Deep Order Flow Imbalance: Extracting Alpha at Multiple

Horizons from the Limit Order Book

Petter Kolm
NYU Courant

[email protected]
linkedin.com/in/petterkolm

Joint work with Jeremy Turiel and Nicholas Westray

Northfield’s Newport Summer Research Seminar

June 14, 2023

1 / 46
Petter Kolm

Clinical Full Professor, NYU Courant

Director of the M.S. in Mathematics in Finance Program

Petter is the Director of the Mathematics in Finance Master’s program and a Clinical Professor of Math-
ematics at the Courant Institute of Mathematical Sciences, New York University. In this role he interacts
with major financial institutions such as investment banks, financial service providers, insurance companies
and hedge funds. Petter worked in the Quantitative Strategies group at Goldman Sachs Asset Manage-
ment developing proprietary investment strategies, portfolio and risk analytics in equities, fixed income
and commodities.
Petter was awarded “Quant of the Year” in 2021 by Portfolio Management Research (PMR) and Journal
of Portfolio Management (JPM) for his contributions to the field of quantitative portfolio theory. Petter
is a frequent speaker, panelist and moderator at academic and industry conferences and events. He
is a member of the editorial boards of the International Journal of Portfolio Analysis and Management
(IJPAM), Journal of Financial Data Science (JFDS), Journal of Investment Strategies (JoIS), and Journal
of Portfolio Management (JPM). Petter is an Advisory Board Member of Alternative Data Group (ADG),
AISignals and Operations in Trading (Aisot), Betterment (one of the largest robo-advisors) and Volatility
and Risk Institute at NYU Stern. He is also on the Board of Directors of the International Association for
Quantitative Finance (IAQF) and Scientific Advisory Board Member of the Artificial Intelligence Finance
Institute (AIFI).
Petter is the co-author of several well-known finance books including, Financial Modeling of the Equity
Market: From CAPM to Cointegration (Wiley, 2006); Trends in Quantitative Finance (CFA Research
Institute, 2006); Robust Portfolio Management and Optimization (Wiley, 2007); and Quantitative Equity
Investing: Techniques and Strategies (Wiley, 2010). Financial Modeling of the Equity Markets was among
the “Top 10 Technical Books” selected by Financial Engineering News in 2006.
As a consultant and expert witness, Petter has provided his services in areas including alternative data, data
science, econometrics, forecasting models, high frequency trading, machine learning, portfolio optimization
with transaction costs, quantitative and systematic trading, risk management, robo-advisory, smart beta
strategies, trading strategies, transaction costs, and tax-aware investing.
He holds a Ph.D. in Mathematics from Yale University; an M.Phil. in Applied Mathematics from the Royal
Institute of Technology, Stockholm, Sweden; and an M.S. in Mathematics from ETH Zurich, Switzerland.
2 / 46
The M.S. in Mathematics in Finance at NYU Courant

◮ Visit our website

◮ Follow us on LinkedIn
◮ Hire our students

3 / 46
Introduction & Motivation

4 / 46
Our Articles Related to This Talk

◮ Kolm, Turiel, and Westray (2021), “Deep Order Flow

Imbalance: Extracting Alpha at Multiple Horizons from the
Limit Order Book”
Available on SSRN:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3900141

◮ Kolm and Westray (2023), “A Bayesian Approach to Analyzing

Information Content of Cross-Sectional and Multilevel Order
Flow Imbalance” (In preparation) 5 / 46
Systematic Trading in U.S. Equities at High Frequency
◮ Generating alpha signals (return predictors) is critical in this
endeavour for both buy and sell sides
◮ Sell Side: Improving performance in next generation trading
algorithms
◮ Buy Side: Developing and deploying profitable HF strategies
◮ The holy grail for many of these firms is to take as input a raw
limit order book (or a collection of order books) and produce
high frequency price/return forecasts

6 / 46
Signal Generation
◮ This type of alpha generation is very challenging for a number
of reasons
◮ Enormous amounts of data (TB & PB sizes)
◮ Specialist infrastructure required to store, process and analyze
◮ Data is noisy, non-stationary and fat tailed
◮ Field is extremely competitive, every single one of your
competitors is trying to do the same thing with the same data
◮ Current state of the art is to employ quants to extract and
handcraft features using expert domain knowledge which then
become high value IP

7 / 46
The Bitter Lesson
The biggest lesson that can be read from 70 years of AI
research is that general methods that leverage computation
are ultimately the most effective, and by a large margin
The ultimate reason for this is Moore’s law, or rather its
generalization of continued exponentially falling cost per
unit of computation.
. . . Researchers seek to leverage their human knowledge
of the domain [to improve performance], but the only thing
that matters in the long run is the leveraging of computa-
tion. These two need not run counter to each other, but
in practice they tend to.
Rich Sutton - The Bitter Lesson, March 20191

8 / 46
Examples of the Bitter Lesson
◮ Chess: Methods that defeated Kasparov were based on
massive deep search, not the special structure of Chess
◮ Go: 20 years later, leveraging learning from self play (to
discover value functions) provided the edge as this allowed
massive computation to be brought to bear
◮ Computer Vision: Early methods searched for edges, cylinders
etc. This has all been discarded in favour of modern
Convolutional Neural Networks (CNNs)
◮ Natural Language Applications: Methods based on linguistics,
knowledge of words/phonemes are also no longer used.
Modern approaches are all deep learning/statistically based
(BERT, FinBERT)

9 / 46
Return Forecasting

10 / 46
Review: Limit Order Book (LOB)

11 / 46
General Problem Statement
◮ Focus on the top 10 (non-zero) levels of the LOB and define
the vector

xt := (at1 , vt1,a , bt1 , vt1,b , . . . , at10 , vt10,a , bt10 , vt10,b )⊤ ∈ R40 (1)

◮ For each stock and time t, fix a horizon h and consider the
time series regression problem

rt,t+h = g (xt , xt−1 , . . . , xt−W ) + εt (2)

◮ rt,t+h are the forward returns (hereafter rt )

◮ The function g is a neural network
◮ W denotes the length of the lookback window (typically 100)

12 / 46
Overview: Neural Networks
◮ The simplest Neural Network is best thought of as being made
of sequential applications (layers) of a linear transform and an
element wise non-linearity

Input Hidden Output

layer layer layer

Input #1

Input #2
Output
Input #3

Input #4

Figure 1: Simple Multilayer Perceptron.

13 / 46
Overview: Multilayer Perceptron (MLP)
◮ More generally, when we have L ≥ 1 hidden layers, each with
Nl ∈ N neurons. If x ≡ h0 ∈ RN0 , is the input then an MLP
with L hidden layers takes the form
! "
h(l) = f W (l) h(l−1) + b (l) , l = 1, 2, . . . , L ,
y = W (L+1) h(L) + b (L+1)

◮ The function f is called the activation (commonly chosen as

f (x) = max{x, 0}, aka ReLU)
◮ Can be easily trained using backpropagation (chain rule), and
there are software packages (Tensorflow, PyTorch) which do
this for you
◮ They are a universal approximator (Hornik, 1991)

14 / 46
Overview: Recurrent Neural Networks (RNNs)
◮ MLP is great for function approximation but has no notion of
order. Recurrent neural networks were designed for sequence
tasks (NLP, time series forecasting etc.)
◮ This time there is an input at each time xt together with a
hidden state ht

ht = fh (xt , ht−1 )
yt = fo (ht )
◮ The non-linear functions fh , fo handle the recurrent
transformation and the output transformation
◮ Trained via backpropagation through time (BPTT) where the
network is unrolled. Issues with vanishing/exploding gradients
◮ Important architecture avoiding this is the Long Short-Term
Memory Network (LSTM) (Hochreiter and Schmidhuber,
1997)
15 / 46
Overview: Convolutional Neural Networks (CNNs)
◮ Originally introduced in computer vision, specifically image
classification
◮ The layers apply (multiple) kernels to locally average parts of
the image, then apply a non-linearity

Figure 2: Example Convolution.

16 / 46
Financial CNNs
◮ We can think of an order book as a matrix, one state per row
◮ Temporal convolutions are interpreted as non-linear moving
averages
◮ Spatial convolutions aggregate information across the different
levels (and sides) of the LOB
◮ The most well known model (Zhang, Zohren, and Roberts,
2019) sequentially applies spatial and temporal convolutions
processing the order book into a series of features
◮ Convolutional layers are a preprocessing step, the resulting
time series is used as the input to an LSTM
◮ The CNN-LSTM can be trained via BPTT

17 / 46
Related Work
◮ Due to success of NNs in classification, existing literature
reformulates the regression problem as one of classification
◮ There are three main groups of authors who have focussed on
this problem, including
◮ Tsantekidis, Passalis, Tefas, Kanniainen, Gabbouj, and Iosifidis
(2020) - Focus on 5 Finnish stocks (FI-2010), investigate
different architectures
◮ Sirignano and Cont (2019) - Focus on S&P 500 with a single
architecture, investigate questions of universality
◮ Zhang, Zohren, and Roberts (2019) - Focus on 5 LSE stocks,
investigate sophisticated CNN-LSTM inception networks
◮ There are still a lot of open practical questions to investigate

18 / 46
Some Important Questions for the Practitioner
◮ Should we transform the raw LOB before inputting into the
NN?
◮ What architecture(s) are best?
◮ Is there a linkage between model predictive performance and
stock characteristics / microstructural properties?
◮ What kind of horizon do these alphas have?

19 / 46
What About Stationarity?
◮ Recall the original regression formulation

rt = g (xt , xt−1 , . . . , xt−W ) + εt (3)

◮ Statistical properties of model inputs and outputs
◮ Prices are not-stationary
◮ Volumes might be stationary - evidence mixed
◮ Returns generally considered stationary
◮ Therefore, we have stationary LHS and non-stationary RHS.
This seems odd
◮ Econometric theory would suggest to difference the input when
we are interested in performing inference

20 / 46
How to Difference a LOB?
◮ Given two snapshots of a limit order book, define the following
transform
# i,b
$
% vt , if bti > bt−1
i ,
bOFt,i := vti,b − vt−1i,b
, i i
if bt = bt−1 ,
$
& i,b
−vt−1 , if bti < bt−1
i
# i,a
$
%−vt−1 , if ati > at−1
i ,
aOFt,i := vti,a − vt−1
i,a
, i i
if at = at−1 ,
$
& i,a
vt , if ati < at−1
i

◮ If the bid price doesn’t move and bid level moves down, that’s
a negative imbalance (signal). Bid price moves up, that’s a
positive imbalance. Bid price moves down that’s a negative
imbalance
◮ Classical transform taken from the market microstructure
literature Cont, Kukanov, and Stoikov (2014)
21 / 46
Order Flows & Order Flow Imbalances
◮ We form order flow (OF) by concatenation and order flow
imbalance (OFI) by subtraction
' (
bOFt
OFt := ∈ R20 ,
aOFt
OFIt := bOFt − aOFt ∈ R10

◮ Both OF and OFI are stationary inputs. In addition, OFI is

well known to be an excellent predictor of contemporaneous
returns (all time scales)
◮ These are ideal candidates to explore as alternative RHS in the
regression

22 / 46
Our Setup & Contribution
◮ Recast the problem as standard regression and include multiple
horizons in the output

rt := (rt,1 , . . . , rt,H )⊤ ∈ RH , where H ≥ 1 (4)

◮ We refer to the forecasts as an alpha term structure at time t.

Our forecasting models take the form

rt = g (xt , xt−1 , . . . , xt−W ) + εt (5)

◮ We are going to
◮ Understand the effects of different RHS in the regression
◮ Compare multiple architectures across a large set of symbols
◮ Explore the relationship between stock characteristics and
model predictive power

23 / 46
Data & Model Fitting

24 / 46
Data Description
◮ We use data for the time period January 1, 2019 through
January 31, 2020 (LOBSTER, WRDS)
◮ Dataset is very large, approximately 10TB uncompressed. We
need specialist infrastructure to process and store
◮ Across our universe we have stocks with a few updates per day
(EBAY) and stocks with many updates per day (MSFT).
Information flows at different rates for different stocks and so
horizon cannot be constant in time across stocks

25 / 46
Defining Return Horizons
◮ We use a stock specific increment defined via

2.34 · 107
∆t := , (6)
N
where the numerator is the number of milliseconds in a trading
day and the denominator N denotes the average number of
non-zero tick by tick mid-price returns
◮ We then define horizons in terms of ∆t or “number of average
price changes.” We focus on a short horizons,
1
5 k∆t , k = 1, . . . , 10 i.e. between 0 and 2 average price
changes
◮ Insert a fixed latency buffer of 10ms for all intervals to mimic
production setting

26 / 46
Model Universe
◮ ARX - Autoregressive with exogenous features (linear model)

100
)
r t = w0 + vi⊤ xt−i + εt
i=1

◮ MLP - Multilayer Perceptron (4 layers) (Briola, Turiel, and

Aste, 2020)
◮ LSTM - Long Short-Term Memory Network (128 hidden units)
◮ LSTM-MLP - LSTM (128 hidden units) mapping to an MLP
(64 hidden units)
◮ LSTM (3) - Deep LSTM (150 hidden units) (Sirignano and
Cont, 2019)
◮ CNN-LSTM - State of the art model proposed by Zhang,
Zohren, and Roberts (2019)

27 / 46
Model Hyperparameters

Number of Number of Learning Batch Training Early

Model Input layers parameters(∗) rate size epochs stopping
ARX OF 1 2.0 × 103 10−4 256 50 Yes
CNN-LSTM OF 27 1.3 × 105 10−3 256 50 Yes
LSTM OF 2 1.0 × 105 10−5 256 50 Yes
LSTM (3) OF 4 4.6 × 105 10−5 256 50 Yes
LSTM-MLP OF 3 8.4 × 104 10−5 256 50 Yes
MLP OF 4 1.3 × 106 10−5 256 50 Yes
ARX LOB 1 4.0 × 103 10−4 256 50 Yes
CNN-LSTM LOB 27 1.4 × 105 10−3 256 50 Yes
LSTM LOB 2 1.1 × 105 10−5 256 50 Yes
LSTM (3) LOB 4 4.7 × 105 10−5 256 50 Yes
LSTM-MLP LOB 3 9.4 × 104 10−5 256 50 Yes
MLP LOB 4 2.3 × 106 10−5 256 50 Yes

Table 1: Summary of the inputs and hyperparameters used. (∗) The

number of parameters are approximated to the nearest order of
magnitude and truncated for readability.

28 / 46
Training
◮ To mimic a real life production setting we perform (for each
symbol) rolling fits over our time period
◮ We choose a (1,4,1) configuration
◮ The first week is for validation (early stopping)
◮ The middle 4 weeks are for training
◮ The final week is for out of sample testing
◮ We then step this forward by 3 weeks (to keep numbers of fits
manageable)
◮ Overall we have 12 x 115 x 18 (models x stocks x time
periods) NNs to fit
◮ We apply winsorization and Z -scoring to all independent and
dependent variables used in the regressions

29 / 46
Network Fitting
◮ Training deep learning models at this scale would likely be
impossible without the use of GPUs which offer ∼ 10×
speedup
◮ Our computations are performed on the NYU Greene2 and
Hudson3 high performance computing environments. Greene
has 32K CPU cores, 332 NVIDIA GPUs and 145TB of RAM
distributed over 568 nodes, Hudson has 960 CPU cores, 160
AMD GPUs and 10TB of RAM distributed across 20 nodes
◮ Code is written in Python and NNs are trained using
Tensorflow 2.3.1 and Keras (ADAM optimizer)
◮ For LSTM/CNN-LSTM-based models, we use a single GPU,
either NVIDIA Quadro RTX8000 (48GB), an NVIDIA V100
(32GB) or an AMD MI-50 (32GB).
◮ Time to train a single model varies in the range of 10-60
minutes, depending on model, stock and GPU
30 / 46
Results

31 / 46
Order Flow Imbalance or Limit Order Book?

1.0

1.25 0.5

1.00
0.0
ARX ARX
0.75 CNN-LSTM CNN-LSTM
(%)

(%)
°0.5
LSTM LSTM
ROS

ROS
0.50 LSTM (3) LSTM (3)
2

2
LSTM-MLP °1.0 LSTM-MLP
MLP MLP
0.25
°1.5
0.00
°2.0
°0.25

0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0

Fraction of an Average Price Change Fraction of an Average Price Change

Figure 3: Left Panel OF. Right Panel LOB.

32 / 46
Discussion
◮ Recall the key questions we set out to address
◮ Understand the effects of different RHS in the regression
◮ Compare multiple architectures across a large set of symbols
◮ OF input is clearly better than LOB - stationarity of inputs is
important
◮ LSTM based models outperform non-LSTM models
◮ Depth/CNN layers do not seem to outperform plain LSTM
after converting to OF
◮ Significant alpha at all horizons for the OF models. Small R 2
but high profitability due to the short horizons

33 / 46
Stock Characteristics
◮ We have seen that a regular LSTM model with OF input is an
excellent (non-complex) model
◮ Use the following stock characteristics to study model
performance
◮ Tick Size - Fraction of time the spread is equal to $0.01 (large
tick stocks approximately 1)
◮ Log Updates - Log Number of updates/day
◮ Log Trades - Log Number of trades/day
◮ Log Price Chg - Log Number of price changes/day
◮ All numbers computed per stock by averaging across the time
period

34 / 46
Characteristics Correlation

Tick Size 1 0.66 0.32 -0.47 0.95

1.0
LogUpdates 0.66 1 0.86 0.29 0.61

0.5
LogTrades 0.32 0.86 1 0.58 0.25
0.0

LogPriceChg -0.47 0.29 0.58 1 -0.59 °0.5

Log(Updates/PriceChg) 0.95 0.61 0.25 -0.59 1

Siz
e tes es
Ch
g
hg
)
ck da rad ice eC
Ti Up Lo
gT
gP
r
Pr
ic
og s/
L Lo
d ate
g(Up
Lo

Figure 4: Dependence Structure of Cross Sectional Characteristics.

35 / 46
Methodology
◮ We fix the model to be (LSTM, OF), average across horizons
and out of sample data points
◮ We are left with a single ROS
2 per stock (115 points)

◮ These are plotted against the stock characteristics in a scatter

◮ Results are almost identical for (CNN-LSTM, OF)

36 / 46
Cross Sectional Performance

(i) (ii)
4 4
(%)

(%)
ROS

ROS
2 2
2

2
0 0

0.2 0.4 0.6 0.8 105 106

Tick Size Updates

(iii) (iv)
4 4
(%)

(%)
ROS

ROS
2 2
2

2
0 0

104 103 104 105

Trades PriceChg

Figure 5: Cross Sectional Performance - (LSTM, OF) Model.

37 / 46
Explaining Performance

5 2
ROS : y = 0.0089log(x) - 0.0089

3
ROS
2 (%)

2 3 4 5
Log(Updates/PriceChg)

Figure 6: Cross Sectional Performance - Log(Updates/PriceChg).

38 / 46
Discussion
◮ There are clear dependencies on updates, tick size and trades
◮ Regression analysis (see article) shows that the best
characteristic is in fact a combination,
Log(Updates/PriceChg). This explains performance (ROS 2 )

with an adj.R 2 of 75%

◮ Why?
◮ When Log(Updates/PriceChg) is large, you have a stable order
book with lots of updates per price change. This is also a
property of large tick stocks (hence the correlation)
◮ You have lots of data and complicated patterns forming
between time series of imbalances and price changes
◮ The LSTM model is able to capture and model this
◮ We have good out-of-sample performance across the vast
proportion of our universe

39 / 46
Additional Results & Robustness Checks
Many additional questions addressed in the article
◮ How far ahead can we predict returns? (about 2-3 price
changes)
◮ How sensitive are the results to the fixed window length
W = 100? (interestingly not very)
◮ What if we use an OFI or volume-only LOB RHS? (removing
prices helps but OF is the best)
◮ Motivation for results in terms of inductive biases - an
interesting new concept from the ML literature

40 / 46
Conclusion & Extensions

41 / 46
Conclusion
◮ We built a framework to evaluate different inputs and deep
learning models for return predictions
◮ We demonstrated that stationarity of the inputs is critical to
getting good outcomes
◮ Our results suggests that universality needs a second
consideration as model results have strong microstructural
dependencies
◮ Simple architectures perform as well as complicated ones

42 / 46
Extensions
◮ Better understand relationship between classification and
regression
◮ How about Bayesian deep nets or other “fancy” models?
◮ Can we use this setup to perform volume prediction?
◮ Can we extend what we have to other venues/asset classes
(futures)?

43 / 46
Contact
Petter Kolm
NYU Courant

[email protected]
https://www.linkedin.com/in/
petterkolm

44 / 46
References I
Briola, Antonio, Jeremy Turiel, and Tomaso Aste (2020). Deep Learning
Modeling Of Limit Order Book: A Comparative Perspective. arXiv:
2007.07319 [q-fin.TR].
Cont, Rama, Arseniy Kukanov, and Sasha Stoikov (2014). “The Price Impact
Of Order Book Events”. In: Journal Of Financial Econometrics 12.1,
pp. 47–88.
Hochreiter, Sepp and Jürgen Schmidhuber (1997). “Long Short-Term Memory”.
In: Neural Computation 9.8, pp. 1735–1780.
Hornik, Kurt (1991). “Approximation Capabilities Of Multilayer Feedforward
Networks”. In: Neural Networks 4.2, pp. 251–257.
Kolm, Petter N., Jeremy Turiel, and Nicholas Westray (2021). “Deep Order
Flow Imbalance: Extracting Alpha from the Limit Order Book”. In: Working
Paper.
Kolm, Petter N. and Nicholas Westray (2023). “A Bayesian Approach to
Analyzing Information Content of Cross-Sectional and Multilevel Order Flow
Imbalance”. In preparation.

45 / 46
References II
Sirignano, Justin A. and Rama Cont (2019). “Universal Features Of Price
Formation In Financial Markets: Perspectives From Deep Learning”. In:
Quantitative Finance 19.9, pp. 1449–1459.
Tsantekidis, Avraam, Nikolaos Passalis, Anastasios Tefas, Juho Kanniainen,
Moncef Gabbouj, and Alexandros Iosifidis (2020). “Using Deep Learning For
Price Prediction By Exploiting Stationary Limit Order Book Features”. In:
Applied Soft Computing 93, p. 106401.
Zhang, Zihao, Stefan Zohren, and Stephen Roberts (2019). “DeepLOB: Deep
Convolutional Neural Networks for Limit Order Books”. In: IEEE
Transactions On Signal Processing 67.11, pp. 3001–3012.

46 / 46