[CACM version]

D OI:10.1145/ 1735223 . 1 73 5 2 47
Censored Exploration
and the Dark Pool Problem
By Kuzman Ganchev, Yuriy Nevmyvaka, Michael Kearns, and Jennifer Wortman Vaughan
Dark pools are a recent type of stock exchange in which
information about outstanding orders is deliberately hidden in order to minimize the market impact of large-volume trades. The success and proliferation of dark pools
have created challenging and interesting problems in algorithmic trading—in particular, the problem of optimizing
the allocation of a large trade over multiple competing
dark pools. In this work, we formalize this optimization as
a problem of multi-venue exploration from censored data,
and provide a provably efficient and near-optimal algorithm for its solution. Our algorithm and its analysis have
much in common with well-studied algorithms for managing the exploration–exploitation trade-off in reinforcement learning. We also provide an extensive experimental
evaluation of our algorithm using dark pool execution data
from a large brokerage.
Dark pools are a relatively new type of exchange designed to
address the problems that arise from the transparent (or
“light”) nature of a typical stock exchange—namely, the difficulty of minimizing the impact of large-volume trades.3, 5, 7 In
a typical exchange, the revelation that there is a large-volume
buyer (seller) in the market can cause prices to rise (fall) at
the buyer’s (seller’s) expense. If the volume is sufficiently
large, and the trading period sufficiently short, such market
impacts remain even if one attempts to fragment the trade
over time into smaller transactions. As a result, there has
been increasing interest in recent years in execution mechanisms that allow full or partial concealment of large trades.
In a typical dark pool, buyers and sellers submit orders
that simply specify the total volume of shares they wish to
buy or sell, with the price of the transaction determined
exogenously by “the market”.a Upon submitting an order
to buy (or sell) v shares, a trader is put in a queue of buyers
(or sellers) awaiting transaction. Matching between buyers and sellers occurs in sequential arrival of orders, similar to a light exchange. However, unlike a light exchange,
no information is provided to traders about how many
parties or shares might be available in the pool at any
given moment. Thus in a given time period, a submission
of v shares results only in a report of how many shares up
to v were executed.
While presenting their own trading challenges, dark pools
have become tremendously popular exchanges, responsible
For our purposes, we can think of the price as the midpoint between the bids
and ask in the light exchanges, though this is a slight oversimplification.
for executing 10–20% of the overall US equity volume. In fact,
they have been so successful that there are now approximately 40+ dark pools for the US Equity market alone. The
popularity of these exchanges has left large-volume traders
and brokerages facing a novel problem: How should one
optimally distribute a large trade over the many independent dark pools?
To answer this question, we analyze a framework and
algorithm for a more general multi-venue exploration problem. We consider a setting in which at each time period,
we have some exogenously determined volume of V units of
an abstract good (for example, shares of a stock that a client would like to sell). Our goal is to “sell” or “consume”
as many of these units as possible at each step, and there
are K abstract “venues” (for example, various dark pools) in
which this selling or consumption may occur. We can divide
our V units into any way we like across the venues in service
of this goal. What differentiates this problem from most
standard learning settings is that if vi units are allocated to
venue i, and all of them are consumed, we learn only that
the total demand at venue i was at least vi, not the precise
number of units that could have been consumed there. This
important aspect of our framework is known as censoring in
the statistics literature.
In this work, we make the natural and common assumption that the maximum amount of consumption available
in venue i at each time step (or the total liquidity available,
in the dark pool problem) is drawn according to a fixed but
unknown distribution Pi. Formally speaking, this means
that when vi units are submitted to venue i, a value si is drawn
randomly from Pi and the observed (and possibly censored)
amount of consumption is min{si, vi}.
A learning algorithm in our framework receives a
sequence of volumes V 1, V 2, … and must decide how to distribute the V t units across the venues at each time step t. Our
goal is to efficiently (in time polynomial in the parameters of
the model) learn a near-optimal allocation policy. There is a
distinct between-venue exploration component to this problem, since the best number of shares to submit to venue i
may depend on both Vt and the distributions for the other
venues, and the only mechanism by which we can discover
the distributions is by submitting allocations. If we routinely
submit too-small volumes to a venue, we receive censored
observations and are underutilizing the venue; if we submit
The original version of this paper was published in
the Proceedings of the 25th Conference on Uncertainty in
Artificial Intelligence, 2009.
MAY 2 0 1 0 | VO L . 5 3 | N O. 5 | C OM M U N IC AT ION S OF T HE ACM
research highlights
too-large volumes, we receive uncensored observations but
have excess inventory.
Our main theoretical contribution is a provably polynomial-time algorithm for learning a near-optimal policy for any
unknown venue distributions Pi. This algorithm takes a particularly natural and appealing form, in which allocation and
distribution reestimation are repeatedly alternated. More precisely, at each time step we maintain estimates of the distributions Pi; pretending that these estimates are in fact exactly
correct, we allocate the current volume V accordingly. These
allocations generate observed consumptions in each venue,
which in turn are used to update the estimates. We show
that when the estimates are “optimistic tail modifications”
of the classical Kaplan–Meier maximum likelihood estimator for censored data, this estimate–allocate loop has provably efficient between-venue exploration behavior that yields
the desired result. Venues with smaller available volumes are
gradually given smaller allocations in the estimate–allocate
loop, whereas venues with repeated censored observations
are gradually given larger allocations, eventually settling on a
near-optimal overall allocation distribution.
Finally, we present an extensive experimental evaluation
of our model and algorithm on the dark pool problem, using
trading data from a large brokerage.
The closest problem to our setting is the widely studied newsvendor problem from the operations research
literature. In this problem, at each time period a player
(representing a newsstand owner) chooses a quantity V
of newspapers to purchase at a fixed per-unit price, and
tries to optimize profit in the face of demand uncertainty
at a single venue (their newsstand).b Huh et al.10 were the
first to consider the use of the Kaplan–Meier estimator in
this class of problems. They use an estimate–allocate loop
similar to ours, and show asymptotic convergence to nearoptimal behavior in a single venue. Managing the distribution of an exogenously specified volume V across multiple
venues (which are the important aspects of the dark pool
problem, where the volume to be traded is specified by a
client, and there are many dark pools) and the attendant
exploration–exploitation trade-off between venues are key
aspects and differentiators of our algorithm and analysis. We also obtain stronger (polynomial time rather than
asymptotic) bounds, which require a modification of the
classical Kaplan–Meier estimator.
Formally, we consider the following problem. At each time
t, a learner is presented with a quantity or volume V t Î {1,
…, V} of units, where V t is sampled from an unknown distribution Q. The learner must decide on an allocation v®t of
these shares to a set of K known venues, with vit Î {0, …, V t}
for each i Î{1, …, K}, and åki = 1 vit = V t. The learner is then
told the number of units rit consumed at each venue i. Here
rit = min{sit, vit}, where sit is the maximum consumption
level of venue i at time t, which is sampled independently
In our setting, it is important that we view V as given exogenously by the
client and not under the trader’s control, which distinguishes our setting
somewhat from the prior works.
COM MUNICATIO NS O F TH E ACM | M AY 2 0 10 | VOL. 53 | NO. 5
from a fixed but unknown distribution Pi. If rit = vit, we say
that the algorithm receives a censored observation because
it is possible to infer only that rit £ sit. If rit < vit, we say that
the algorithm receives a direct observation because it must
be the case that rit = sit.
The goal of the learner is to discover a near-optimal
one-step allocation policy, that is, an allocation policy that
approximately optimizes the expected number of units out
of Vt consumed at each time step t. (We briefly discuss other
objectives at the end of Section 4.4.)
Throughout the remainder of the paper, we use the shorthand Ti for the tail probabilities associated with Pi. That
is, Ti(s) = ås¢ ³ s Pi(s¢). c Clearly Ti(0) = 1 for all i. We use T̂it(s) for
an empirical estimate of Ti(s) at time t.
Before tackling the full exploration–exploitation problem,
we must examine a more basic question: Given estimates T̂i
of the tail probabilities Ti for each venue i, how can we maximize the (estimated) expected number of units consumed
on a single time step? It turns out that this can be accomplished using a simple greedy allocation scheme. The greedy
algorithm allocates one unit at a time. The venue to which
the next unit is allocated is chosen to maximize the estimated probability that the unit will be consumed. It is easy
to see that if vi units have already been allocated to venue i,
then the estimated probability that the next allocated unit
will be consumed is simply T̂i (vi + 1). A formal description of
the Greedy algorithm is given in Figure 1.
Theorem 1. The allocation returned by Greedy maximizes
the expected number of units consumed in a single time step,
where the expectation is taken with respect to the estimated tail
probabilities {T̂i }iK = 1.
The proof of this theorem is fairly simple. Using the fact
that tail probabilities must satisfy T̂i (s) ³ T̂i (s¢) for all s £ s¢, it
is easy to verify that by greedily adding units to the venues in
decreasing order of T̂i (s), the algorithm returns
The remainder of the proof involves showing that the
expression being maximized here equivalent to the
expected number of units consumed. This can be done
We now present our main theoretical result, which is a
In the early literature on censored estimation, these tail probabilities were
referred to as survival probabilities, as T(s) usually represented the probability
that a patient in a particular medical study survived for at least s years past the
start of the study. In this setting, observations were frequently censored when
researchers lost track of a patient midway through the study and knew only
that the patient lived at least until the point at which contact was broken.1
The curious reader can find more details of this and other omitted proofs
in the original version of this paper.9
Figure 1. Optimal allocation algorithm Greedy.
Input: Volume V, tail probability estimates {T̂i }iK=1
Output: An allocation v
v ¬ 0;
for ! ¬ 1 to V do
i ¬ argmaxi T̂i (vi + 1);
vi ¬ vi + 1;
return v
polynomial-time, near-optimal algorithm for multi-venue
exploration from censored data. The analysis of our algorithm bears strong resemblance to the exploration–exploitation arguments common in the E3 and RMAX family of
algorithms for reinforcement learning.4, 12 In particular,
there is an analogy to the notion of a known state inherent in
those earlier algorithms, along with an exploitation lemma
(proving that expected payoffs from known states are high)
and an exploration lemma (proving that extended periods of
low payoffs must result in more states becoming known).
In our setting, however, the number of states is exponential
and thus the special structure of our problem is required
to obtain a polynomial time algorithm. We first provide an
overview of the algorithm and its analysis before examining
it in more detail.
At the highest level, the algorithm is quite simple and
natural. It maintains estimates T̂it for the true unknown
tail probabilities Ti for each venue i. These estimates
improve with time in a particular quantifiable sense which
drives between-venue exploration. At any given time t, the
current volume Vt is allocated across the venues by simply
calling the optimal greedy allocation scheme from Figure
1 on the current set of estimated tail probabilities T̂it. This
results in new censored observations from each venue,
which in turn are used to update the estimates T̂it + 1 used at
the next time step. Thus the algorithm, which is formally
stated in Figure 2, implements a continuous allocate–
reestimate loop.
Note that we have not yet described the algorithm’s
subroutine OptimisticKM, which specifies how we estimate T̂it from the observed data. The most natural choice
would be the maximum likelihood estimator on the data.
This estimator is well-known in the statistics literature
as the Kaplan–Meier estimator. In the following section,
we describe Kaplan–Meier and derive a new convergence
result that suits our particular needs. This result in turn
lets us define an optimistic tail modification of Kaplan–
Meier that becomes our choice for OptimisticKM. Figure 3
shows the full subroutine.
The analysis of our algorithm, which is developed in more
detail over the next few sections, proceeds as follows:
Step 1: We first review the Kaplan–Meier maximum likelihood estimator for censored data and provide a new
finite sample convergence bound for this estimator. This
bound allows us to define a cut-off for each venue i such
that the Kaplan–Meier estimate of the tail probability Ti(s)
for every value of s up to the cut-off is guaranteed to be
close to the true tail probability. We then define a lightly
modified version of the Kaplan–Meier estimates in which
the tail probability of the next unit above the cut-off is
modified in an optimistic manner. We show that in conjunction with the greedy allocation scheme, this minor
modification leads to increased exploration, since the next
unit beyond the cut-off always looks at least as good as the
cut-off itself.
Step 2: We next prove our main Exploitation Lemma (Lemma
3). This lemma shows that at any time step, if it is the case
that the number of units allocated to each venue by the
greedy algorithm is strictly below the cut-off for that venue
(which can be thought of as being in a known state in the
parlance of reinforcement learning) then the allocation is
provably e -optimal.
Step 3: We then prove our main Exploration Lemma (Lemma
4), which shows that on any time step at which the allocation
made by the greedy algorithm is not e-optimal, it is possible
to lower bound the probability that the algorithm explores.
Thus, any time we cannot ensure a near-optimal allocation,
we are instead assured of exploring.
Step 4: Finally, we show that on any sufficiently long
sequence of time steps (where sufficiently long is polynomial in the parameters of the model), it must be the case
that either the algorithm has already implemented a nearFigure 2. Main algorithm.
Input: Volume sequence V 1, V 2, V 3,...
Arbitrarily initialize T̂ i for each i;
for t ¬ 1, 2, 3, ... do
% Al locat ion Step:
vt ¬ Greedy (V t, T̂1,...,T̂K);
for i ¬ {1,...,K} do
Submit Vit units to venue i;
Let rit be the number of shares sold;
% Rees t imat ion Step:
T̂i ¬ OptimisticKM ({(v it, rit )}tt = 1);
Figure 3. Subroutine OptimisticKM. Let Mi,s'
and Ni,s'
be defined in
Section 4.1, and assume that e, d > 0 are fixed parameters.
Input: Observed data ({(v it, rit )}tt = 1) for venue i
Output: Modified Kaplan–Meier estimators for i
% Calculate the cut -of f :
ci t¬ max{s : s = 0 or Ni,s–1 ³ 128 (sV/e )2 ln(2V/d)};
% Compute Kaplan–Meier tai l probabi l i t ies :
T̂i (0) = 1;
for s = 1 to V do
T̂i (s) ¬ Õs¢ = 0 (1–(Mi,s¢t /Ni,s¢t ));
% Make the opt imist ic modi f icat ion:
if cit < V then
T̂i (cit + 1) ¬ T̂i (cit );
return T̂i ;
MAY 2 0 1 0 | VO L . 5 3 | N O. 5 | C OM M U N IC AT ION S OF T H E ACM
research highlights
optimal solution at almost every time step (and thus will
continue to perform well in the future), or the algorithm has
explored sufficiently often to learn accurate estimates of
the tail distributions out to V units on every venue. In either
case, we can show that with high probability, at the end of
the sequence, the current algorithm achieves an e -optimal
solution at each time step with probability at least 1 – e.
4.1. Convergence of Kaplan–Meier estimators
We begin by describing the standard Kaplan–Meier maximum likelihood estimator for censored data,11, 13 restricting
our attention to a single venue i. Let zi,s be the true probability that the demand in this venue is exactly s units given that
the demand is at least s units. Formally,
of the tail probability for s shares rapidly improves.
To prove this theorem, we must first show that the estit
converge to the true probabilities zi,s. In an i.i.d.
mates ẑ i,s
setting, this could be accomplished easily using standard
concentration results such as Hoeffding’s inequality.
In our setting, we instead appeal to Azuma’s inequality
(see, for example, Alon and Spencer2), a tool for bounding martingales, or sequences X1, X2, … such that for each
n, |Xn – Xn+1| £ 1 and E [Xn+1|Xn] = Xn. In particular, we show
(z i,s - ẑi,ts) can be expressed as the final term
that the value Ni,s
of a martingale sequence, allowing us to bound its absolute value. This in turn implies that bound on |z i,s - ẑi,ts| that
we need, and all that remains is to show that these bounds
imply a bound on the discrepancy between Ti(s) and the
estimator T̂i(s).
with T̂it(0) = Ti (0) = 1 for all t.
Previous work has established convergence rates for the
Kaplan–Meier estimator to the true underlying distribution in
the case that each submission in the sequence v1i ,...,vti is independently and identically distributed (i.i.d.),8 and asymptotic
convergence for non-i.i.d. settings.10 We are not in the i.i.d.
case, since the submitted volumes at one venue are a function
of the entire history of allocations and executions across all
venues. In the following theorem, we give a new finite sample
convergence bound applicable to our setting.
4.2. Modifying Kaplan–Meier
In Figure 3, we describe the minor modification of Kaplan–
Meier necessary for our analysis. As described above
(Step 1), the value cit in this algorithm can intuitively be
viewed as a cut-off up to which we are guaranteed to have
sufficient data to accurately estimate the tail probabilities
using Kaplan–Meier; this is formalized in Lemma 1. Thus
for every quantity s < cit, we simply let T̂it(s) be precisely the
Kaplan–Meier estimate as in Equation 1.
However, to promote exploration, we set the value
of T̂it(cit + 1) optimistically to the Kaplan–Meier estimate of
the tail probability at cit (not at cit + 1). This optimistic modification is necessary to ensure that the greedy algorithm
explores (i.e., has a chance of making progress towards
increasing at least one cut-off value) on every time step for
which it is not already producing an e -optimal allocation.
In particular, suppose that the current greedy solution
allocated no more than cit units to any venue i and exactly
cjt units to some venue j. Using the standard Kaplan–Meier
tail probability estimates, it could be the case that this
allocation is suboptimal (there is no way to know if it
would have been better to include unit cit + 1 from venue j
in place of a unit from another venue since we do not have
an accurate estimate of the tail probability for this unit),
and yet no exploration is taking place. By optimistically
modifying the tail probability T̂it(cjt + 1) for each venue, we
ensure that no venue remains unexplored simply because
the algorithm unluckily observes a low demand a small
number of times.
We now formalize the idea of cit as a cut-off up to which
the Kaplan–Meier estimates are accurate. In the results that
follow, we think of e > 0 and d > 0 as fixed parameters of the
Theorem 2. Let T̂it be the Kaplan–Meier estimate of Ti as given
in Equation 1. For any d > 0, with probability at least 1 – d, for
every s Î {1, …, V},
Lemma 1. For any s £ V, let T̂it(s) be the Kaplan–Meier estimator
for Ti(s) returned by OptimisticKM. With probability at least
1 – d, for all s £ cit, |Ti(s) - T̂it(s)| £ e /(8V).
It is easy to verify that for any s > 0,
At a high level, we can think of Kaplan–Meier as first computing a separate estimate of zi,s for each s and then using
these estimates to compute an estimate of Ti (s).
be the number of direct obserMore specifically, let Mi,s
vations of s units up to time t, that is, the number of time
steps at which strictly more than s units were allocated to
venue i and exactly s were consumed. Let Ni,st be the number
of either direct or censored observations of at least s units
on time steps at which strictly more than s units were allocated to venue i. We can then naturally define our estimate
= Mi,s
/Ni,st , with ẑ i,s
= 0 if Ni,st = 0. The Kaplan–Meier estimaẑ i,s
tor of the tail probability for any s > 0 after t time steps can
then be expressed as
T̂it(s) = Õ (1 - ẑ i,s¢
s¢= 0
Proof. It is always the case that Ti(0) = T̂it(0) = 1, so the result
This result shows that as we make more and more direct
or censored observations of at least s – 1 units on time steps
at which at least s units are allocated to venue i, our estimate
CO MM UNICATIO NS O F T H E ACM | M AY 2 0 10 | VOL. 53 | NO. 5
In particular, e corresponds to the value e specified in Theorem 3, and d
corresponds roughly to that d divided by the polynomial upper bound on
time steps.
holds trivially unless cit > 0. Suppose this is the case. Recall that
Ni,st is the number of direct or censored observations of at least
s units on time steps at which strictly more than s units were
allocated to venue i. By definition, it must be the case that
Ni,st ³ Ni,st , whenever s £ s¢. Thus by definition of the cut-off cit in
Figure 3, for all s < cit, Ni,st v128(sV /e)2 ln(2V/e ). The lemma then
follows immediately from an application of Theorem 2. !
Lemma 2 shows that it is also possible to achieve additive
bounds on the error of tail probability estimates for quantities s much bigger than cit as long as the estimated tail probability at cit is sufficiently small. Intuitively, this is because
the tail probability at these large values of s must be smaller
than the true tail probability at cit, which, in this case, is
known to be very small already.
Lemma 2. If T̂it (cit) £ e /(4V) and the high probability
event in Lemma 1 holds, then for all s such that cit < s £ V,
|Ti (s) - T̂it (s)| £ e /(2V).
4.3. Exploitation and exploration lemmas
We are now ready to state our main Exploitation Lemma
(Step 2), which formalizes the idea that once a sufficient
amount of exploration has occurred, the allocation output by the greedy algorithm is e -optimal. The proof of
this lemma is where the optimistic tail modification to
the Kaplan–Meier estimator becomes important. In particular, because of the optimistic setting of T̂it(cit + 1), we
know that if the greedy policy allocates exactly cit units to
a venue i, it could not gain too much by reallocating additional units from another venue to venue i instead. In this
sense, we create a buffer above each cut-off, guaranteeing
that it is not necessary to continue exploring as long as one
of the two conditions in the lemma statement is met for
each venue.
The second condition in the lemma may appear mysterious at first. To see why it is necessary, notice that the rate at
which the estimate T̂it(cit + 1) converges to the true tail probability Ti(cit + 1) implied by Theorem 2 depends on the number
of times that we observe a consumption of cit or more units. If
Ti(cit) is very small, then the consumption of this many units
does not frequently occur. Luckily, if this is the case, then
we know that Ti(cit + 1) must be very small as well, and more
exploration of this venue is not needed.
Lemma 3 (Exploitation Lemma). Assume that at time t, the
high probability event in Lemma 1 holds. If for each venue i,
either (1), vit £ cit or (2), T̂it(cit) £ e/(4V), the difference between the
expected number of units consumed under allocation v t and the
expected number of units consumed under the optimal allocation is at most e.
Proof Sketch. The proof begins by creating an arbitrary
one-to-one mapping between the units allocated to different
venues by the algorithm and an optimal allocation. Consider
any such pair in this mapping.
If the first condition in the lemma holds for the venue
i to which the unit was allocated by the algorithm, we can
use Lemma 1 to show that the algorithm’s estimate of the
probability of this unit being consumed is close to the true
probability; in particular, the algorithm is not overestimating
this probability too much. If the second condition holds, then
the algorithm’s estimate of the probability of the share being
consumed is so small that, again, the algorithm cannot possibly be overestimating it too much (because the lowest the
probability could be is zero). This follows from Lemma 2.
Now consider the venue j to which unit was allocated by
the optimal allocation. If the number of units vjt allocated to
this venue by the algorithm is strictly less than the cut-off cjt,
then by Lemma 1, the algorithm could not have underestimated the probability of additional units being consumed
by too much. Furthermore, because of the optimistic tail
modification of the Kaplan–Meier estimator, this also holds
if vjt = cjt. Finally, if it is instead the case that the second condition in the lemma statement holds for venue j, then the
algorithm again could not possibly have underestimated the
probability of the unit being consumed too much because
the true probability is so low.
Putting these pieces together, we can argue that for each
pair in the matching (of which there are no more than V),
since the algorithm did not overestimate the probability of
unit it chose being consumed by too much (in this case, too
much means more than e /(2V) ) and did not underestimate
the probability of the corresponding unit in the optimal
allocation by too much (again, by e /(2V) ), the difference in
expected units consumed between the optimal allocation
and the algorithm’s is at most e . !
Finally, Lemma 4 presents the main exploration lemma
(Step 3), which states that on any time step at which the allocation is not e -optimal, the probability of obtaining a useful
observation is at least e /(8V).
Lemma 4 (Exploration Lemma). Assume that at time t, the
high probability event in Lemma 1 holds. If the allocation is
not e -optimal, then for some venue i, with probability at least
e /(8V),
Proof. Suppose the allocation is not e -optimal at time
t. By Lemma 3, it must be the case that there exists some
venue i for which vit > cit and T̂it(cit) > e /(4V), i.e., a venue in
which the algorithm has allocated units past the cut-off
but for which the tail probability at the cut-off is not too
close to zero. Let be a venue for which this is true. Since
v!t > c!t, it will be the case that the algorithm obtains a useful
observation for exploration of this venue (i.e., an observat
tion causing N!,c
t to be incremented) if the number of units
consumed at this venue is sufficiently high (specifically, if
r!t > c!t). Since T̂ !t (c!t) > e /(4V), Lemma 1 implies that T!(c!t) >
e /(8V), which in turn implies that the number of units
consumed is high enough to constitute a useful observation with probability at least e /(8V). !
4.4. Putting it all together
With the exploitation and exploration lemmas in place, we
are finally ready to state our main theorem.
Theorem 3 (Main Theorem). For any e > 0 and d > 0, with
probability 1 − d (over the randomness of draws from Q and
{Pi}), after running for a time polynomial in K, V, 1/ e , and
ln(1/d ), the algorithm in Figure 2 makes an e-optimal allocation
MAY 2 0 1 0 | VO L . 5 3 | N O. 5 | C OM M U N IC AT ION S OF T H E ACM
research highlights
on each subsequent time step with probability at least 1 – e .
Proof Sketch. Suppose that the algorithm runs for R time
steps, where R is a (specific, but unspecified for now) polynomial in the model parameters K, V, 1/e , and ln(1/d ). If it
is the case that the algorithm was already e -optimal on a
fraction (1 – e ) of the R time steps, then we can argue that
the algorithm will continue to be e -optimal on at least a fraction (1 – e ) of future time steps since the algorithm’s performance should improve on average over time as estimates
become more accurate.
On the other hand, if the algorithm chose sub-optimal
allocations on at least a fraction e of the R time steps, then
by Lemma 4, the algorithm must have incremented Ni,ct it for
some venue i and cut-off cit approximately e 2R/(8V) times. By
definition of the cit, it can never be the case that Ni,ct it was incremented too many times for any fixed values of i and cit (where
too many is a polynomial in V, 1/e , and ln(1/d )); otherwise the
cut-off would have increased. Since there are only K venues
and V possible cut-off values to consider in each venue, the
total number of increments can be no more than KV times
this polynomial, another polynomial in V, 1/e , ln(1/d ), and
now K. If R is sufficiently large (but still polynomial in all of
the desired quantities) and approximately e 2 R/(8V) increments were made, we can argue that every venue must have
been fully explored, in which case, again, future allocations
will be e -optimal. !
We remark that our optimistic tail modifications of the
Kaplan–Meier estimators are relatively mild. This leads us to
believe that using the same estimate–allocate loop with an
unmodified Kaplan–Meier estimator would frequently work
well in practice. We investigate a parametric version of this
learning algorithm in the experiments described below.
The remainder of this article is devoted to the application
of our techniques to the dark pool problem. We begin with
a description of the trading data we used, and go on to
describe a variety of experiments we performed.
5.1. Summary of the dark pool data
Our data set is from the internal dark pool order flow for a
major US broker–dealer. Each (possibly censored) observation is of the form discussed throughout the paper—a triple
consisting of the dark pool name, the number of shares
sent to that pool, and the number of shares subsequently
executed within a short time interval. It is important to highlight some limitations of the data. First, note that the data set
conflates the policy the brokerage used for allocation across
the dark pools with the liquidity available in the pools themselves. For our data set, the policy in force was very similar
to the bandit-style approach we discuss below. Second, the
“parent” orders determining the overall volumes to be allocated across the pools were determined by the brokerage’s
trading needs, and are similarly out of our control.
The data set contains submissions and executions
for four active dark pools: BIDS Trading, Automated
Trading Desk, D.E. Shaw, and NYFIX, each for a dozen of
relatively actively-traded stocks,f thus yielding 48 distinct
CO MM UNICATIO NS O F T H E AC M | M AY 2 0 10 | VOL. 53 | NO. 5
stock–pool data sets. The average daily trading volume
of these stocks across all exchanges (light and dark)
ranges from 1 to 60 million shares, with a median volume of 15 million shares. Energy, Financials, Consumer,
Industrials, and Utilities industries are represented. Our
data set spans 30 trading days. For every stock–pool pair
we have on average 1,200 orders (from 600 to 2,000), which
corresponds to 1.3 million shares (from 0.5 to 3 million).
Individual order sizes range from 100 to 50,000 shares,
with 1,000 shares being the median. Sixteen percent of
orders are filled at least partially (meaning that fully 84%
result in no shares executed), 9% of the total submitted
volume was executed, and 11% of all observations were
5.2. Parametric models for dark pools
The theory and algorithm we have developed for censored
exploration permit a very general form for the venue distributions Pi. The downside of this generality is that we are
left with the problem of learning a very large number of
parameters. More parameters generally mean that more
data is necessary to guarantee that the model will generalize
well, which means more rounds of exploration are needed
before the algorithm’s future performance is near-optimal.
In some applications, it is therefore advantageous to employ
a less general but more simple parametric form for these
We experimented with a variety of common parametric
forms for the distributions. For each such form, the basic
methodology was the same. For each of the 4 × 12 = 48
venue–stock pairs, the data for that pair was split evenly
into a training set and a test set. The training data was
used to select the maximum likelihood model from the
parametric class. Note that we can no longer directly apply
the nonparametric Kaplan–Meier estimator—within each
model class, we must directly maximize the likelihood on
the censored training data. This is a relatively straightforward and efficient computation for each of the model
classes we investigated. The test set was then used to measure the generalization performance of each maximum
likelihood model.
Our investigations revealed that the best models maintained a separate parameter for the probability of zero
shares being available (that is, Pi(0) is explicitly estimated)—
a zero bin or ZB parameter. This is due to the fact that the
vast majority of submissions (84%) to dark pools result in
no shares being executed. We then examined various parametric forms for the nonzero portions of the venue distributions, including uniform (which of course requires
no additional parameters), and Poisson, exponential and
power law forms (each of which requires a single additional
parameter); each of these forms were applied up to the largest volume submitted in the data sets, then normalized.
The generalization results strongly favor the power law
form, in which the probability of s shares being available
is proportional to 1/sb for real b—a so-called heavy-tailed
Tickers represented are AIG, ALO, CMI, CVX, FRE, HAL, JPM, MER, MIR,
NOV, XOM, and NRG.
Train Loss
Test Loss
ZB + Uniform
ZB + Power Law
ZB + Poisson
ZB + Exponential
distribution when b > 0. Nonparametric models trained
with Kaplan–Meier are best on the training data but overfit badly due to their complexity relative to the sparse
data, while the other parametric forms cannot accommodate the heavy tails of the data. This is summarized
in Table 1. Based on this comparison, for our dark pool
study we investigate a variant of our main algorithm, in
which the estimate–allocate loop has an estimation step
using maximum likelihood estimation within the ZB +
Power Law model, and allocations are done greedily on
these same models.
In terms of the estimated ZB + Power Law parameters
themselves, we note that for all 48 stock–pool pairs the
Zero Bin parameter accounted for most of the distribution (between a fraction 0.67 and 0.96), which is not surprising considering the aforementioned preponderance
of entirely unfilled orders in the data. The vast majority
of the 48 exponents b fell between b = 0.25 and b = 1.3—so
rather long tails indeed—but it is noteworthy that for one
of the four dark pools, 7 of the 12 estimated exponents
were actually negative, yielding a model that predicts
higher probabilities for larger volumes. This is likely an
artifact of our size- and time-limited data set, but is not
entirely unrealistic and results in some interesting behavior in the simulations.
5.3. Data-based simulation results
As in any control problem, the dark pool data in our possession is unfortunately insufficient to evaluate and compare
different allocation algorithms. This is because of the aforementioned fact that the volumes submitted to each venue
were fixed by the specific policy that generated the data,
and we cannot explore alternative choices—if our algorithm chooses to submit 1000 shares to some venue, but in
the data only 500 shares were submitted, we simply cannot
infer the outcome of our desired submission.
We thus instead use the raw data to derive a simulator
with which we can evaluate different approaches. In light of
the modeling results of Section 5.2, the simulator for stock S
was constructed as follows. For each dark pool i, we used all
of the data for i and stock S to estimate the maximum likelihood Zero Bin + Power Law distribution. (Note that there is
no need for a training-test split here, as we have already separately validated the choice of distributional model.) This
results in a set of four venue distribution models Pi that form
Figure 4. Sample learning curves. For the stock AIG (left panel),
the naive bandits algorithm (labeled blue curve) beats uniform
allocation (dashed horizontal line) but appears to asymptote short
of ideal (solid horizontal line). For the stock NRG (right panel),
the bandits algorithm actually deteriorates with more episodes,
underperforming both the uniform and ideal allocations. For both
stocks (and the other 10 in our data set), our algorithm (labeled red
curve) performs nearly optimally.
Fraction executed
the simulator for stock S. This simulator accepts allocation
vectors (v1, v2, v3, v4) indicating how many shares some algorithm wishes to submit to each venue, draws a “true liquidity” value si from Pi for each i, and returns the vector (r1, r2, r3,
r4), where ri = min(vi, si) is the possibly censored number of
shares filled in venue i.
Across all 12 stocks, we compared the performance of
four different allocation algorithms. The (obviously unrealistic) ideal allocation is given the true parameters of the
ZB + Power Law distributions used by the simulator and
allocates shares optimally (greedily) with respect to these
distributions. The uniform allocation divides any order
equally among all four venues. Our learning algorithm
implements the repeated allocate–reestimate loop as in
Figure 2, using the maximum likelihood ZB + Power Law
model for the reestimation step. Finally, the simple (and
fairly naive) bandit-style algorithm maintains a weighting
over the venues and chooses allocations proportional to
the weights. It begins with equal weights assigned to all
venues, and each allocation to a venue which results in
any nonzero number of shares being executed causes that
venue’s weight to be multiplied by a constant factor a.
(Optimizing a over all stock–pool pairs resulted in a value
of a = 1.05.)
Some remarks on these algorithms are in order. First,
note that the ideal and uniform allocation methods are
nonadaptive and are meant to serve as baselines—one of
them the best performance we could hope for (ideal), and
the other the most naive allocation possible (uniform).
Second, note that our algorithm has a distinct advantage in
the sense that it is using the correct parametric form, the
same being used by the simulator itself. Thus our evaluation of this algorithm is certainly optimistic compared to
what should be expected in practice. Finally, note that the
bandit algorithm is the crudest type of weight-based allocation scheme of the type that abounds in the no-regret literature6; we are effectively forcing our problem into a 0/1 loss
setting corresponding to “no shares” and “some shares”
being executed. Certainly more sophisticated bandit-style
approaches can and should be examined.
Fraction executed
Table 1. Average per-sample log-loss (negative log likelihood) for
each venue distribution models. The “Wins” column shows the number of stock-venue pairs where a given model beats the other four on
the test data.
MAY 2 0 1 0 | VO L . 5 3 | N O. 5 | C OM M U N IC AT ION S OF T H E ACM
research highlights
COMM UNICATIO NS O F T H E AC M | M AY 2 0 10 | VOL. 53 | NO. 5
Figure 5. Comparison of our learning algorithm to the three
baselines. In each plot, the performance of the learning
algorithm is plotted on the y-axis, and the performance of one
of the baselines on the x-axis. Left column: Evaluated by the
fraction of submitted shares executed in a single time step;
higher values are better, and points above the diagonal are wins
for our algorithm. Right: Evaluated by order half-life; lower
values are better, and points below the diagonal are wins for
our algorithm. Each point corresponds to a single stock and
order size; small orders (red plus signs) are 1000 shares, large
orders (blue squares) are 8000 shares.
Fraction executed
Order half-life
Each algorithm was run in simulation for some number of episodes. Each episode consisted of the allocation
of a fixed number V of shares—thus the same number of
shares is repeatedly allocated by the algorithm, though
of course this allocation will change over time for the
two adaptive algorithms as they learn. Each episode of
simulation results in some fraction of the V shares being
executed. Two values of V were investigated—a smaller
value V = 1000, and the larger and potentially more difficult V = 8000.
We begin by showing full learning curves over 2000
episodes with V = 8000 for a couple of representative stocks
in Figure 4. Here the average performance of the two nonadaptive allocation schemes (ideal and uniform) are represented as horizontal lines, while learning curves are
given for the adaptive schemes. Due to high variance of the
heavy-tailed venue distributions used by the simulator, a
single trial of 2000 episodes is extremely noisy, so we both
average over 400 trials for each algorithm, and smooth the
resulting averaged learning curve with a standard exponential decay temporal moving average.
We see that our learning algorithm converges towards
the ideal allocation (as suggested by the theory), often
relatively quickly. Furthermore, in each case this ideal
asymptote is significantly better than the uniform allocation strawman, meaning that optimal allocations
are highly nonuniform. Learning curves for the bandit
approach exhibit one of the three general behaviors over
the set of 12 stocks. In some cases, the bandit approach
is quite competitive with our algorithm, though converging to ideal perhaps slightly slower (not shown in Figure
4). In other cases, the bandit approach learns to outperform uniform allocation but appears to asymptote short
of the ideal allocation. Finally, in some cases the bandit
approach appears to actually “learn the wrong thing”, with
performance decaying significantly with more episodes.
This happens when one venue has a very heavy tail, but
also a relatively high probability of executing zero shares,
and occurs because the very naive bandit approach that we
use does not have an explicit representation of the tails of
the distribution.
The left column of Figure 5 shows more systematic
head-to-head comparisons of our algorithm’s performance versus the other allocation techniques after 2000
episodes for both small and large V. The values plotted are
averages of the last 50 points on learning curves similar to
Figure 4. These scatterplots show that across all 12 stocks
and both settings of V, our algorithm competes well with
the optimal allocation, dramatically outperforms uniform, and significantly outperforms the naive bandit allocations (especially with V = 8000). The average completion
rate across all stocks for the large (small) order sequences
is 10.0% (13.1%) for uniform and 13.6% (19.4%) for optimal allocations. Our algorithm performs almost as well as
optimal—13.5% (18.7%)—and much better than bandits at
11.9% (17.2%).
In the right column, we measure performance not by
the fraction of V shares filled in one step, but by the natural alternative of order half-life—the number of steps of
repeated resubmission of any remaining shares to get the
total number executed above V/2. Despite the fact that
our algorithm is not designed to optimize this criterion
and that our theory does not directly apply to it, we see the
same broad story on this metric as well—our algorithm
competes with ideal, dominates uniform allocation and
beats the bandit approach on large orders. The average
order half-life for large (small) orders is 7.2 (5.3) for uniform allocation and 5.9 (4.4) for the greedy algorithm on
the true distributions. Our algorithm requires on average
6.0 (4.9) steps, while bandits uses 7.0 (4.4) to trade the large
(small) orders.
While there has been longstanding interest in quantitative finance in the use of models from machine learning and related fields, they are often applied towards the
attempt to predict directional price movements, or in the
parlance of the field, to “generate alpha” (outperform the
market). Here we have instead focused on a problem in
what is often called algorithmic trading—where one seeks
to optimize properties of a specified trade, rather than
decide what to trade in the first place—in the recently
introduced dark pool mechanism. In part because of the
constraints imposed by the mechanism and the structure
of the problem, we have been able to adapt and blend
methods from statistics and reinforcement learning
in the development of a simple, efficient, and provably
effective algorithm. We expect there will be many more
applications of machine learning methods in algorithmic
trading in the future.
We are grateful to Curtis Pfeiffer and Andrew Westhead
for valuable conversations and to Bobby Kleinberg
for introducing us to the literature on the newsvendor
1. Akritas, M.G. Nonparametric survival
analysis. Stat. Sci. 19, 4 (2004), 615–623.
2. Alon, N., Spencer, J. The Probabilistic
Method, 2nd Edition. Wiley, New York,
3. Bogoslaw, D. Big traders dive into dark
pools. Business Week article, available
at: http://www.businessweek.com/
394204.htm, 2007.
4. Brafman, R., Tennenholtz, M.
R-MAX—a general polynomial
time algorithm for near-optimal
reinforcement learning. J. Mach.
Learn. Res. 3 (2003), 213–231.
5. Carrie, C. Illuminating the new dark
influence on trading and U.S. market
structure. J. Trading 3, 1 (2008), 40–55.
6. Cesa-Bianchi, N., Lugosi, G. Prediction,
Learning, and Games. Cambridge
University Press, 2006.
7. Domowitz, I., Finkelshteyn, I.,
Yegerman, H. Cul de sacs and
highways: an optical tour of dark pool
trading performance. J. Trading 4, 1
(2009), 16–22.
8. Foldes, A., Rejto, L. Strong uniform
consistency for nonparametric
survival curve estimators from
randomly censored data. Ann. Stat. 9,
1 (1981), 122–129.
9. Ganchev, K., Kearns, M. Nevmyvaka, Y.,
Vaughan, J.W. Censored exploration and
the dark pool problem. In Proceedings
of the 25th Conference on Uncertainty
in Artificial Intelligence, 2009.
10. Huh, W.T., Levi, R., Rusmevichientong, P.,
Orlin, J. Adaptive data-driven inventory
control policies based on Kaplan–Meier
estimator. Preprint available at http://
psfiles/km-myopic.pdf, 2009.
11. Kaplan, E.L., Meier, P. Nonparametric
estimation from incomplete
observations. J. Am. Stat. Assoc. 53
(1958), 457–481.
12. Kearns, M., Singh, S. Near-optimal
reinforcement learning in polynomial
time. Mach. Learn. 49 (2002), 209–232.
13. Peterson, A.V. Kaplan-Meier estimator.
In Encyclopedia of Statistical
Sciences. Wiley, 1983.
Kuzman Ganchev (kuzman@cis.upenn.
edu), University of Pennsylvania.
Michael Kearns (mkearns@cis.upenn.
edu), University of Pennsylvania.
Yuriy Nevmyvaka (yuriy@cs.cmu.edu),
University of Pennsylvania.
Jennifer Wortman Vaughan (jenn@seas.
harvard.edu), Harvard University.
© 2010 ACM 0001-0782/10/0500 $10.00
ACM’s interactions
Magazine Seeks Its Next
For more information, see
To apply, please send
a résumé, letter of motivation,
and statement of your vision
for the magazine to:
Dan Olsen,
Search Committee Chair
As ACM’s premier magazine on applied computer-human interaction
(CHI), interactions is designed to keep developers, designers, managers,
researchers, and users abreast of the latest tools and ideas emerging from the CHI community—and beyond. This colorful, bi-monthly
magazine shares timely articles, stories, and practical content related
to the interactions between experiences, people, and technology.
Its primary objective is to trace innovative technologies from their
R&D beginnings to real-world applications and future potential.
interactions is also the membership magazine for ACM’s SIGCHI and
as such is distributed to more than 7,000 members worldwide.
The role of co-editor-in-chief is a three-year volunteer position and is
assisted by ACM professional staff for the production of the magazine.
The position starts in the third quarter of 2010 in preparation for the
January-February 2011 issue.
MAY 2 0 1 0 | VO L . 5 3 | N O. 5 | C OM M U N IC AT ION S OF T H E ACM