---
title: 'Market Split Problem: Models & Applications'
url: https://www.emergentmind.com/topics/market-split-problem
type: topic
---

# Market Split Problem: Models & Applications

Searching arXiv for recent and foundational papers on “market split problem” and closely related formulations.
The expression **Market Split Problem** is not unitary in the technical literature. It denotes several distinct but structurally related problems in optimization, economics, market microstructure, electricity-market design, and platform systems. In its most classical operations-research sense, it is a binary feasibility or deviation-minimization problem of the form \(Ax=d\) or \(\min \sum_i |s_i|\) subject to \(\sum_j a_{ij}x_j+s_i=d_i\) with \(x_j\in\{0,1\}\) [2507.05045]. In game-theoretic work, it denotes the endogenous partition of consumers across competing producers in a two-stage noncooperative game [1403.7536]. In financial microstructure, it denotes the splitting of a large hidden parent trade into aggressive child orders and the reconstruction of that parent order from execution data, particularly in options markets [1902.05418]. In electricity and platform markets, it denotes the division of a previously integrated market into separate price areas or platform silos because of congestion or institutional boundaries [1809.10554; 2204.03616]. In experimental design for two-sided marketplaces, it denotes a deliberate splitting of budgets and traffic into independent sub-markets to remove interference between treatment and control [2012.08724]. The common element is a constrained partition of demand, supply, exposure, or feasible matches.

## 1. Scope and principal usages

The literature uses the term across several domains with different state variables, equilibrium concepts, and objectives. The table summarizes the main usages represented in the cited work.

| Usage | Core object | Representative paper |
|---|---|---|
| Classical MSP | Binary feasibility or deviation minimization \(Ax=d\) | [2507.05045] |
| Market allocation game | Quantile-like partition of consumers across producers | [1403.7536] |
| Options execution | Reconstruction and impact of split parent orders | [1902.05418] |
| Zonal electricity market | Congestion-induced price-area separation | [1809.10554] |
| Shared mobility segmentation | Platform-induced restriction of feasible rider-driver matches | [2204.03616] |
| Marketplace experimentation | Budget and traffic split into two independent sub-markets | [2012.08724] |

This plurality matters because the phrase can otherwise be misleadingly treated as if it referred to a single canonical model. In fact, some papers study **optimization under exact balance constraints**, others study **equilibrium market allocation**, and still others study **institutional or physical segmentation**. A useful way to read the term is as an umbrella for problems in which an integrated market, order, or matching pool is partitioned, either endogenously or by design.

## 2. Classical binary optimization problem

In the formulation emphasized by Cornuéjols and Dawande and revisited in recent algorithmic work, the Market Split Problem is the optimization problem
\[
\begin{aligned}
\min \quad & \sum_{i=1}^m |s_i| \\
\text{s.t.}\quad & \sum_{j=1}^n a_{ij} x_j + s_i = d_i, \qquad i=1,\dots,m,\\
& x_j \in \{0,1\}, \qquad j=1,\dots,n,\\
& s_i \in \mathbb{Z}, \qquad i=1,\dots,m.
\end{aligned}
\]
Its **feasibility version** asks whether there exists \(x\in\{0,1\}^n\) such that
\[
\sum_{j=1}^n a_{ij}x_j=d_i,\qquad i=1,\dots,m,
\]
or, in matrix form,
\[
Ax=d,\qquad x\in\{0,1\}^n.
\]
The feasibility problem is thus a multidimensional subset-sum problem [2507.05045].

A standard surrogate reduction maps the \(m\)-dimensional system to a one-dimensional subset-sum instance. If \(D>a_{ij}\) for all \(i,j\), then
\[
\sum_{j=1}^n \left( \sum_{i=1}^m (nD)^{i-1} a_{ij} \right) x_j
=
\sum_{i=1}^m (nD)^{i-1} d_i
\]
is equivalent to the original system. This mixed-radix encoding is classical, but recent work stresses that it is not necessary to solve the full problem: one can instead exploit the multidimensional structure directly [2507.05045].

The reason MSP became a hard benchmark is not merely NP-completeness. The cited algorithmic paper states that such instances have **a vast number of basic LP solutions with objective value \(0\)** and **very little pruning during tree exploration**, which makes LP-based branch-and-cut and branch-and-bound perform poorly [2507.05045]. The standard hard family is defined by
\[
n=10(m-1), \qquad d_i=\left\lfloor \frac12 \sum_{j=1}^n a_{ij}\right\rfloor,
\]
with \(a_{ij}\) uniform in \([0,99]\) in the original benchmark construction.

A recent exact method derives from Schroeppel–Shamir’s subset-sum algorithm but uses it only to enumerate solutions of a single constraint. Variables are partitioned into four blocks, heaps generate sorted partial sums lazily, and the remaining \(m-1\) constraints are validated in parallel on the GPU. The resulting hybrid CPU-GPU implementation is reported to solve instances with up to **10 constraints and 90 variables**, including \((9,80)\) in **less than fifteen minutes** and \((10,90)\) in **up to one day** [2507.05045]. This places the classical MSP squarely in exact exponential-time combinatorial optimization rather than in market design in the economic sense.

## 3. Endogenous market allocation in a two-stage game

A distinct usage appears in the noncooperative model of service providers and consumers in which a market split can look collusive while arising from decentralized optimization. There are finitely many producers
\[
P=\{0,1,\dots,n-1\},
\]
a continuum of consumers with types \(d\in T=[0,1]\), and a finite measure \(\mu\) over \(T\), where a lower \(t\) means better quality of service. Consumer type \(d\) accepts only producers with \(t_j\le d\). In the second stage, consumers choose among acceptable producers with minimal load; in the first stage, producers choose QoS levels \(t_j\in[0,1]\) [1403.7536].

A central technical result is that although the stage-2 consumer game may have many Nash equilibria, the induced producer loads are unique across them. This makes the producer game well defined. Under the economically relevant **fine preferences**, each producer first maximizes load and then, conditional on load, prefers larger \(t_j\), that is, worse and therefore cheaper QoS. The paper’s main theorem states that a unique pure-strategy Nash equilibrium exists, up to permutations. When the equilibrium QoS vector is sorted,
\[
t_0\le t_1\le \cdots \le t_{n-1},
\]
it satisfies
\[
t_j=\Max\left\{t\in T \mid \mu([0,t))\le \frac{j}{n}\mu(T)\right\}\qquad \forall j\in P,
\]
and each producer’s equilibrium load is
\[
\ell_j=\frac{\mu(T)}{n}.
\]
Moreover, the equilibrium is **super-strong** [1403.7536].

The market-split theorem sharpens this into an allocation statement. If \(\mu\) is atomless and \(\bar t\) is a producer Nash equilibrium, then in any mixed-consumption Nash equilibrium, producer \(j\) serves almost all consumers in the quantile slice
\[
\mu([0,d))\in \left(\frac{j}{n}\mu(T),\frac{j+1}{n}\mu(T)\right).
\]
Hence each producer serves one contiguous block of consumer types ordered by minimum acceptable quality, and each producer chooses precisely the **worst quality still acceptable** to its own customer slice [1403.7536].

This result is important because it directly addresses an apparent misconception: a split market need not indicate coordination or signaling. The paper proves that the outcome is the unique and highly robust consequence of noncooperative behavior, and it further shows that sequential \(\delta\)-better-response dynamics converge and that best-response dynamics do so very quickly [1403.7536]. The “split” here is therefore an equilibrium partition of the consumer space.

## 4. Order splitting and market impact in options markets

In financial microstructure, the market split problem is the execution problem created by limited displayed liquidity. A large desired trade is fragmented into many smaller executions over time; the hidden parent order is called a **metaorder**. In equities, same-day sequential orders with the same agent, product id, and direction are a natural definition. In options, the object of interest is usually not a single listed contract but exposure to a parameter of the implied-volatility surface, so the definition must be adapted [1902.05418].

For a volatility-surface parameter \(\theta\), an options metaorder is defined as a series of same-day sequential orders with the same agent, the same underlying product id, and the same direction in the sign of sensitivity to \(\theta\). The directional state variable is
\[
\mathcal{S}^{\theta}:=Q\times \frac{\partial \mathcal O}{\partial \theta},
\]
where \(Q\) is the algebraic quantity, positive for a buy and negative for a sell, and \(\mathcal O\) is the option price. The relevant parent intentions are therefore long or short **ATM forward volatility** or **ATM forward skew**, not merely long or short one option. Detection must be done in **greek / volatility-parameter space** rather than by ticker and buy/sell flag alone [1902.05418].

The empirical study uses a proprietary BNP Paribas trading-desk dataset on **KOSPI 200 options** over a **2-year period**, from **June/July 2016 through June 2018**, and focuses on **aggressive limit orders**. Because parameter recalibration must occur at a cadence comparable to trade arrivals, the analysis concentrates mainly on short-maturity options. To keep only meaningful split executions, the paper requires at least **5 executed transactions** per metaorder. The reported sample sizes are **1,026,197 orders** and **149,441 metaorders** for ATMF volatility, and **1,304,714 orders** and **174,091 metaorders** for ATMF skew [1902.05418].

The core empirical message is that the reconstructed hidden parent orders in options behave much like the metaorders documented in equities. Their impact is **increasing and concave during execution**, **partially decays afterward**, **satisfies an approximate fair-pricing relation**, and **scales roughly with the square root of participation** [1902.05418]. The study also emphasizes a striking timescale difference: whereas the earlier equity study cited by the authors had metaorders lasting hours, the options metaorders here often last only **a few tenths of seconds**. In this usage, the market split problem is fundamentally about execution fragmentation and impact measurement.

## 5. Congestion, network constraints, and segmented matching markets

In zonal day-ahead electricity markets, market split refers to **price-area separation** generated by transmission constraints. The market is divided into geographic zones, and the operator clears all zones jointly subject to inter-zonal transfer capacities. If transmission is not congested, connected zones trade freely and prices remain equal. If a line reaches its transfer limit, the market splits into separate price areas, with the importing zone tending to have a higher market-clearing price and the exporting zone a lower price [1809.10554].

The cited Turkish day-ahead market model is a **Mathematical Program with Complementarity Constraints (MPCC)** classified as an **MIQCQP**. Market splitting is encoded through the interaction of zonal balance constraints, line-capacity bounds, complementarity conditions, and price-flow equations. For \(t=1,\dots,23\),
\[
\overline{\mu}_{l,t}-\underline{\mu}_{l,t}
+\overline{\rho}_{l,t}-\underline{\rho}_{l,t}
-\overline{\rho}_{l,t+1}+\underline{\rho}_{l,t+1}
=
p_{e_l,t}-p_{s_l,t},
\]
with a corresponding terminal-period condition at \(t=24\). If no capacity or ramping constraint binds, the multipliers vanish and
\[
p_{e_l,t}=p_{s_l,t},
\]
so there is **no market split**. If a transmission bound binds, the relevant multiplier becomes positive and creates a price difference across the line [1809.10554].

An institutional rather than physical segmentation appears in shared mobility. There the split market problem is the inefficiency created when riders and drivers are partitioned across platforms that do not coordinate. The paper formalizes market structure as a restriction on the feasible matching graph:
\[
G_{\mathcal Q^\mu}=(V_{\mathcal Q^\mu},E_{\mathcal Q^\mu}),
\qquad
E_{\mathcal Q^\mu}\subseteq E_{\mathcal Q},
\qquad
V_{\mathcal Q^\mu}\subseteq V_{\mathcal Q}.
\]
Segmentation partitions riders and drivers by platform, so some cross-platform edges become institutionally infeasible [2204.03616].

This graph-theoretic restriction reduces matching thickness, pooling opportunities, and vehicle utilization. The paper studies four market structures—**Bilateral Trading Market**, **Central Trading Market**, **Cooperative Market**, and **Shared Mobility Marketplace**—and evaluates them with a simulator using Manhattan ride-hailing demand data from **7–10 a.m.**, **Wednesday, October 2, 2019**. The headline result is that the proposed market designs can reduce total VMT by **6%** while serving more customers with **8.4% fewer total number of trips**, and customers receive on-average **5.4% shorter waiting time** [2204.03616]. A plausible implication is that, in both electricity and shared mobility, market splitting can be read as the removal of feasible exchange edges, whether by network congestion or by platform boundaries.

## 6. Experimental, sequential, and adjacent extensions

A further usage appears in causal inference for two-sided online marketplaces. Standard Bernoulli A/B testing is biased because treatment and control contaminate each other through competition for shared budgets and seller attention. The proposed **budget-split** design resolves this by splitting each buyer’s budget proportionally across two simultaneously running sub-markets and splitting the seller-side traffic into two buckets. If members are partitioned into sizes \(N^{(0)}\) and \(N^{(1)}\), then each campaign \(j\) is duplicated with budgets
\[
B_j^{(0)}=\frac{N^{(0)}}{N}B_j,
\qquad
B_j^{(1)}=\frac{N^{(1)}}{N}B_j,
\]
and the two copies target only their assigned buckets. Under **Limited interference** and **Stable system**, and when
\[
N^{(0)}=N^{(1)}=N/2,\qquad p=0.5,
\]
the estimator is unbiased for the full-market total effect [2012.08724]. The paper states that naive marketplace experiments can yield **one to two times overestimation of the treatment effect**, and reports that for the effect size at which budget-split reaches **80% power**, campaign-level experiments have **5.2%** and **12%** power in two marketplaces, while switchback has **5.1%** and **5.2%** power [2012.08724]. Here the market split is deliberately engineered to create two independent sub-markets.

Several adjacent literatures use closely related partitioning ideas. In market definition, Yan Yang argues that the relevant task is to partition products into one or more **relevant product markets** directly from quantified **substitutability**, using a multi-dimensional feature space and clustering, chiefly \(k\)-means / \(k\)-means++ [1906.10030]. In the **global robust newsvendor** problem, the split is sequential rather than simultaneous: production serves a primary market first, leftover inventory
\[
x=(q_1-D_1)^+
\]
can be transshipped to a later secondary market, and the make-versus-transfer rule is governed by the threshold
\[
c_2\frac{\theta}{\theta+\Delta\theta}\ \lessgtr\ C+v_1
\]
[2207.03801]. In conceptual work on the **separation of market and price**, the split is functional: one may have a discovery mechanism without a positive price or a price without an effective discovery facility, with the job application market and the real-estate-agent market as the canonical examples [2106.05972].

More distant but still informative uses of “split” appear in stochastic portfolio theory and multiscale control. A market model with **competing Brownian particles** allows firms to split when their market weight reaches a threshold and to merge when an exponential clock rings; the paper proves nonexplosiveness and diversity while finding no relative arbitrage over finite horizons [1404.0748]. In degradation-aware battery market participation, the key contribution is an offline/online **market/degradation split** that separates long-term aging from short-term market operation via a health-indexed value function approximation [2603.21089]. These usages are not the classical MSP, but they reinforce a general theme: splitting is a device for representing constrained interaction across heterogeneous timescales, customer groups, resources, or feasible matches.

Across these literatures, the Market Split Problem is best understood not as one theorem or one optimization model, but as a family of formal problems about how integrated markets become partitioned and what that partition does to feasibility, equilibrium, prices, impact, experimentation, or welfare.

Source: https://www.emergentmind.com/topics/market-split-problem