market-level trading activity (total buys and sells) is an easier task

and high frequency data on total trading volume is readily available, while data on individual

traders is not. Moreover, as we will find, machine learning techniques can significantly improve our

ability to forecast volume, due in part to the non-linear nature of volume and its relation to trading

costs. With these insights, we model expected trading costs solely using forecasted total dollar

volume for a stock. Although this exercise is only a partial solution to the trading cost problem, it

is a general one, and it allows us to showcase the economic value of predicting trading volume in

an optimal portfolio framework.10

2.2 Data

We compile a data panel of daily stock-level dollar trading volume (Vei,t ) and 175 predictors (Xi,t ).

The unit of observation is stock-day (i, t). We adopt the convention that Xi,t is observed by day

t − 1, whereas the associated prediction target Vei,t is observed until the end of day t. We use a tilde

to denote a random variable conditional on the information before day t.

The sample period is 2018 to 2022, or 1,258 days. The cross-section covers around 4,700 stocks,

with an average of 3,500 stocks per day, or 4,400,000 observations in total. We split the data into
10
Korajczyk and Sadka (2004), Novy-Marx and Velikov (2016), and DeMiguel et al. (2020) evaluate the after-cost
profitability of factor portfolios, rather than actively seeking cost mitigation in portfolio optimization. Jensen et al.
(2022) take price impact transaction costs as given in portfolio optimization. We tackle the forecasting problem given
uncertain transaction costs and apply it within a portfolio optimization framework.

9
a 3-year training sample and a 2-year testing sample. All models are trained once in the training

sample and evaluated out of the sample. We avoid re-sampling methods such as cross-validation

and rolling-window re-estimations.

Our analysis focuses on predicting out-of-sample trading volume. Reasonable in-sample fit,

often evaluated in the literature (Chordia, Huh, and Subrahmanyam, 2007; Chordia, Roll, and

Subrahmanyam, 2011), does not often lead to good out-of-sample (OOS) performance, which is of

primary interest to evaluate the robustness and economic impact of volume predictability.

While it is common in the stock return predictability literature to use large sets of variables

(e.g., “factor zoo”) to identify the best predictors, it is less common to use large data sets to predict

market microstructure variables such as trading volume. We show that using big data improves

the precision of volume forecasts significantly.

The main variable we aim to predict is dollar trading volume, which we measure as the natural

logarithm of end-of-day transacted total dollar trading volume for each stock. This variable is

highly persistent. We focus on predicting innovations in trading volume as well, which we show

has significant impact on trading costs. When volume is suddenly much lower than expected,

an investor will “overpay” in transaction costs. If volume is higher than expected, then there is

more liquidity, and an investor incurs opportunity costs from not trading aggressively enough. We

examine the predictive content of several sets of predictors, including technical, fundamental, and

event-based variables, which we describe below.

2.3 Prediction objects: daily stock trading volume

Daily dollar trading volume Vei,t ranges widely (from thousands to billions of dollars) across stocks

and is highly skewed. We take the log of dollar volume, ve = log Ve , whose distribution, shown in

Figure 1, is relatively well-behaved, being close to normal and symmetrical.

Log dollar volume is highly persistent in the time series and cross-section of stocks, and can

be easily predicted by lagged moving averages of various frequencies. The five-day moving average

predicts log dollar volume with an R2 of 93.68%, higher than the one-day lag (92.53%), moving

average of 22 days (92.60%), or 252 days (86.12%).

10
 Figure 1: Distributions of daily stock-level log dollar volume (e
v ) and its shock (e
η ), panel pooled

Histograms of vei,t and ηei,t in the full sample of around 4,400,000 stock-day observations. The
second horizontal axis in the left panel is dollar volume Vei,t in the log scale. Log dollar
volume shock ηe is daily log dollar volume minus the moving average in the past five days:
ηei,t := vei,t − 15 (e
vi,t−1 + · · · + vei,t−5 ).

We focus on predicting the log dollar volume shock defined as daily log dollar volume minus
1
the moving average in the past five days, ηei,t := vei,t − 5 vi,t−1 + · · · + vei,t−5 ) := vei,t − [ma5 ]i,t .
(e

This exercise is comparable to predicting asset returns (change in log price) instead of price levels.

Figure 1 (right panel) shows the pooled distribution of ηe, which is relatively symmetric, centered

around zero, and has long tails.

2.4 Predictors

We use a total of 175 predictors from various sources, including technical signals, firm fundamentals,

and market and corporate events. We show that the virtue of complexity approach in return

prediction (Kelly, Malamud, and Zhou, 2024) is also useful for volume prediction. We find that

each subset of variables provides incremental improvement to predicting volume, while using all

variables has the greatest OOS predictability. We list the sets of predictors that are cumulatively

added to the prediction model.

1. Technical signals (“tech”): lagged moving averages of returns and log dollar volume over the

past 1, 5, 22, and 252 days. (8 predictors.)

2. A small set of commonly used fundamental firm characteristics (“fund-1”): market equity,

standardized earnings surprise, book leverage, book-to-market equity, Dimson beta, and firm

11
 age. (6 predictors.)11

3. The remaining firm characteristics from the JKP dataset (“fund-2”), which are merged and

transformed in the same way as fund-1 variables. (147 predictors.)

4. Calendar dates with large effects on trading volume (“calendar”). We hard code four binary

features based on the dates of the following four types of events.

- Early closing days for the exchanges (July 3rd, Black Friday, Christmas Eve, and New

Year’s Eve).

- Triple witching days (four times a year when the index futures contract expires at the

same time as the index option contract and single stocks options).

- Double witching days (eight times a year when two of the three events above, the single

stock options and index options expiration, coincide).

- Russell index re-balancing (once a year, the fourth Friday in June).12

Early closing days have substantially less trading volume, while the other three are associated

with positive spikes in trading volume.

5. Earnings release schedule (“earnings”): We construct 10 categorical dummy variables (one-

hot encoding) indicating whether the firm has an upcoming earnings release or just had one

in the past few days. We first construct the number of days until the next known scheduled

earnings release event. For example, a value of zero implies the current day is previously

known to have a scheduled release. A negative value means there are no known scheduled

events in the future and indicates how many days since the last event. We convert this

number into 10 dummy variables of categorical bins: ≤ −4, −3, −2, −1, 0, 1, 2, 3, 4, ≥ 5. The

data source is the Capital IQ Key Developments dataset. (10 predictors.)

This list of variables for predicting volume is not exhaustive. Other variables that could add
11
These predictors are from the JKP dataset (Jensen, Kelly, and Pedersen, 2022). We forward fill the monthly firm
characteristics in time when merging to the daily panel. Hence, the characteristics are still always ex-ante available.
On each day, we rank standardize in the cross-section each characteristic to a uniform distribution from -1 to +1.
12
Triple witching happens on the third Friday in March, June, September, and December. Double witching is on
the third Friday in the other eight months. Russell index re-balancing is on the fourth Friday of June when the Russell
1000, Russell 2000, Russell 3000, and other Russell indexes are reconstituted, and has “often been one of the highest-
volume trading days of the year” for the exchange, due to indexes tracking funds adjusting their holdings to reflect
the updates. Rerferenes: [Link] and
[Link]

12
predictive power are microstructure variables, intraday observations, and lead-lag relations across

stocks in terms of trading (e.g., large to small stocks, within industry, etc.). As stated previously,

we do not attempt to provide the best volume prediction model. Rather, we translate the volume

prediction problem into an economic problem whose objective is after-cost portolio performance.

Using our framework, future work can add further predictors for volume that may provide even

larger economic benefits than we show here. However, our framework provides a way to assess

those predictive contributions in economic terms.

3 Volume prediction from a statistical perspective

We start with a statistical prediction of daily trading volume using various subsets of predictors

and a variety of methods, including machine learning techniques.

3.1 Prediction methods

We run predictive regressions of ηe (changes in daily dollar volume) on a set of predictors, X,

in the training sample panel to estimate the models.13 We compare linear models (ols), with

neural networks (nn) that allow for non-linear transformations and complex interactions, as well

as recurrent neural networks (rnn) that, in addition, allow for state variables to incorporate time

series dynamics. The simplest baseline is predicting vbi,t = [ma5 ]i,t , or in other words, ηbi,t = 0. (The

“hat” denotes predicted values.) Linear regression is also a simple benchmark comparison.

The neural network implementation is kept simple, standard, and fixed throughout the paper in

order to facilitate transparency. The network architecture has three fully-connected hidden layers

of 32, 16, and 8 ReLU nodes, respectively, and one linear output node. The size of the input layer

is the number of predictors supplied.

Recurrent neural network architecture is particularly appealing for this application as it is

designed to capture time-series dynamics. An rnn is analogous to state space models like GARCH,

in which the forecast ηbi,t is not only a (non-linear) function of the concurrent predictors Xi,t , but
13
As explained above, predicting ηe or ve are essentially the same: predicting ηe as nn(X) is just predicting ve as
nn(X) + ma5 , where ma5 is one of the predictors. From the machine learning perspective, this is implementing a
simple residual connection (ResNet, He et al. 2016) as illustrated in Figure 4.

13
also of a set of state variables that is the output of the network applied to the previous data

point {i, t − 1}. That is, (b

ηi,t , statei,t ) = rnn(Xi,t , statei,t−1 ), where rnn represents the neural

network function and state are the state variables. The recurrent neural network processes data

sequentially, and recursively passes the state variables to the next time period. Essentially, rnn

extracts predictive information from concurrent and lagged predictors Xi,t , Xi,t−1 , Xi,t−2 , . . . , in

contrast to a nn that uses only the concurrent predictors but nothing from the past. Although Xi,t

contains moving averages of vei,t−1 , vei,t−2 , vei,t−3 . . . , for example, the way such lagged information

enters the model without an rnn architecture is highly restrictive. With rnn, the model can “learn”

flexible dynamics, where time-series dependencies are parameterized by trainable network weights.

We implement the rnn with the popular and standard lstm (long short-term memory) architecture.

The number of layers and neurons are kept the same as nn, but the total number of parameters

increases by four times (due to the flow of lagged information).14

Appendix A.1 contains other details on implementing the machine learning methods, includ-

ing the optimizer, training scheme, and infrastructure. We do not tune or optimize the hyper-

parameters, the architecture, or the training scheme to improve the results.

3.2 Prediction results

Table 1 reports the OOS prediction accuracy of each method. We cumulatively add sets of predictors

in the columns from left to right to highlight the prediction improvement from using larger data

sets. The rows correspond to different prediction methods. We express the accuracy in two R2 ’s.

Panel A reports the explained percentage of the total sum of squared errors of log dollar volume

relative to ηe, the residual after controlling for the five-day moving average. Panel B converts that

to the percentage in the total sum of squared of log of dollar volume (e

v ). Lastly, Panel C reports

the number of parameters estimated in each model with each set of predictors.

Volume is highly predictable. The most sophisticated model using all predictors can predict
14
Specifically, the bottom hidden layer in the aforementioned 3-layer network is upgraded to an lstm layer with
32 hidden states and cell states, with the rest of the two layers unchanged. Lstm is a standard and popular type of
rnn with four specific internal mechanisms, or gates, that control the flow of information from both the short- and
long-term past (Hochreiter and Schmidhuber, 1997). See Kelly and Xiu (2023) for a general reference and Appendix
A.1 for our specifications.

14
 Table 1: Prediction accuracy

cumulatively adding predictor sets tech fund-1 fund-2 calendar earnings

total number of predictors 8 14 161 165 175
A: R2 relative to ηe (e
v − ma5 )
ma5 0
ols 12.09 12.26 12.27 14.85 15.99
nn 14.31 14.90 14.42 17.13 18.45
rnn 15.80 16.25 15.47 18.12 19.86
B: R2 relative to ve (log dollar volume)
ma5 93.68
ols 94.44 94.45 94.45 94.62 94.69
nn 94.58 94.62 94.59 94.76 94.85
rnn 94.68 94.69 94.64 94.86 94.93
C: number of parameters
ma5 0
ols 9 15 162 166 176
nn 961 1,153 5,857 5,985 6,305
rnn 6,049 6,817 25,633 26,145 27,425

Each row represents a prediction model, and each column cumulatively adds to the set of predictors.
Panels A and B respectively express the OOS prediction accuracy in two different R2 ’s. The R2
relative to ηe is 1 − MSE/avg(e v − ma5 )2 ; R2 relative
P to ve is 1 − MSE/avg(e v ))2 , where
v − avg(e
1
MSE := avg(e v − vb)2 = avg(e
η − ηb)2 and avg := |OOS| i,t∈OOS is the OOS average. Each reported
R2 value is the average across five independent runs initialized with different random seeds to ensure
the results’ robustness and reproducibility. Panel C reports the number of parameters, for which
rnn is about four times of nn due to the four gates in lstm, see exact formulas in Appendix A.1.

nearly 20% of future variation in daily trading volume changes. In comparison, daily stock returns

are hardly predictable with a positive OOS R2 , even with state of the art models and predictors,

and monthly returns are only slightly predictable (Gu, Kelly, and Xiu, 2020).

Even simple ols models can deliver double-digit R2 , especially when using the largest set of

predictors, but the nn and rnn add an additional 3-4% of OOS R2 predictability.15

Adding more predictors improves accuracy in general. All 175 predictors can increase the R2

by more than three percentage points compared to just the eight technical signals. The exception

is that adding the large set of fundamental signals (fund-2) makes the methods (even ML methods)
15
We experimented with regularization on the linear model (lasso and ridge regressions), and did not find significant
improvements.

15
perform a little worse. This may be due to overfitting when the number of features increases and

where we do not use regularization techniques to try to correct for that. The predictors associated

with expected returns do not necessarily work for predicting volume. Market-wide calendar events

are quite effective in capturing volume changes, however. Scheduled earning announcements also

add a sizable gain in prediction accuracy.

The results show that machine learning is critical, and that complexity has its virtue in the

context of predicting volume. The prediction accuracy of the rnn is better than the nn, which

in turn is better than ols, uniformly across each configuration of included predictors. Panel C

shows the improved prediction accuracy is achieved through a significant increase in the number of

parameters, a measure of model complexity. Appendix A.1 shows the computational costs of the

complex models are higher but manageable.

Panel B reports an alternative R2 measured as the explained percentage of the total variation of

ve. Under this metric, the R2 is always high since the trailing moving average explains ve to a large

extent already, and the gain is at most around 1.2 percentage points on top of ma5 , which does

not look impressive. This begs the question: which one is the right metric to gauge the economic

value of this statistical task? Similarly, why take the log of volume and not predict dollar volume

Ve or some other transformation of it? In the next section, we will formulate an economic objective

as a function of volume by modeling a portfolio problem and show the economic value from the

seemingly small gain in R2 is indeed large and valuable.

Appendix C.1 reports that larger firms have higher prediction accuracy than smaller firms,

while the overall patterns of predictability across methods are robust in each firm size sub-sample.

The R2 ’s evaluated in the mega firms are roughly twice those of the nano firms. Smaller firms

have a greater magnitude of unexpected trading volume shocks that are hardest to predict. This

result makes sense since small firms are volatile and have low trading volume, hence unexpected

events that give rise to volume spikes are more likely for these firms. This finding indicates that

in addition to small firms being less liquid on average, their liquidity is also less predictable and

more volatile. In other words, their trading costs are less predictable. We examine whether firms

of different size groups should be modeled differently by implementing a mixture of experts (moe)

16
method, which is shown to be beneficial for the linear model but not for the neural networks.

4 Volume alpha: the economic value of volume forecasting

To quantify the economic value of predicting volume, we set up a portfolio problem that features

a tradeoff between tracking a target portfolio versus minimizing trading costs. The key choice is

whether to trade aggressively toward the target or passively to avoid trading costs. The optimal

balancing point depends on the volume forecast and the trading costs associated with it. We

evaluate to what extent more accurate volume forecasts translate to better trading execution.

The incentive to trade is modeled with an objective that penalizes the tracking error toward

a target portfolio. The tracking target, potentially informed by the various return forecasting

signals, is taken as given since the target itself plays a tangential role in the core tradeoff analysis.

Thus, we fix and set aside the return prediction problem to focus on the improvement afforded

by the volume prediction problem. Once this problem is solved, we experiment with a range of

pre-specified tracking targets to evaluate the economic benefit achieved in different tracking tasks,

such as implementing short-term reversal factor portfolios or quantitative strategies.

4.1 Tracking error optimization and its portfolio microfoundation

The tracking error objective is modeled with a simple mean-variance portfolio optimization frame-

work. In this framework, the tracking target is rooted in return forecasts. Regardless of the

mean-variance microfoundation, the tracking error objective is also relevant for circumstances such

as tracking a benchmark index.

Consider a portfolio manager who chooses dollar portfolio positions xi,t in stock i at the start

of day t to maximize a mean-variance certainty equivalence, adjusted for trading costs:

X γ X X
2
A(1 + rf,t ) + xi,t mi,t − xi,t xj,t σij,t − T radingCosti,t , (1)
2A
i i,j i

where A is the manager’s assets under management (AUM) or fund size and γ is her risk aversion

17
coefficient.16 To improve the investment outcome, much empirical work has been devoted to pre-
2 ). Instead, we assume a simple
dicting mean excess returns (mi,t ) and the variance-covariances (σij,t

form for the return moments and take them as given in order to illustrate the economic value solely

of the trading cost term. Assuming Varri = σ 2 , with zero covariances, the objective function is

 2 X  
γσ 2 X A A X 2
− xi,t − mi,t − T radingCosti,t + A(1 + rf,t ) + mi,t . (2)
2A γσ 2 2γσ 2
i i i

The task is to balance the tradeoff between the first term, which is the tracking error penalty as the

result of the mean-variance optimization, and the second term, the transaction cost, to be detailed

below. The third term can be ignored in the optimization since it is irrelevant to the x choices.

The tracking error penalty (first term) is quadratic,

1
T rackingErrori,t := µ(xi,t − x∗i,t )2 , (3)
2

where the target x∗ is the before-cost mean-variance efficient portfolio position, which increases in

the asset’s return expectations as well as the total portfolio size A. In implementation, the trading

target is formed in a separate process without immediate trading cost consideration. We analyze

a general strategy that optimizes the trading rate toward the target based on volume predictions.
γσ 2
Parameter µ := A controls the strength of the tracking error penalty. In later empirical analyses,

we do not calibrate µ from risk coefficients but instead treat it as a hyperparameter and tune the

optimal µ under various AUM levels according to investment performance. Still, the qualitative

relationship is preserved – a larger investor penalizes tracking error (measured in dollars) less,

meaning they trade less aggressively toward the target in general.

Trading costs are modeled as,

1e 0 2
T radingCosti,t := λ i,t (xi,t − xi,t ) , (4)
2
16
We consider a simple situation where the AUM is constant due to an immediate payout program, see Eq. 13 for
detail. In the second term, the risk aversion coefficient is explicitly adjusted by A such that the before-cost Markowitz
optimal dollar position, x∗ , scales up with A.

18
 e as a function of volume: λ
where x0 is the starting position. We specify λ e = 0.2/Ve = 0.2 exp(−e
v ),

following Frazzini, Israel, and Moskowitz (2018). Underlying the quadratic functional form, it is

assumed that the price impact is linear in the trade’s size relative to the volume of the day (a.k.a.
0
participation rate): P riceImpact = 0.1 x−x (Kyle, 1985). For example, buying (or selling) 10% of
e V

the daily volume would move the price by 1% (or −1%). And the trading cost is the price impact

multiplied by the dollar trade size: T radingCost = P riceImpact · (x − x0 ).17

The aggregate tracking error optimization problem is,

X
min (T rackingErrori,t + T radingCosti,t ) . (5)
{xi,t }
i,t

ei,t , is not known when choosing xi,t (emphasized by

Central to our paper, vei,t , or equivalently λ

the tilde). One must predict them based on available conditioning information represented by

predictors Xi,t . Taking x0i,t and x∗i,t as given, the problem becomes {i, t}-separable.18 Then, the

problem is

 
1e 0 2 1 ∗ 2 19
min E λ(x − x ) + µ(x − x ) .
x∈σ(X ) 2 2

4.2 Normalized tracking error and trading rate (z)

To implement this problem empirically, we first normalize the problem by the target trade size

x∗ − x0 so that we analyze the loss for a one-dollar target trade. The problem scales quadratically:

a $1,000-dollar target trade will incur 106 times the loss of a $1-trade. In detail, let the choice
17
For simplicity, we assume away cross-impact on λ e from related stocks as well as other determinants of λ. e Also,
0.2
other functional forms for λ, e= √
e such as the quadratic form often shown empirically, λ = 0.2 exp(− 12 ve), (Frazzini,
V
e
Israel, and Moskowitz, 2018) can also be used. In this case, all the analyses carries through but ve will be twice
as large. We stick with the linear specification consistent with theory (Kyle, 1985) and empirical evidence on the
unexpected component of trading volume (see Frazzini, Israel, and Moskowitz 2018).
18
Ideally, x0i,t should not be taken as given, as it is affected by the choice on the previous day, but we are not
considering the dynamics of the problem here. By taking x0i,t and x∗i,t as given, the problem is {i, t}-separable
and easier to solve. In the trading experiments (Section 6), however, we consider the dynamics by evaluating the
recursively traded portfolios.
19
Here “x ∈ σ(X )” restricts x as a random variable that depends only on predictive information available at the
time of the choice, withh Xi,t := [Xi,t , Xi,t−1 , Xi,t−2i, . . . ]. This unconditional expectation minimization is equivalent
to solving minx∈R E λ(x − x0 )2 + 1 µ(x − x∗ )2 X for each X realization.
2
1e
2

19
 x−x0
variable be trading rate z := x∗ −x0
, then the minimization objective becomes

1e 1 1
2

λ(x − x0 )2 + µ(x∗ − x)2 = x∗ − x0 λze 2 + µ(1 − z)2 . (6)
2 2 2

2
Since the factor, x∗ − x0 , does not matter for the choice of z, define the economic loss as

lossecon (e e 2 + µ(1 − z)2 ,

v , z; µ) := λz (7)

and the normalized problem as

min E [lossecon (e
v , z; µ)] , (8)
z∈σ(X )

which is the main focus of economic machine learning. Being able to separate x∗ , x0 affords many

conveniences. It means the core problem is independent of the target strategy or fund size. It

allows us to look at each i, t observation independently in a volume prediction setting. After the

prediction task is done, we evaluate the investment performance under various pre-specified target

strategies (x∗ ) with different AUM levels.

4.3 The optimal policy ignoring forecast error (function s)

Suppose ve = v and we ignore the inaccuracy in the prediction, the optimal policy is then,

µ 1
s(v; µ) := arg min lossecon (v, z; µ) = = . (9)
z µ+λ 1 + exp (−v + log 0.2 − log µ)

We plot this function in Figure 2. It is a sigmoid function with a horizontal offset determined by

µ. The optimal trading rate z ranges from 0 to 1 as v increases. Then, the optimal dollar position

choice is x = x0 + s(v; µ)(x∗ − x0 ). This is the finding of Gârleanu and Pedersen (2013) that the

optimal strategy should “trade partially toward the aim.”20 In their setting, the trading rate is a
20
The other finding in Gârleanu and Pedersen (2013) is that the optimal strategy should also “aim in front of
the target.” This dynamic effect is abstracted away in our problem since we are considering the {i, t}-separable
optimization. One interpretation is that x∗ is already the aim that is in front of the target that implicitly embeds
the dynamic effect.

20
 Figure 2: The policy function (s) that maps log volume (v) to trading rate (z)

1.0
s(v; µ)
s(v; µ) with a larger µ
0.8

trading rate z 0.6

0.4

0.2

0.0
0 4 8 12 16 20 24
v

The solid curve uses a µ value relevant for $1b AUM. The dashed curve changes to a greater µ
(AUM = $100m), increasing z across the spectrum. The background is the histogram of ve data
repeated from Figure 1, to show that a typical ve corresponds to a ze somewhere in the middle
between 0 and 1 given the chosen µ’s.

fixed constant. Here, it is still irrelevant to either the trading target or the starting point (x∗ , x0 ),

but importantly, it depends on the volume prediction, v. Instead of assuming liquidity as a constant

known by the agent, the innovation and emphasis of this paper is that volume prediction alters the

tradeoff between the cost and benefit of trading, and hence more accurate predictions lead to more

efficient portfolio implementation with a z that varies with forecasted volume. Additionally, the

optimal z also does not explicitly depend on the scale of the fund. If the AUM doubles, and both x∗

and x0 double, the optimal z remains the same, while the dollar position choice x doubles. However,

a smaller fund will find a larger µ more relevant for their investment performance optimization (µ-

tuning detailed further below). In that case, they will trade more aggressively uniformly across v,

as illustrated by the upward displacement of the dashed curve in Figure 2.

5 Machine learning for the economic value of volume prediction

We provide empirical methods to construct the policy of choosing z given X .

21
5.1 The statistical and economic tasks of volume prediction

We consider two ways of approaching the portfolio optimization problem. The first conducts

a statistical prediction of volume and then plugs the volume forecasts into the optimal trading

policy s(v; µ) to form a trading plan. This indirect approach we call “statistical” learning. The

second is an economic learning approach, which instead learns trading rate z as a function of

conditioning information directly to minimize the economic loss. We argue directly choosing z is

also “predicting volume,” but with a different optimization objective rather than the least squares

loss commonly used in statistical predictions. We term this approach “economic” learning. We

show their theoretical differences: the economic loss penalizes inaccuracies in volume forecasting

asymmetrically, where overestimating volume is more costly. Therefore the model optimized for the

economic goal should be more conservative, at the expense of compromising on accuracy measured

in terms of squared errors.

• Approach 1, statistical learning:

Step 1: run machine learning regressions of ve onto X in the training sample as in Section 3

X
v ∗ (·) = arg min vi,t − v(Xi,t ))2 .
(e (10)
v(·)
i,t∈train

Step 2: plug the OOS predictions vbi,t := v ∗ (Xi,t ) into policy equation (9) to trade zbi,t =

s(b
vi,t ; µ).

• Approach 2, economic learning:

parameterize z as a neural network, optimize the economic objective in the training sample

X
z ∗ (·) = arg min lossecon (e
vi,t , z(Xi,t ); µ) (11)
z(·)
i,t∈train

and trade zbi,t = z ∗ (Xi,t ) in the testing sample.21

The difference between the two approaches boils down to the different loss functions deployed in

penalizing volume prediction errors. For example, a trading action zb = z(X ) implies an underlying
21
For both approaches, the machine learning framework can be rnn or nn, in which Xi,t includes concurrent and
lagged predictors (X) or only the concurrent.

22
volume forecast vb = s−1 (b
z ; µ), and equivalently the economic loss can be written as a function of

the z-implied v instead of z itself: lossecon

vv (ev , v; µ) := lossecon (e
v , s(v; µ); µ). Hence, the economic

learning approach is equivalent to first solving

X
min lossecon
vv (evi,t , v(Xi,t ); µ) (12)
v(·)
i,t∈train

followed by the s( · ; µ) transformation, which is also required in the first approach.

Given that the two approaches are only different in the loss functions, we compare lossecon
vv
1
with the least squares loss function, lossls (e
v , v) := 2 (v − ve)2 , used in statistical predictions. To

understand how the two approaches behave differently, first note that the two functions are the

same for the smallest possible loss being attained, which is when the forecast, v exactly equals

the target, ve. In empirical experiments, we label the strategy made with perfect foresight “oracle”

vi,t = vei,t ), which yields the smallest mean squared error (MSE) and
as the unattainable ideal (b

smallest mean economic loss (MEL).

Second, both loss functions monotonically increase as the forecast v deviates away from the true

ve. Therefore, it is intuitive to think that making forecasts that are close to ve in the least squares

sense, will translate to better portfolio implementation as evaluated by the economic loss.22 How-

ever, forecast errors will not guarantee this outcome because of the differences between the two

loss functions. From a theoretical perspective, it is well known that the conditional expectation,
 
v |X ], is the minimizer of the problem minv∈σ(X ) E lossls (e
E[e v , v) , so that the statistical learn-

ing method recovers the conditional expectation with the neural network tools. However given a

different economic loss function,

 
v |X ] = arg minv∈σ(X ) E lossls (e
Proposition 1. The least squares minimizer, E[e v , v) , does not

optimize the economic loss minimization problem minv∈σ(X ) E [lossecon

vv (ev , v; µ)].

Even with unlimited data, the “perfect” statistical learning would not optimize the economic prob-
22
The above mentioned properties expressed mathematically are arg minv loss(v, ve) = ve, ∀e v ; and loss(v, ve) is in-
creasing in |v − ve| , ∀v, ve. Both lossecon
vv and lossls
satisfy these properties. In Figure 3, the dots mark the minimums.
In the right panel, the dips around the minimums are too shallow to be noticeable, though analytically they are
indeed the minimums. See Appendix B.4 for more details on the local curvature around the minimum points.

23
 Figure 3: The statistical (least squares) and economic loss functions

30
lossls (e
v , v) lossecon
vv (ev , v; µ)
10−3
25 minimum point minimum point
10−4 greater µ
20
10−5
15 10−6

10 10−7

10−8
5
10−9
0
0 4 8 12 16 20 24 0 4 8 12 16 20 24
v v

The two panels visualize lossls (e v , v) and lossecon

vv (ev , v; µ), respectively. We pick five different true
values ve = 4, 8, 12, 16, 20 (in five colors), and respectively plot the loss curves for v ∈ [0, 24]. The
dots mark the minimums of the loss curves, attained at v = ve. The right panel is in the log scale.
The solid and dash curves use the µ values in columns 2 and 3 in Table 2 (corresponding to $1b
and $100m AUM), respectively.

lem. The theoretical foundation of this claim is that E[e

v |X ] = arg minv∈σ(X ) E [ϕ(e
v , v)] if and only

if the generic loss function ϕ is in the Bregman class (Banerjee, Guo, and Wang, 2005; Patton,

2020). As is well known, the least squares loss belongs to the Bregman class, the economic loss

function does not (proofs in Appendix B.3).

More intuitively, we plot and compare the two loss functions in Figure 3. We pick five different

true values for ve = 4, 8, 12, 16, 20, and respectively plot the loss curves for a range of forecasts v ∈

[0, 24]. The distinguishing feature of the economic loss is its asymmetric penalty for overestimating

volume when actual volume is low. For example, the blue loss curve (actual ve = 4) is very high if

the forecast is mistakenly large (v > 12). The economic intuition for why this particular forecasting

error is so costly is that it implies trading aggressively (z close to 1) when actual liquidity is low.

Analytically, limz↑1 lossecon (e e which increases exponentially if ve is small.

v , z; µ) = λ,

In contrast, errors in the other direction are much more forgiving (e.g., the purple curve

with ve = 20).23 Trading too little when actual liquidity turns out to be ample delivers a loss
23
Notice the vertical axis is in log scale, meaning the purple is much more flat compared to blue in terms of the
difference of its right and left ends. The same plot in linear scale is in Appendix Figure B.2.

24
equal to the opportunity cost of not better tracking the target portfolio, which is fixed at µ

(limz↓0 lossecon (e
v , z; µ) = µ, ∀e
v ). In summary, for low levels of true volume, overestimating vol-

ume is very costly, but for high levels of volume, the economic cost of volume forecast error is

relatively small.

We state the asymmetric property formally in Proposition 4 in Appendix B.4, and provide

further analysis. The analytical results rely on the quadratic functional forms assumed in equations

(3) and (4). However, the qualitative points carry over to more general settings.

These findings have important implications. Off-the-shelf machine learning tools are not the

most suitable for specific portfolio problems because they minimize statistical error rather than

economic error. Hence, an altered financial machine learning design that accounts for economic loss

can be more effective. The ranking of forecasts evaluated by the two loss functions can be reversed

– a set of forecasts with a smaller sample MSE might have a greater economic loss – something we

will see empirically in Table 2. The asymmetry in the economic loss has important implications for

implementing a model optimized for the statistical criteria to the trading task. The model should

“learn” to be more careful about the potential of a liquidity dry-up and be conservative in avoiding

overestimating volume. When necessary, it should compromise least squares accuracy in favor of

minimizing economic losses.

Both the statistical and economic learning approaches are implemented with neural networks,

given the many benefits of deep learning such as the ability to handle high-dimensional data and

non-linear relations. Next, we emphasize one aspect of the implementation that is particularly

relevant for transferring statistical learning results to finance applications such as portfolio opti-

mization.

5.2 Transfer learning via pre-training and fine-tuning

We implement a transfer learning paradigm in which the statistical and economic learning models

are trained sequentially, as illustrated in Figure 4 right panel.

The statistical volume prediction is an upstream task. It learns valuable but generic information

on the predictive relationship and serves as the foundation model (center node in Figure 4 right

25
panel). It enjoys off-the-shelf machine learning programs optimized for this typical task. The

background knowledge can be transferred to more specific downstream tasks such as portfolio

optimization in our case. The finance-motivated tasks benefit from a good “foundation”. Fine-

tuning the pre-trained foundation model per the economic loss objective further improves the

economic performance. The different specific downstream tasks require separate economic training

routines. (The tasks are different because the economic loss function is modulated by µ.) The

same pre-trained foundation model serves as the common starting point for different downstream

fine-tuning routines.24

Pre-training optimizes a nn or rnn from X to ηb (the blue part in Figure 4 left panel) as described

in the previous section. Using the pre-trained network as is, flowing the output ηb through additional

transformations, “+ma5 ” and “sigmoid s( · ; µ)” shown in black, the resulting zb implements

the plug-in step. Taking this foundation model as the initialization, fine-tuning conducts further

stochastic gradient descent in the economic loss function evaluated at the same training sample. The

resulting fine-tuned network implements the economic learning approach. Only the trainable part

(in blue) is updated in fine-tuning. The economic approach only innovates on the loss function,

not the network architecture or data. Neural networks tackle the non-linearity not only in the

predictive relationship but also in the (marginal) economic loss function. Experiments show fine-

tuning requires only a small number of epochs of training to significantly improve OOS economic

performance compared to the pre-trained foundation model.25

Many other finance problems share a similar structure, in which statistical results are transferred

into actionable strategies that are applied towards an economically motivated problem. Markowitz

portfolio optimization is a classic example. Another example is financial risk management, which

relies on volatility forecasts. The transfer learning procedure adopted here provides a unified

framework for applying machine learning techniques in these scenarios.26 The procedure is also
24
An even lower-level downstream task is to make decisions given target position x∗ . We do not directly train for
that, but do evaluate the performance in such trading experiments further below.
25
Alternatively, side-stepping pre-training but directly training for the economic loss from scratch is generally less
robust and takes more time (epochs over the sample) to train, probably because the machine learning program is not
optimized for such a loss. Not to mention that each µ would require a separate training routine as they do not share
the common pre-trained baseline.
26
Some financial machine learning studies, including Jensen et al. (2022), Chen et al. (2023), Cong et al. (2021)
and Chen, Pelger, and Zhu (2023) involve directly training for the economic target.

26
 Figure 4: Network architecture and transfer learning procedure

training start

statistical learning
sigmoid:
(pre-training for upstream task)

trainable ma5 foundation model

plus economic learning (generic predictive model)
network
(fine-tuning for downstream tasks)
fully-connected each requires different training
residual connection
hidden layers

input layer (predictors )

trading models
The left illustration shows the network architecture. The blue part “trainable network” is a stan-
dard feed-forward neural network, specified with three fully connected hidden layers with 32, 16,
and 8 neurons, respectively. The black parts are non-trainable transformations from ηb to log volume
prediction vb and ultimately trading rate zb. The recurrent connections in rnn are omitted in this
illustration. The right figure illustrates the training procedure for transfer learning. Each dot repre-
sents a trained model, i.e., a parameterization of the network. The arrows show economic learning
as specific fine-tuning steps based on the common foundation model pre-trained statistically.

similar to how GPT models are “P”re-trained on language representations and applications like

ChatGPT are fine-tuned for generating conversational responses or other tasks.

5.3 Economic and statistical prediction results

We present a systematic comparison of how different predictor sets and models perform in terms of

both statistical and economic performance. The results show the economic performance is indeed

optimized by fine-tuning the training process on the same objective, and that more predictors and

the network model lead to improved performance.

Table 2 Panel A evaluates the OOS mean economic loss (MEL), and Panel B the mean squared

error (MSE). For ease of comparison, Panels A′ and B′ normalize these metrics as a percentage

loss reduction such that the “oracle” (which represents the perfect prediction vb = ve) is at 100%

accuracy and the baseline ma5 at 0%. (The percentage reduction in MSE is then the R2 predicting

ηe as in Table 1 Panel A.) The four columns correspond to different µ values that modulate the

economic loss function, representing four different downstream economic tasks relevant to different

27
AUM magnitudes. As the AUM decreases, the relevant economic objective is parameterized by a

greater µ, such that overall trading becomes more aggressive. The average trading rate ranges from

trading conservatively near the starting position (13% for AUM = $10b) to trading almost all the

way to the target position (95% for AUM = $10m).

The rows include the statistical prediction methods described in Section 3 (ma5 , ols, nn, and

rnn). Additionally, the economic learning approaches (those ending with “.econ”) conduct fine-

tuning on top of the statistical forecasts to optimize the economic loss function under the four

different µ’s. The lines spanning the columns indicate that the statistical forecasts do not vary

with µ, whereas economic learning generates different (z-implied) volume forecasts as µ varies.

The predictor sets include the 8 “tech” signals or “all” of the 175 signals.

Looking first at the statistical performances of the forecasts, model complexity and feature rich-

ness help reduce MSE. Configuration rnnall yields the highest R2 ≈ 20%. However, smaller MSEs

do not necessarily lead to better economic performance. For example, nntech accrues greater MELs

than olstech , albeit nn is more accurate in terms of R2 (comparing Panels A and B). For the low

AUM task in particular, the various volume forecasts are even worse than the baseline ma5 , likely

due to the need to trade aggressively in this task and the disproportional penalty when aggressive

trades are the result of overestimating volume. This result is intuitive given the theoretical analysis,

and argues why economic fine-tuning produces better OOS portfolio performance.

The economic training methods lead to better economic outcomes. The fine-tuned networks

with all the predictors yield the best performance, reaching an OOS economic performance that

is about 43%∼70% of the unattainable oracle benchmark at various AUM scales. These improved

OOS outcomes are not mechanically guaranteed since the fine-tuning is to minimize in-sample

loss. The empirical results demonstrate the validity of the economic learning design. Looking at

Panel B′ , the statistical accuracy retreats after fine-tuning, often to levels even worse than the ma5

baseline resulting in negative R2 . This means to achieve improvements in economic outcomes, the

models compromise MSE. Optimal trading actions are conservative against volume overestimation

given the economic loss function.

Furthermore, for a smaller AUM or, equivalently, a higher µ value, the economic models tend

28
 Table 2: Economic and statistical performance of different methods

1 2 3 4 1 2 3 4
µ 1.2e-9 6.3e-8 4.7e-7 9.4e-6 1.2e-9 6.3e-8 4.7e-7 9.4e-6
avg ze 0.13 0.57 0.78 0.95 0.13 0.57 0.78 0.95
relevant AUM $10b $1b $100m $10m $10b $1b $100m $10m
A. Mean economic loss (MEL) (×10−8 ) A′ . % reduction in mean economic loss
ma5 0.1046 3.163 15.41 93.0 0.0 0.0 0.0 0.0
olstech 0.1041 3.011 14.81 93.2 27.9 29.6 11.1 -0.3
nntech 0.1043 3.100 14.97 99.7 19.2 12.3 8.0 -12.8
rnntech 0.1040 2.955 14.37 102.2 33.6 40.8 19.2 -17.7
[Link] 0.1041 2.991 12.35 66.8 31.3 33.6 56.5 50.2
[Link] 0.1039 2.855 11.78 64.2 39.7 60.4 67.0 55.3
olsall 0.1040 3.024 14.97 94.7 32.4 27.2 8.1 -3.3
nnall 0.1040 3.019 15.07 106.5 33.3 28.2 6.2 -25.9
rnnall 0.1040 3.012 14.78 109.8 34.8 29.5 11.7 -32.1
[Link] 0.1039 2.810 11.56 61.9 39.6 69.2 70.9 59.6
[Link] 0.1038 2.812 11.60 66.4 43.7 68.8 70.3 51.0
oracle 0.1029 2.653 9.99 40.8 100 100 100 100

B. Mean squared error (MSE) B′ . R2 (% reduction in MSE)

ma5 0.437 0.0
olstech 0.385 12.1
nntech 0.375 14.3
rnntech 0.368 15.8
[Link] 0.389 0.449 0.457 0.630 11.2 -2.6 -4.5 -44.1
[Link] 0.392 0.492 0.487 0.481 10.3 -12.4 -11.3 -10.0
olsall 0.367 16.0
nnall 0.357 18.4
rnnall 0.350 19.9
[Link] 0.394 0.555 0.590 1.979 10.0 -26.8 -34.9 -352.5
[Link] 0.377 0.440 0.477 0.785 13.9 -0.6 -9.0 -79.5
oracle 0.00 100
Panel A: Mean economic loss (MEL) := avg lossecon (e v , zb; µ); A′ : % reduction in mean economic
loss := (MELma5 − MELm )/(MELma5 − MELoracle ); B: MSE := avg (e v − vbm )2 ; B′ : R2 := 1 −
avg (ev − vbm )2 /avg (e
v − vbma5 )2 = (MSEma5 − MSEm )/MSEma5 , for each method m and µ (“avg” is
the OOS average over i, t). zb varies over each method and µ. For statistical methods, vb does not
depend on µ, hence the horizontal lines indicate the MSE and R2 do not depend on µ for these
methods. These R2 numbers repeat those from Table 1 row “e η ” by construction. In the header,
the two additional rows help interpret the µ values: avg ze = avg s(e v ; µ) is the average trading rate
given true volume; Under each “relevant AUM”, the corresponding µ is backed out in portfolio
optimization hyperparameter tuning (see Footnote 28 for details on tuning µ).

29
introduce more statistical bias, as indicated by the increasingly negative R2 values. This can be

explained by the changes in the economic loss functions. With a smaller µ, trading is more intensive

in general (s curve in Figure 2 shift to the left). That means the penalty for overestimating low

volume is more stringent and takes effect earlier (dashed curves in Figure 3 shift to the left). Thus,

for smaller AUM, there is a greater difference between the least squares loss and the economic loss.

6 Investment performance in trading experiments

We now apply the analysis to real-world investment portfolios in a set of trading experiments.

6.1 Trading experiment design

The trading experiment forms a set of trades xi,t by applying the various trading strategies detailed

in the previous section to dynamically track a set of given target positions x∗i,t . The target positions

are not optimized on trading cost considerations, but formed in a separate process focused solely

on return prediction. We evaluate the outcome of the implemented trades xi,t , including the

mean return, Sharpe ratio, and turnover, in the OOS period. We examine whether a volume

prediction method brings improved investment performance net of trading costs and tracking error

considerations.

Various sets of target positions {x∗i,t } are exogenously supplied to mimic realistic trading tasks.

The first set of experiments mimic a quantitative strategy. We simulate an extremely profitable

before-cost trading strategy assuming the agent can, with some probability, forecast the realized

direction of stock price change. We experiment with different AUM levels, in which the dollar

positions scale linearly while the trading costs increase quadratically. As a result, the optimal

µ, which controls the overall trading aggressiveness, varies. The second set of experiments tracks

monthly-rebalanced factor portfolios sorted on firm characteristics from the literature as the trading

targets. These experiments reveal the effectiveness of volume prediction across the spectrum of

investment styles.

Given the target {x∗i,t }, the implemented trading outcome {xi,t } is constructed following the

30
trading rate strategy, such as zbi,t = [Link](Xi,t ; µ), which is formed in the training sample.

In particular, xi,t = x0i,t + zbi,t (x∗i,t − x0i,t ), where the starting position is formed recursively as

x0i,t := xi,t−1 Ri,t

raw and Rraw = 1 + r
i,t f,t−1 + ri,t is the arithmetic raw return accrued on day t − 1.
27

Note the dynamic effect here: the portfolio choice matters for the starting position on the next day,

although the optimization does not explicitly consider it. Additionally, the target amount to trade

x∗ − x0 , varies across i, t, which is another aspect abstracted from in the theoretical analysis.

Given the resulting trades {xi,t }, the accounting of the investment outcome is standard. The

daily dollar payoff is such that the AUM is fixed over time:

X X
payofft = A(1 + rf,t ) + xi,t rei,t − T radingCosti,t − A. (13)
i i

We normalize by A to get the net-of-cost excess return of the implementation:

payofft X Xλ
ei,t

0 2
reimplemented,t := − rf,t = wi,t rei,t − A wi,t − wi,t , (14)
A 2
i i

xi,t 0 := x0i,t
where wi,t := A and wi,t A are portfolio weights as a ratio of AUM. We have the familiar

result that the before-cost return (the first term) is scale-invariant while the percentage trading

cost due to price impacts (the second term) scales with the AUM linearly. We report the mean and

Sharpe ratio of reimplemented,t . Additionally, we also evaluate the annualized turnover as

1 X xi,t − x0i,t 1 X 0
Turnover := × 252 = wi,t − wi,t × 252. (15)
T 2A 2T
i,t i,t

The second equation indicates turnover is scale invariant.

6.2 Implementing a simulated quantitative strategy

We consider a set of trading targets {x∗i,t } that simulate a quantitative investment strategy. We

simulate a trading signal that, with 1% chance, perfectly forecasts whether a stock goes up or down

over the next five days. The signal is independent across i, t. Following the signal, all stocks are
27
To initiate the recursive calculation, let x0i,t = 0 on the first day a stock appears in the sample.

31
allocated to either the long or the short group. When no signal is received, the stock position stays

the same. We let, x∗i,t be the equal-weighted long-short strategy in which each leg sums up to 50%

of the AUM.

This portfolio has an unrealistically high before-cost OOS Sharpe ratio of around 7, which is

brought down substantially (to more reasonable levels) after trading costs. We experiment with a

sequence of AUM magnitudes up to $10 billion which command varying levels of trading costs. We

first evaluate across a grid of µ values that vary the aggressiveness of trades, and then perform the

µ-tuning method that allows for the comparison of the highest investment performance attained

by each volume prediction method.

We run a trading experiment for every method and each µ value for various AUM levels. Each

experiment plots one dot in Figure 5, with OOS turnover versus mean return (or Sharpe ratio) as

the coordinates. Each curve corresponds to one method, connecting the dots with varying µ values.

To understand the general shape of the curves, first, note that under different µ values, the

strategies vary from always passively holding (z = 0) to trading all the way to the target (z = 1).

As the implemented portfolio becomes more aggressive (µ increase), the turnover increases. On the

vertical axis, the mean return (or Sharpe ratio) first increases as a result of active trading for profit

and then bends downward due to trading costs, whose effect shows up with high turnover. The two

opposing forces result in the inverted U-shaped curves. For higher AUM, the trading costs’ effect is

stronger, resulting in lower inverted U’s that peak at smaller turnover levels (i.e., lower µ values).

In the next subsection, We implement the highest attainable investment outcome by selecting the

µ value from the peak of the in-sample curves.

The investment gain of a better trading rate strategy z(X ) is shown in the vertical displacement

of the curves. The improvement comes from two aspects. First, more closely tracking the target

portfolio delivers a higher return. Second, reducing trading costs when liquidity is expected to

be low, for a given amount of turnover, lowers the trading cost drag on the portfolio. These two

objectives are contradictory in our framework, hence the value of predicting volume more accurately

is to strike a better balance in the tradeoff between trading costs and tracking error. At the two

extremes of the curves (µ = 0 or +∞), there is no room for tradeoff, and hence predicting volume

32
 Figure 5: Trading experiment performance, with only tech predictors
Mean return after tcost (%, annualized) AUM = $10b AUM = $1b AUM = $100m

A B C
10

6 D E F method:
oracle
ma5
Sharpe ratio (annualized)

olstech
nntech
rnntech
3 [Link]
[Link]

0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5

Turnover (annualized) Turnover (annualized) Turnover (annualized)

Figure 6: Trading experiment performance, with all predictors

AUM = $10b AUM = $1b AUM = $100m
Mean return after tcost (%, annualized)

A B C
10

6 D E F method:
oracle
ma5
Sharpe ratio (annualized)

olsall
nnall
rnnall
3 [Link]
[Link]

0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5

Turnover (annualized) Turnover (annualized) Turnover (annualized)

Each dot plots the outcome of one trading experiment: the horizontal coordinate is turnover,
vertical is the mean return or Sharpe ratio. Each curve is for a method with varying µ values.
Figures 5 and 6 differ in using only the tech or all sets of predictors, while the benchmarks curves
“ma5 ” and “oracle” are the same.

33
 Table 3: Investment performance in trading experiments

A. Mean return (%, annualized) B. Sharpe ratio (annualized)

AUM $10b $1b $100m $10m $10b $1b $100m $10m
ma5 3.88 6.47 11.19 13.20 2.00 2.21 5.47 6.55
olstech 3.82 7.60 11.28 13.14 2.16 3.32 5.59 6.59
nntech 3.76 7.30 11.32 13.13 2.14 2.79 5.63 6.59
rnntech 3.74 7.84 11.33 13.13 2.18 3.58 5.64 6.59
[Link] 4.60 7.20 11.29 13.22 2.13 2.57 5.59 6.63
[Link] 4.67 8.69 11.59 13.30 2.50 4.32 5.73 6.66
olsall 3.82 7.60 11.28 13.14 2.17 3.35 5.60 6.59
nnall 3.86 7.44 11.28 13.13 2.19 3.09 5.60 6.58
rnnall 3.79 7.55 11.25 13.09 2.18 3.26 5.59 6.56
[Link] 4.64 8.87 11.61 13.29 2.18 4.24 5.74 6.66
[Link] 4.68 8.95 11.77 13.30 2.50 4.53 5.85 6.68
oracle 6.47 9.89 12.54 13.56 3.05 4.97 6.28 6.80
For each combination of AUM and method, we report the OOS mean return (Panel A) and Sharpe
ratio (Panel B) at the tuned µ, which is selected to maximize the in-sample mean return (or Sharpe
ratio) over a grid of µ values.28

has no value, in which case all curves converge.

Comparing the methods, we can analyze the sources of improvement. First, using the full set of

predictors is better than using only the set of technical predictors, which is still much better than

no additional information besides the five-day moving average. Second, holding the information

set constant, the neural network model predicts volume better than linear regression. Lastly, given

the pre-trained neural network, further fine-tuning the trading strategy by directly optimizing the

investment performance provides further improvement.

Finally, we tune the hyperparameter µ to report each method’s highest attainable investment

performance. We choose the µ value that maximizes the in-sample expected return (or Sharpe

ratio) and report the out-of-sample performance at the tuned µ. Effectively, µ is selected from the

peaks of the in-sample version of the curves (not plotted) and then applied to the OOS curves (as

plotted). We expect the in-sample and OOS curves to peak at relatively close µ ranges so that the

method attains an OOS performance close to the peak of the OOS curves.
28
The “relevant AUM” row in Table 2 is calculated according to the µ-tuning result under each AUM level under
method ma5 .

34
 The results are reported in Table 3. Applying better prediction methods and using a larger

set of predictors improves investment performance uniformly across AUM levels. The economic

magnitude of improvement is significant, comparable to if not more than the marginal improvement

from innovating on return prediction signals. For $10b of AUM, the average annual return increases

from 3.88% when using the baseline prediction method of only a 5-day moving average volume, to

4.68% when making volume predictions with an economic objective optimization imbedded within

a rnn. For $1b AUM, the magnitude of improvement is even greater, going from 6.47% to 8.98%,

and more than doubling the Sharpe ratio from 2.21 to 4.53.

For smaller AUM ($10m), the improvement is still noticeable but smaller, because price impact

shrinks and all methods therefore prescribe trading very aggressively. The investment performance

converges to the high before-cost level regardless of the prediction method. For more realistic

considerations, future research could consider per-unit trading costs such as bid-ask spread in

addition to price impacts, which tend to show up as dollar trade sizes shrink.

6.3 Implementing factor zoo portfolios

As another set of trading experiments, we use as trading targets the portfolios sorted on character-

istics in the JKP dataset that come from the asset pricing literature. The goal is to examine the

improvement in implementation outcomes across different investment styles.

For each of the 153 characteristics, the target {x∗i,t } is formed in a standard fashion without

considering trading costs: at the start of every month, the cross section of stocks is split at the 50th

quantile into equal-weighted long-short portfolios based on each characteristic. We fix the AUM at

$10 billion ($5 billion for each of long and short legs).

We fix µ across the 153 factors for consistent comparison, and because factor-specific µ tuning

is likely to be unstable. For example, consider factors that happen to earn negative realized returns

during the training sample period, the factor-specific optimal µ would be zero – no trading at all.

By evaluating at a fixed positive µ, we can address whether this factor that happens to lose money

in the sample, would lose less with a better implementation. We pick the µ that optimizes the

average gain across all factors, with results robust to perturbations in µ.

35
 Figure 7: Mean return improvements in implementing each factor portfolio

A. implementing each factor portfolio B. averaged by theme clusters

ret_1_0
Mean return gain, [Link] vs. ma5 (%, annualized)

1.0 zero_trades_21d
iskew_ff3_21d
1.0 Short-Term Reversal
rskew_21d
iskew_hxz4_21d
rmax5_rvol_21d iskew_capm_21d
turnover_var_126d
dolvol_var_126d
lti_gr1a
seas_1_1an
0.8 0.8
inv_gr1a
ret_3_1
ivol_ff3_21d rmax1_21d coskew_21d
taccruals_ni ivol_hxz4_21d seas_2_5an
rvol_21d rmax5_21d
ivol_capm_21d
0.6 age
cash_atturnover_126d
aliq_mat
zero_trades_126d
taccruals_at niq_be_chg1 resff3_6_1 beta_dimson_21d 0.6
at_be
oaccruals_ni
z_score oaccruals_at
op_at dsale_dsga
niq_at_chg1
ret_12_7 Profitability Low Risk
ope_bel1
qmj_prof
op_atl1
ebit_sale
ni_ar1
mispricing_perf
eqnetis_at
qmj_safety prc_highprc_252d
bidaskhl_21d seas_16_20an
seas_11_15an Accruals
Low Leverage Profit Growth Seasonality
ebit_bev
netdebt_me
ivol_capm_252d netis_at
dsale_drec
niq_at betadown_252d
qmj_growth
cop_at niq_be nfna_gr1a
ni_me be_me f_score
ocfq_saleq_std resff3_12_1
ocf_at
ni_be at_me cowc_gr1a
chcsho_12mocf_at_chg1
cop_atl1
zero_trades_252d
qmj
beta_60m
o_score eqnpo_12m
dgp_dsale seas_6_10an Quality Momentum
0.4 eqnpo_me
ope_be bev_mev
sale_emp_gr1
ebitda_mev
gp_atl1 eqpo_me
earnings_variability
capx_gr2 seas_16_20na
fnl_gr1a
mispricing_mgmt dbnetis_at 0.4 Value
saleq_gr1 niq_su
debt_gr3 coa_gr1a
at_turnover
debt_me
opex_ateq_dur
ocf_me
fcf_me
noa_at
col_gr1a
sale_gr1
dsale_dinv ret_12_1 Debt Issuance
gp_at
kz_index
sale_bev
tangibilityppeinv_gr1a
noa_gr1a
ival_me
capex_abn
emp_gr1
be_gr1a
lnoa_gr1a
saleq_su
seas_11_15na ret_6_1 Investment
dolvol_126d inv_gr1
capx_gr3 capx_gr1
prcaliq_at
rd_saleni_ivol
pi_nix
ami_126d
sale_gr3
market_equity at_gr1
sti_gr1a seas_2_5na
seas_6_10na
ret_9_1
Size
0.2 sale_me
rd5_atrd_me ret_60_12 nncoa_gr1a
seas_1_1na 0.2
betabab_1260d
ni_inc8q
ncoa_gr1atax_gr1a
corr_1260d

ncol_gr1a
0.0 div12m_me 0.0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Turnover (annualized) Turnover (annualized)
Each dot implements one JKP factor portfolio. The y-axis is the difference in after-cost mean
excess return between implementing with the [Link] and the ma5 . The x-axis is the turnover
of the factor portfolio target (i.e., Eq. 15 with xi,t = x∗i,t , x0i,t = x∗i,t−1 ). Panel B averages the points
in A by style clusters (from JKP).

Figure 7 plots the gain in mean return when implementing the factor portfolios with the

[Link] volume prediction compared with using the ma5 volume prediction. The horizontal

axis is the turnover of the target factor portfolio. Panel B averages the points by investment style

clusters (from JKP). The plots show volume prediction from the [Link] benefits portfolio imple-

mentation across the factors. The average gain in mean after-cost return across factors is 0.44% per

year from volume prediction alone using the [Link] model versus the simple moving average.

Almost all of the 153 factors have positive gains. With the $10b AUM scale, this translates into

an additional $44m per year in implementation cost-saving from improved volume prediction.

Across factors, the gain is larger for those factors with higher turnover. In the right region of

Figure 7, some raw factors have a turnover approaching six (600% per year or roughly turning over

half of the AUM every month). The gains for these factors, including various short-term reversal

strategies, are around 0.5% to 1.0%. These factors are constructed with technical signals over a

shorter window.29 In the left part of the figure, even factors with low turnover (those relying on
29
Examples: ret 1 0 short term reversal; iskew capm 21d idiosyncratic skewness from the CAPM; iskew ff3 21d
idiosyncratic skewness from the FF3F; rmax5 rvol 21d highest 5 days of return; rskew 21d return skewness 21d;
seas 1 1an 1 Year Annual Seasonality; and coskew 21d coskewness.

36
quarterly fundamental signals and signals with greater persistency) show gains that range from

0.2% to 0.6% per year from volume prediction.

Appendix Figure C.3 plots the same gains in the vertical axis but changes the horizontal axis to

the mean return attained by ma5 , i.e., the baseline level in the gain calculation. The figure shows

that, regardless of the baseline, the gain is independently distributed around a positive center.

That is, a better volume prediction is uniformly effective, and the improvement is not concentrated

on factors that have positive (or negative) realized returns. Appendix Figure C.4 reports similar

plots with gains measured in Sharpe ratio space instead of mean returns.

7 Conclusion

We translate volume predictability into net-of-cost portfolio performance by linking it to expected

trading costs – a term we call “trading volume alpha.” Volume is highly predictable, especially

when using machine learning techniques, large data signals, and exploiting the virtue of complexity

in prediction. We find that volume prediction can be as valuable as return prediction in achieving

optimal mean-variance portfolios net of trading costs.

We find that incorporating an economic objective function directly into machine learning is

even more effective for obtaining useful predictions. This feature may be general to many finance

applications of machine learning, where incorporating the economic objective directly may dominate

a two-step process that first satisfies some statistical objective and then incorporates that statistical

object into an economic framework. For volume prediction, the asymmetric cost of overestimating

versus underestimating volume is captured (ignored) by an economic (statistical) objective, and

delivers sizeable economic impact.

While we find substantial economic benefits from volume predcition using our framework and

methods, there is much room for improvement. Our goal is not to develop the best trading cost

model or even the best volume prediction model, but rather to translate the prediction problem into

economic consequences, which yield interesting insights. A more exhaustive search for prediction

variables and models that forecast volume more accurately could translate into even larger economic

37
benefits than we show here. Some promising candidates for additional features and methods are

lead-lag volume relations across stocks, more seasonal indicators, other market microstructure

variables, and more complex nn and rnn models.

Our simple framework for predicting and characterizing volume alpha also has limitations. For

one, we study a very simple functional form for trading costs that maps volume prediction directly

into costs. Other functional forms and other determinants of costs beyond volume may lead to

novel results. In addition, we separate the volume prediction problem from the expected return

and variance/covariance modeling problem. Combining all three could generate further portfolio

improvements and the interaction between these three prediction problems could be enlightening.

Lastly, our trading experiments are merely a tool to illustrate some possible applications of our

insights, but are not designed to optimize any performance outcome. Specifically, two things not

considered in our design are: dynamic effects from trading and heterogeneous trade tasks. Adding

these more complex features would be an interesting area for future research.

Trading volume prediction in general is an interesting research area worthy of further explo-

ration. While we have couched the volume prediction problem into a portfolio context to translate

the problem into economic consequences, understanding the role of volume more generally – its

causes and consequences – is interesting. Using some of our techniques may shed light on this

question and may help pinpoint what components of volume are most valuable to trading costs

and, as a byproduct, portfolio construction. For example, informed versus uninformed volume,

volume with temporary versus permanent price impact, and short- versus long-term volume may

be interesting research pursuits. Examining various aspects of volume could be very useful for

improving portfolio optimization and understanding trading activity more broadly. We leave these

issues for future work.

38
References
Amihud, Y. 2002. Illiquidity and stock returns: Cross-section and time-series effects. Journal of
Financial Markets 5:31–56.

Balduzzi, P., and A. W. Lynch. 1999. Transaction costs and predictability: Some utility cost
calculations. Journal of Financial Economics 52:47–78.

Banerjee, A., X. Guo, and H. Wang. 2005. On the optimality of conditional expectation as a
Bregman predictor. IEEE Transactions on Information Theory 51:2664–9. Conference Name:
IEEE Transactions on Information Theory.

Benston, G. J., and R. L. Hagerman. 1974. Determinants of bid-asked spreads in the over-the-
counter market. Journal of Financial Economics 1:353–64.

Brennan, M. J., and A. Subrahmanyam. 1995. Investment analysis and price formation in securities
markets. Journal of Financial Economics 38:361–81.

Campbell, J. Y., S. J. Grossman, and J. Wang. 1993. Trading volume and serial correlation in
stock returns. The Quarterly Journal of Economics 108:905–39.

Çetin, U., R. A. Jarrow, and P. Protter. 2004. Liquidity risk and arbitrage pricing theory. Finance
and Stochastics 8:311–41.

Chen, H., Y. Cheng, Y. Liu, and K. Tang. 2023. Teaching economics to the machines. Available
at SSRN 4642167 .

Chen, L., M. Pelger, and J. Zhu. 2023. Deep learning in asset pricing. Management Science .

Chordia, T., S.-W. Huh, and A. Subrahmanyam. 2007. The cross-section of expected trading
activity. The Review of Financial Studies 20:709–40.

Chordia, T., R. Roll, and A. Subrahmanyam. 2011. Recent trends in trading activity and market
quality. Journal of Financial Economics 101:243–63.

Cong, L. W., K. Tang, J. Wang, and Y. Zhang. 2021. AlphaPortfolio: Direct construction through
deep reinforcement learning and interpretable AI. Available at SSRN 3554486 .

Datar, V. T., N. Y. Naik, and R. Radcliffe. 1998. Liquidity and stock returns: An alternative test.
Journal of Financial Markets 1:203–19.

DeMiguel, V., A. Martin-Utrera, F. J. Nogales, and R. Uppal. 2020. A transaction-cost perspective

on the multitude of firm characteristics. The Review of Financial Studies 33:2180–222.

Engle, R. 2004. Risk and volatility: Econometric models and financial practice. American Economic
Review 94:405–20.

Frazzini, A., R. Israel, and T. J. Moskowitz. 2012. Trading costs of asset pricing anomalies.
Fama-Miller Working Paper, Chicago Booth Research Paper .

———. 2018. Trading Costs. doi:10.2139/ssrn.3229719.

39
Glosten, L. R., and L. E. Harris. 1988. Estimating the components of the bid/ask spread. Journal
of Financial Economics 21:123–42.

Glosten, L. R., and P. R. Milgrom. 1985. Bid, ask and transaction prices in a specialist market
with heterogeneously informed traders. Journal of financial economics 14:71–100.

Goldstein, I., C. S. Spatt, and M. Ye. 2021. Big data in finance. The Review of Financial Studies
34:3213–25.

Grossman, S. J., and M. H. Miller. 1988. Liquidity and market structure. Journal of Finance
43:617–33.

Gu, S., B. Kelly, and D. Xiu. 2020. Empirical asset pricing via machine learning. The Review of
Financial Studies 33:2223–73.

Gârleanu, N., and L. H. Pedersen. 2013. Dynamic trading with predictable

returns and transaction costs. Journal of Finance 68:2309–40. eprint:
[Link]

———. 2016. Dynamic portfolio choice with frictions. Journal of Economic Theory 165:487–516.

Harvey, C. R., Y. Liu, and H. Zhu. 2016. . . . and the cross-section of expected returns. The Review
of Financial Studies 29:5–68.

He, K., X. Zhang, S. Ren, and J. Sun. 2016. Deep residual learning for image recognition. In
Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 770–8.

Ho, T. S., and H. R. Stoll. 1983. The dynamics of dealer markets under competition. Journal of
Finance 38:1053–74.

Hochreiter, S., and J. Schmidhuber. 1997. Long short-term memory. Neural Computation 9:1735–
80.

Jensen, T. I., B. Kelly, and L. H. Pedersen. 2022. Is there a replication crisis in finance? Journal
of Finance .

Jensen, T. I., B. T. Kelly, S. Malamud, and L. H. Pedersen. 2022. Machine learning and the
implementable efficient frontier. doi:10.2139/ssrn.4187217.

Kelly, B., S. Malamud, and K. Zhou. 2024. The virtue of complexity in return prediction. Journal
of Finance 79:459–503.

Kelly, B., and D. Xiu. 2023. Financial machine learning. Foundations and Trends in Finance
13:205–363.

Korajczyk, R. A., and R. Sadka. 2004. Are momentum profits robust to trading costs? Journal of
Finance 59:1039–82.

Kyle, A. S. 1985. Continuous auctions and insider trading. Econometrica 53:1315–35. Publisher:
[Wiley, Econometric Society].

40
McLean, R. D., and J. Pontiff. 2016. Does academic research destroy stock return predictability?
Journal of Finance 71:5–32.

Novy-Marx, R., and M. Velikov. 2016. A taxonomy of anomalies and their trading costs. The
Review of Financial Studies 29:104–47.

Patton, A. J. 2020. Comparing Possibly Misspecified Forecasts. Journal of

Business & Economic Statistics 38:796–809. Publisher: Taylor & Francis eprint:
[Link]

Shleifer, A., and R. W. Vishny. 1997. The limits of arbitrage. Journal of Finance 52:35–55.

Stoll, H. R. 1978. The supply of dealer services in securities markets. Journal of Finance 33:1133–51.

41
 Internet Appendix

Trading Volume Alpha

Ruslan Goyenko Bryan Kelly Tobias Moskowitz Yinan Su Chao Zhang

Yale, AQR, and Yale, AQR, and

McGill Johns Hopkins HKUST (GZ)
NBER NBER

A Technical details

A.1 Neural network implementation details

The nn architecture consists of three fully-connected hidden layers with 32, 16, and 8 neurons,

respectively.

The rnn architecture is similar to that of the nn. The first (bottom) hidden layer in the 3-

layer network is upgraded to an lstm layer with 32 hidden states and cell states, respectively. The

remaining two layers are unchanged: fully connected with 16 and 8 neurons, respectively.

The formulas for the number of parameters in nn and rnn, as reported in Table 1 Panel C, are

stated below. For nn (three fully-connected hidden layers with 32-16-8 neurons), the formula is

(# of predictors + 1) × 32 + (32 + 1) × 16 + (16 + 1) × 8 + (8 + 1). In rnn, the first hidden layer

has 32 hidden states and 32 cell states with four gates, changing the formula to (# of predictors +

32 + 1) × 32 × 4 + (32 + 1) × 16 + (16 + 1) × 8 + (8 + 1).

In training the rnn, we implement the “many-to-one” type data pipeline, where the model

recursively processes a sequence of 10 inputs, Xi,t−9 , . . . , Xi,t , and produces a single output ηbi,t to

calculate the training loss at each data point {i, t}. (When increasing the sequence length from

10 to 50, the results had minimum improvements but required much larger GPU memory and

longer training time.) For data points at the beginning of a stock’s observed period where lagged

predictors (e.g., Xi,t−9 ) are not available, we fill in with zero vectors.

42
 In training both nn and rnn models, we use the Adam optimizer, with default learning rates and

other parameters. The batch size is 1024. The ηe prediction models are trained with 50 epochs. For

the sake of clear benchmarking, we do not adopt early stopping, weights dropout, or hyperparameter

tuning with cross-validation, though these techniques could further boost the prediction accuracy.

The machine learning program is implemented with the PyTorch package.

Figure A.1: Learning curves

A. nnall B. rnnall
22 In-Sample In-Sample
Out-of-Sample 25 Out-of-Sample
21 24

20 23

19 22
R2

R2
18 21

20
17
19
16

0 10 20 30 40 50 0 10 20 30 40 50
Epoch Epoch

R2 of the ηe prediction models as training progresses (epochs).

The learning curves show the gap between the IS and OOS R2 is relatively small, and does not

widen with continued training. This indicates limited in-sample overfitting at this neural network

configuration. The rnn learning curves show slightly more severe overfitting, though the OOS

learning curve is still relatively stable as training continues. The learning curves display fluctuation

due to the randomness of the stochastic gradient descent, though the extent to which is not severe.

We also note the exact results depend on the inherent randomness of the training program. We find

the quantitative results are insensitive to random seeds and report the average of five independent

runs to obtain a robust evaluation outcome.

Table A.1 summarizes computational costs in terms of the training times and memory usage

for nn and rnn with all predictors for a single random seed. These experiments were conducted on

a system equipped with an Nvidia A100 GPU with 40 GB of GPU memory, an AMD EPYC 7713

64-Core Processor @ 1.80GHz with 128 cores, and 1.0TB of RAM, running Ubuntu 20.04.4 LTS.

43
 Table A.1: Training time and memory usage

Training time (hours) CPU memory usage (GB) GPU memory usage (GB)

nnall 0.48 10.87 1.22

rnnall 0.63 144.98 1.48

B Additional theoretical analysis

B.1 Microfoundation of the tracking error part in the portfolio objective

The track error penalty term in the portfolio optimization objective function can be economically

founded. We show the quadratic tracking error penalty can be derived from a mean-variance utility

function. This analysis connects the target positions x∗i to the before-cost mean-variance efficient

portfolio weight, which increases in the asset’s return expectations as well as the total portfolio size

(AUM). It also provides a microfoundation of the hyperparameter µ and the quadratic penalty: as

the position deviates away from the target, the mean-variance loss increases in a quadratic fashion.
γ
Assume the agent has a mean-variance utility function, U = EA′ − 2A VarA
′ − T Cost, where

A′ is the before-cost investment outcome, A is the initial wealth (AUM), and γ is the risk aversion

coefficient, T Cost is the term for the objective of minimizing transaction costs (not the focus

here as we are justifying the tracking error part). A′ is the outcome of the portfolio strategy:
P
A′ = A(1 + rf ) + i xi ri , where xi is the dollar position in risky asset i with excess return ri .

Assume Eri = mi , Varri = σ 2 , and zero covariances. The agent’s portfolio optimization problem

is choosing {xi } to maximize U .

Then, the objective function is

X γ X 2 2
U = A(1 + rf ) + xi mi − xi σ − T Cost (16)
2A
i i
X  γσ 2 
2
= − x + mi xi + A(1 + rf ) − T Cost (17)
2A i
i
 
γσ 2 X 2 2A
=− xi − mi xi + A(1 + rf ) − T Cost (18)
2A γσ 2
i

44
  2 !
γσ 2 X A A X 2
=− xi − mi + mi + A(1 + rf ) − T Cost (19)
2A γσ 2 2γσ 2
i i

The first term matches the tracking error term modeled in Eq. 3. The second term is the before

TCost utility at the zero tracking error portfolio (x = x∗ ). This term is constant of the x choice,

hence can be ignored in the optimization problem.

Comparing the first term with the tracking error modeled in Eq. 3, we see the target portfolio

is x∗i = A
m.
γσ 2 i
As expected, the target positions are the result of Markowitz mean-variance opti-

mization. They are proportional to the return expectations in the cross section. They scale linearly

with the AUM (A) and inversely with the risk aversion coefficient and volatility. Additionally, the
γσ 2
overall tracking error penalizing coefficient µ = A , which is decreasing in AUM. The rationale is

that the quadratic penalty stems from the quadratic risk change as the position deviates away from

the target, and that the absolute risk aversion coefficient decreases with the wealth level. Although

the main analysis takes µ as a hyperparameter ignoring its microfoundation, we still observe the

negative relationship between the tuned µ and the economically relevant AUM (e.g., in Section

5.3).

B.2 The economic task as predicting ze

P econ (e
We have shown the economic task of choosing the z strategy, minz(·) i,t∈train loss vi,t , z(Xi,t ); µ),
P
can be seen as a prediction task of predicting the ve with the economic loss function: minv(·) i,t∈train

lossecon
vv (evi,t , v(Xi,t ); µ). In this appendix, we provide the equivalent representation as a prediction

problem of the oracle trading rate ze := s(e

v ; µ).

Define lossecon
zz (ez , z; µ) := lossecon (s−1 (e
z ; µ), z; µ), where s−1 ( · ; µ) is inverse of s( · ; µ) function.

Under this definition, the economic task can be viewed as the problem of looking for a function

z(·) that maps X into z to minimize the training sample average loss:

X
min lossecon
zz (ezi,t , z(Xi,t ); µ) (20)
z(·)
i,t∈train

45
According to the definition, the analytical expression of lossecon
zz (ez , z; µ) is

µ
lossecon
zz (ez , z; µ) = (z − ze)2 + µ(1 − ze). (21)
ze

µ
In this expression, the loss can be seen as the squared z prediction error weighted by ze . The

last term is constant of choice z so can be ignored in optimization. It equals lossecon

zz (ez , ze; µ), the

baseline loss incurred even with the perfect prediction.

e or ve.
To derive this expression, notice ze is already defined as the perfect trading given λ

µ e= µ
ze = =⇒ λ −µ (22)
e
µ+λ ze

e with ze and then complete the square:

Then, start from Eq. 7, represent λ

lossecon e 2 + µ(1 − z)2

zz (ez , z; µ) = λz
µ 
= − µ z 2 + µ(1 − z)2
ze
µ
= (z − ze)2 + µ(1 − ze)
ze

We show lossecon
zz (ez , z; µ) is still meaningfully different from the standard squared error loss

function, and that E [e

z |X ] will not be the optimal choice either further below.

B.3 Economic loss functions are not in Bregman class

We show functions lossecon

vv (ev , v; µ) and lossecon
zz (ez , z; µ) are not in the Bregman class, for all µ.

Without loss of generality, any loss function F (p, q) can be normalized as F̄ (p, q) := F (p, q) −

F (p, p), such that that the function acquires the convenient property that F̄ (p, p) = 0, and that

the solution to the optimization problem minq∈σ{X } E [F (p, q)] does not change. Therefore, in the
econ
v , v; µ) := lossecon
following propositions, we normalize accordingly and consider lossvv (e vv (ev , v; µ) −
econ
lossecon
vv (e z , z; µ) := lossecon
v , ve; µ) and losszz (e zz (ez , z; µ) − lossecon
zz (ez , ze; µ).

The definition of the Bregman loss function is (Banerjee, Guo, and Wang, 2005):

46
Definition 1. Let ϕ : Rd → R be a strictly convex differentiable function, then, the Bregman loss

function Dϕ : Rd × Rd → R is defined as:

Dϕ (p, q) := ϕ(p) − ϕ(q) − ⟨p − q, ▽ϕ(q)⟩ (23)

We consider the simpler case where p, q are scalars. In this case, a Bregman function has

the property that its partial second derivative in the first argument is independent of the second

argument.

∂ 2 Dϕ (p, q)
= ϕ′′ (p) (24)
∂p2

The propositions below rely on this property.

econ
Proposition 2. Function losszz (e
z , z; µ) is not in the Bregman class, for all µ.

econ
Proof. We verify that losszz (e
z , z; µ) violates the property in Eq. 24.

econ µ
losszz (e
z , z; µ) = (z − ze)2 (25)
ze
econ
∂ 2 losszz (e z , z; µ) 2µz 2
2
= 3 (26)
∂e
z ze

It is clear that this is not irrelevant to z.

econ
Proposition 3. Function lossvv (e
v , v; µ) is not in the Bregman class, for all µ.

Proof. Given the property in Eq. 24, a Bregman loss function must be unbounded as p → +∞.

This is because for any fixed q, when p > q, Dϕ (p, q) is increasing and convex in p.

However, we verify that lossecon

vv (ev , v; µ) is bounded by showing it converges to a finite number

as ve → +∞.

econ 0.2 exp (−ev ) + µ (exp (−v + log 0.2 − log µ))2
lossvv (e
v , v; µ) :=
(1 + exp (−v + log 0.2 − log µ))2
0.2 exp (−e v + log 0.2 − log µ))2
v ) + µ (exp (−e
−
(1 + exp (−e v + log 0.2 − log µ))2

47
 (0.2µ)2 exp (−e v ) − exp (v))2
v ) (exp (e
= (27)
v ) + 0.2) (µ exp (v) + 0.2)2
(µ exp (e

By L’Hôpital’s rule, we have the limit of it as:

econ (0.2µ)2 (exp (e

v ) − exp (2v − ve))
lim lossvv (e
v , v; µ) = lim
ṽ→∞ e→∞
v µ (µ exp (v) + 0.2)2 exp (e
v)
0.04µ
= (28)
(µ exp (v) + 0.2)2

B.4 Further analysis on the loss functions

The following is a visualization of the economic loss function in addition to the one in Figure 3

Panel B. The vertical axis is changed to the linear scale from the log scale. For large ve (12, 16, 20,

in green, red, purple), the curves are indistinguishable from a flat line because the blue curve is at

a much greater magnitude, which is for the loss of overestimating low actual volume (e
v = 4).

Figure B.2: Economic loss function in linear scale

lossecon
vv (ev , v; µ)
0.0035 minimum point
greater µ

0.0030

0.0025

0.0020

0.0015

0.0010

0.0005

0.0000

0 4 8 12 16 20 24
v

Note: same as Figure 3 Panel B but with the vertical axis in linear scale.

The following proposition formally states the asymmetric property of the loss function for

over/under-estimating volume.

48
Proposition 4. Consider two symmetrical cases with low and high liquidity ve1 and ve2 such that

ze1 = 1 − ze2 < 0.5. Suppose one makes an overestimation in the low-volume case vb1 = ve1 + ε,

comparing with an equal amount of underestimation in the high-volume case vb2 = ve2 − ε, the

additional loss incurred in the first case is greater than the second:

lossecon
vv (ev1 , vb1 ; µ) − lossecon
vv (ev1 , ve1 ; µ) > lossecon
vv (ev2 , vb2 ; µ) − lossecon
vv (ev2 , ve2 ; µ), ∀µ > 0. (29)

Proof. We first show that, given ze2 = 1 − ze1 , as well as vb1 − ve1 = ve2 − vb2 = ε, we have zb2 = 1 − zb1 ,

and that ze1 − zb1 = zb2 − ze2 .

1
We know s(v; µ) = 1+exp(−v+log 0.2−log µ) . From ze1 = 1 − ze2 , we have:

1 1
=1−
1 + exp(−e
v2 + log 0.2 − log µ) 1 + exp(−e
v1 + log 0.2 − log µ)

v1 − ve2 + 2 log 0.2 − 2 log µ) = 1

exp(−e

ve2 = 2(log 0.2 − log µ) − ve1 (30)

Then we have:

1 1
zb1 + zb2 = +
1 + exp(−e
v1 − ε + log 0.2 − log µ) 1 + exp(−ev2 + ε + log 0.2 − log µ)
1 1
= + =1
1 + exp(−e
v1 − ε + log 0.2 − log µ) 1 + exp(e
v1 + ε − log 0.2 + log µ)

1 1
The last equation comes from the fact that 1+exp(x) + 1+exp(−x) = 1. Then we make use of the loss

function expressed in terms of ze and z, as defined in Eq. 21:

µ
lossecon
zz (ez , z; µ) = (z − ze)2 + µ(1 − ze) (31)
ze

µ µ
Since ze1 − zb1 = zb2 − ze2 and ze1 < ze2 , that is ze1 > ze2 , we have the required result:

lossecon
vv (ev1 , vb1 ; µ) − lossecon
vv (ev1 , ve1 ; µ) > lossecon
vv (ev2 , vb2 ; µ) − lossecon
vv (ev2 , ve2 ; µ), ∀µ > 0. (32)

49
C Additional empirical results

C.1 Prediction results in firm size groups and “mixture of experts” forecasts

Table C.2 Panel A provides additional assessments of prediction accuracy by evaluating volume

forecasts in different size groups. We use the five groups from the JKP data sorted on the firms’

market capitalization.30

Table C.2: Prediction accuracy (R2 in %) in different size groups and “mixture of experts”

size group jointly nano micro small large mega

training obs 2,522,619 300,790 797,880 680,209 479,839 263,901
testing obs 1,893,067 273,792 467,413 552,503 384,819 214,540

A: pooled training evaluated in size groups and jointly (same models as in Table 1)
olsall 15.99 13.32 12.60 20.90 25.49 26.16
nnall 18.45 15.80 14.86 23.71 27.76 29.12
rnnall 19.86 16.63 16.14 26.00 30.50 32.02

B: size group training evaluated in size groups and jointly (mixture of experts)
ols+moeall 16.34 13.68 12.73 21.43 25.93 27.47
nn+moeall 17.78 15.29 14.43 22.69 26.57 27.71
rnn+moeall 18.26 15.24 14.71 24.76 29.02 30.99

Panel A evaluates the benchmark models (pooled training) in the five size groups, respectively, in
the OOS period. Each model uses “all” 175 predictors. Column “jointly” repeats Table 1, Panel
A, last column. Panel B trains “expert” models for each size group separately and evaluates them
in their corresponding size groups in the OOS period. Column “jointly” evaluates the mixture of
experts (moe) model, which predicts with the corresponding expert model trained on the same size
group for each OOS data point. Each R2 value is the average of five runs, same as Table 1.

The prediction accuracy increases as firm size increases, regardless of the prediction method.

The R2 ’s evaluated in the mega firms are roughly twice those of the nano firms. As explained in

the main text, smaller firms have a greater magnitude of unexpected trading volume shocks that

are hardest to predict. This result makes sense since small firms are volatile and have low trading
30
The five groups are defined according to the market capitalization breakpoints of NYSE stock percentiles: mega
stocks, greater than the 80th percentile; large, 50–80; small, 20–50; micro, 1-20; and nano, below the 1st percentile.

50
volume, hence unexpected events that give rise to volume spikes are more likely for these firms.

This finding also indicates that in addition to small firms being less liquid on average, their liquidity

is also less predictable and more volatile. Hence, tiny firms are not only costly to trade in general,

but their costs are less predictable. These results are intuitive and suggest that our prediction

models are capturing true variation in volume and not simply noise.

We also examine whether firms of different size groups should be modeled differently. We train a

model on each size group separately to attempt to better capture the heterogeneity across the firm

size dimension rather than pooling all firms in the same model. We implement a simple mixture

of experts (moe) method, where each size group is trained separately to form “expert” models and

then compared against the pooled training model in Panel A.

The moe improves performance of the ols method. Comparing the first lines in Panels A and

B, linear models catered to different size groups are more accurate than the pooled ols. For nn

and rnn, however, the sample size reduction outweighs the potential benefits of separate training,

making the mixture of expert models less accurate (either conditioning on size groups or jointly).

Since the potential non-linear effects of firm size are already allowed in the neural networks, forcing

size groups into different models is less effective. For this reason, we stick with pooled training

samples.

C.2 Additional results of implementing factor zoo portfolios

We provide results on the trading experiments that implement the factor zoo portfolios in addition

to Subsection 6.3.

Figure C.3 has the same vertical axis as Figure 7, which is the improvement in after-cost mean

excess return from ma5 to [Link] . The horizontal axis changes to the mean excess return

achieved with the ma5 method, i.e., the baseline level of the improvement. The plot shows the gain

is distributed around a positive center uncorrelated with the baseline. That is, a better volume

prediction is uniformly effective, and the improvement is not concentrated on factors that have

positive (or negative) realized returns.

Figure C.4 is the Sharpe ratio version of Figure 7 by showing the gain in Sharpe ratio instead

51
of the mean return. The plot shows a similar pattern to the one reported in the main text. The

gains in Sharpe ratios are larger for those factors with higher turnover, reaching around 0.3 to 0.4

per year.

Figure C.3: Mean return improvements in implementing each factor portfolio

A. implementing each factor portfolio B. averaged by theme clusters

ret_1_0

1.0
Mean return gain, [Link] vs. ma5 (%, annualized)

1.0 iskew_ff3_21d
zero_trades_21d Short-Term Reversal
rskew_21d
iskew_hxz4_21d
rmax5_rvol_21d
iskew_capm_21d
turnover_var_126d
dolvol_var_126d
lti_gr1a
0.8
seas_1_1an
0.8
inv_gr1a
ret_3_1
coskew_21d rmax1_21d ivol_ff3_21d
taccruals_ni ivol_hxz4_21d
seas_2_5an
rvol_21d
0.6 cash_at aliq_mat
niq_be_chg1
turnover_126d
zero_trades_126d
rmax5_21d
ivol_capm_21d 0.6
age niq_at_chg1taccruals_at
resff3_6_1
at_be
dsale_dsga
z_score oaccruals_at
ret_12_7
beta_dimson_21d
oaccruals_ni
op_at
Accruals Low Risk Profitability
bidaskhl_21d ni_ar1
seas_16_20an ope_bel1 mispricing_perf
op_atl1 ebit_sale qmj_prof eqnetis_at
seas_11_15an
netdebt_me
qmj_growth
dsale_drec qmj_safety
nfna_gr1abetadown_252d cop_at
prc_highprc_252d
ebit_bev
ivol_capm_252d
niq_at netis_at Profit Growth
Low Leverage Seasonality
seas_6_10an beta_60m
ocf_at_chg1
ocfq_saleq_std f_score
qmj resff3_12_1 cowc_gr1a
be_me niq_be
cop_atl1 zero_trades_252dni_be
ni_me
ocf_at
at_me Momentum
0.4 seas_16_20na
dgp_dsale
sale_emp_gr1
earnings_variability capx_gr2 fnl_gr1a
chcsho_12m
eqnpo_12m o_score
bev_mev
mispricing_mgmt
ope_be
dbnetis_at eqpo_me
eqnpo_me
ebitda_mev 0.4 Quality Value
dsale_dinv
debt_gr3 niq_su
opex_at
gp_atl1
saleq_gr1col_gr1acoa_gr1a
ret_12_1
eq_dur
sale_gr1 at_turnover debt_me fcf_me ocf_me Debt Issuance
seas_11_15na
noa_at
saleq_su
capex_abn ival_me tangibility sale_bev
kz_index
ppeinv_gr1a
emp_gr1
noa_gr1a
gp_at ret_6_1
be_gr1a
Investment
inv_gr1 dolvol_126d lnoa_gr1a
ni_ivol prc
pi_nix
capx_gr3
rd_sale
ami_126d market_equity
sale_gr3
seas_2_5na sti_gr1a
capx_gr1
ret_9_1 at_gr1
aliq_at
Size
0.2
seas_6_10na
rd5_at
ret_60_12
seas_1_1na
rd_me nncoa_gr1a
sale_me 0.2
betabab_1260d
ni_inc8q tax_gr1a ncoa_gr1a
corr_1260d

ncol_gr1a
0.0 div12m_me 0.0
15 10 5 0 5 10 15 15 10 5 0 5 10 15
Mean return implemented with ma5 (%, annualized) Mean return implemented with ma5 (%, annualized)
The same plot as Figure 7, but changing the x-axis to mean return achieved with the ma5 method,
i.e., the baseline of the gain.

Figure C.4: Sharpe ratio improvements in implementing each factor portfolio

A. implementing each factor portfolio B. averaged by theme clusters

iskew_ff3_21d
iskew_hxz4_21d

0.4 0.4
Sharpe ratio gain, [Link] vs. ma5 (annualized)

iskew_capm_21d

rskew_21d
Short-Term Reversal

0.3 0.3
inv_gr1a
nfna_gr1a seas_1_1an
rmax5_rvol_21d
taccruals_ni
coskew_21d
lti_gr1a dsale_drec ret_1_0
0.2 cowc_gr1a
niq_be_chg1
resff3_6_1 0.2
niq_su
niq_at_chg1
capx_gr2 fnl_gr1a
oaccruals_ni turnover_var_126d
dolvol_var_126d
taccruals_at seas_2_5an Accruals
cop_ataliq_mat
sale_bev
ocfq_saleq_std
z_score
kz_index
debt_gr3
cop_atl1earnings_variability
be_gr1a
lnoa_gr1a resff3_12_1
ret_3_1 Profit Growth
oaccruals_at
dsale_dsga
ni_ar1
capx_gr1
sale_emp_gr1
ppeinv_gr1a ocf_at_chg1 dbnetis_at
ret_12_7
Quality Debt Issuance Seasonality
0.1
op_at
op_atl1
opex_atdgp_dsale
qmj
qmj_growth
qmj_safety
at_turnover
sti_gr1a
f_score
coa_gr1a
mispricing_mgmt
col_gr1a
noa_gr1a
saleq_gr1
sale_gr1mispricing_perf
emp_gr1
zero_trades_21d
beta_dimson_21d 0.1 Investment Momentum
gp_atl1 qmj_prof
beta_60m
ope_bel1
tangibility
gp_at
at_be
ebit_bev
netdebt_me
noa_at
eq_dur
inv_gr1
capx_gr3
ocf_atival_me
ebit_sale at_gr1
be_me
saleq_su
netis_at
dsale_dinv
eqnetis_at
capex_abnniq_at
ivol_ff3_21d
ivol_hxz4_21d
ivol_capm_21d rmax5_21d
rvol_21d
rmax1_21d seas_16_20an
seas_11_15an
seas_6_10an Profitability
Low Leverage Low Risk
niq_be
age cash_atturnover_126d prc_highprc_252d ret_6_1
ret_12_1 ret_9_1
rd_sale o_score
ni_be
ope_beeqpo_me
eqnpo_me
rd_me
zero_trades_126d
bev_mev
ivol_capm_252d
ni_me chcsho_12m
debt_me
fcf_me
sale_gr3 nncoa_gr1a
eqnpo_12m
zero_trades_252d
ebitda_mev
dolvol_126d
ocf_me
at_me
seas_16_20na
betadown_252d
tax_gr1a seas_2_5na
seas_11_15na
Value
aliq_at
pi_nix
rd5_at
sale_me
market_equity
ncoa_gr1a
ni_inc8q seas_1_1na bidaskhl_21d Size
ami_126d
ni_ivolbetabab_1260d seas_6_10na
prc ret_60_12
0.0 div12m_me
corr_1260d
ncol_gr1a
0.0

0 1 2 3 4 5 6 0 1 2 3 4 5 6
Turnover (annualized) Turnover (annualized)
The same plot as Figure 7, but showing the gain in Sharpe ratio instead of mean return.

52