MaxBalanceCore: Heuristic for LSCBM Detection

• MaxBalanceCore is a heuristic method that detects the largest strong-correlation balanced module (LSCBM) in statistically validated stock networks, addressing the subjectivity of threshold-based methods.
• It converts validated Pearson correlations into a ternary signed adjacency matrix and uses seed selection, pruning, and greedy expansion to enforce structural balance.
• Empirical applications on Chinese stock data reveal that the detected core size varies with market conditions, offering insights into market cohesion during crises and stability periods.

Searching arXiv for the exact paper and closely related “balanced core” usages to ground the article. MaxBalanceCore is the heuristic algorithm introduced to detect the largest strong-correlation balanced module (LSCBM) in a statistically validated signed stock network. In that setting, the pipeline begins with daily stock prices, converts them to daily log-returns, retains only statistically significant Pearson correlations, and then searches the resulting sparse signed network for the maximum-size subset that is simultaneously pairwise strongly connected and structurally balanced. The exact identifier “MaxBalanceCore” belongs to this stock-network setting; elsewhere, the literature contains closely related but non-identical balanced-core constructions in payment systems, blockchain resource allocation, payment-channel rebalancing, and cooperative game theory (Qing et al., 7 Aug 2025).

1. Formal object and network construction

MaxBalanceCore operates on a statistically validated correlation network derived from stock returns. Given daily prices $P_i(\tau)$, returns are defined as

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$

where $\triangle\tau$ is one day. Pairwise dependence is measured by the Pearson correlation coefficient $\mathbf{C}_{i,j}$, and statistical validation is performed by testing

$$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$

with

$t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$

Under $H_0$, this follows a Student $t$-distribution with $\nu=T-2$ degrees of freedom, and significance is declared when

$|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$

The statistically validated matrix is then

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$0

with $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$1 for convenience (Qing et al., 7 Aug 2025).

Within this network, a subnetwork $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$2 is a strong-correlation balanced module (SCBM) if it satisfies two conditions. The strong-correlation condition requires

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$3

for any pair $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$4. The structural balance condition requires

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$5

for every three distinct nodes $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$6. Operationally, this allows exactly two balanced triangle types: all three edges positive, or one positive and two negative. The optimization target is the largest strong-correlation balanced module

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$7

The paper treats this construction as a response to two limitations of traditional threshold-based stock networks. First, the threshold $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$8 in

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$9

is subjective. Second, binarization removes both correlation magnitude and correlation sign, even though negative dependence may encode hedging or diversification relations. By contrast, $\triangle\tau$0 preserves sign and validated strength, while the later threshold $\triangle\tau$1 is used only for defining “strong-correlation” modules rather than for constructing a crude binary graph (Qing et al., 7 Aug 2025).

A key equivalent characterization of structural balance is also used. In the random signed-graph analysis, an SCBM can be written as a partition $\triangle\tau$2 such that all edges within $\triangle\tau$3 are positive, all edges within $\triangle\tau$4 are positive, and all edges between $\triangle\tau$5 and $\triangle\tau$6 are negative. This two-faction form is the combinatorial basis of MaxBalanceCore’s search logic.

2. Algorithmic structure of MaxBalanceCore

MaxBalanceCore is the algorithmic procedure for approximating $\triangle\tau$7. The paper is explicit that exact LSCBM detection is NP-hard, so MaxBalanceCore is a heuristic, not an exact solver (Qing et al., 7 Aug 2025).

The algorithm first converts the validated weighted network into a ternary signed adjacency matrix: $\triangle\tau$8 This enforces the strong-correlation requirement and retains only edge sign for the balance check. Although the input network is weighted, MaxBalanceCore itself operates on $\triangle\tau$9.

For each node $\mathbf{C}_{i,j}$0, the algorithm computes an impact score

$\mathbf{C}_{i,j}$1

Nodes are sorted in descending order of impact, and only the top $\mathbf{C}_{i,j}$2 seeds are explored. This top-100 cap is a substantive heuristic restriction: the search assumes that large balanced modules are likely to be centered on high-degree seeds.

For a chosen seed, its strong signed neighborhood is

$\mathbf{C}_{i,j}$3

The neighborhood is then split into two provisional factions: $\mathbf{C}_{i,j}$4

$\mathbf{C}_{i,j}$5

This directly encodes the two-faction structural-balance characterization.

The algorithm next prunes these factions to enforce sign consistency. Inside each group, any node $\mathbf{C}_{i,j}$6 is removed if there exists $\mathbf{C}_{i,j}$7 in the same group with

$\mathbf{C}_{i,j}$8

Thus every surviving intra-faction edge must be positive. Across the two factions, nodes are removed whenever a cross edge satisfies

$\mathbf{C}_{i,j}$9

so every surviving inter-faction edge must be exactly $$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$0. After pruning, the current candidate module is

$$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$1

A greedy expansion step then tries to grow the seed-centered balanced core. For a node outside the module, the intended requirement is that it have strong signed edges to all current module members. The pseudocode writes

$$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$2

which is slightly inconsistent with $$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$3; the intended logic is that all such edges are nonzero. A node can join faction $$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$4 if

$$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$5

and can join faction $$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$6 if

$$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$7

If canJoinA holds, the node is added to $$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$8; otherwise, if canJoinB holds, it is added to $$H_{0}: \mathbf{C}_{i,j} = 0 \quad \text{vs.} \quad H_{1}: \mathbf{C}_{i,j} \neq 0$$9. The algorithm tracks the largest module found across all processed seeds and returns that incumbent.

The implicit invariant after pruning and successful growth is exact: all pairs inside $t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$0 are $t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$1, all pairs inside $t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$2 are $t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$3, and all pairs across $t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$4 are $t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$5. This is why the paper interprets MaxBalanceCore as searching for a balanced module rather than for a generic dense or community-like subgraph.

The paper also states what MaxBalanceCore is not. It is not the name of the optimization problem, which is the LSCBM problem; and it is not a literal $t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$6-core or degeneracy-based core decomposition method. The “core” in the name refers to a balanced module, not to a graph-theoretic core decomposition (Qing et al., 7 Aug 2025).

3. Theoretical setting and asymptotic results

The theoretical analysis abstracts the validated stock network into a random signed graph $t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$7 with adjacency matrix $t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$8, where for each pair $t_{i,j} = \mathbf{C}_{i,j} \sqrt{\frac{T - 2}{1 - \mathbf{C}_{i,j}^2}.$9,

• $H_0$0 with probability $H_0$1,
• $H_0$2 with probability $H_0$3,
• $H_0$4 with probability $H_0$5.

In this setting, an SCBM is a subset in which every pair has nonzero edge and every triangle has positive sign product. The optimization statement remains

$H_0$6

The paper does not provide a mixed-integer exact formulation; the analysis is probabilistic and asymptotic rather than exact-combinatorial (Qing et al., 7 Aug 2025).

The first result is asymptotic non-emptiness. If $H_0$7, $H_0$8, and $H_0$9, then

$t$0

Thus balanced strong modules are not exceptional in large random signed networks.

For fixed $t$1, the main size law is

$t$2

where

$t$3

The theory also proves multiplicity: $t$4 where $t$5 is the number of SCBMs of size $t$6. The expected number of SCBMs of size $t$7 is

$t$8

and the associated rate function is

$t$9

Two special regimes are highlighted. In the dense positive-dominated regime

$\nu=T-2$0

the paper proves

$\nu=T-2$1

and states that with high probability the LSCBM is an all-positive module. In the negative-dominated regime

$\nu=T-2$2

the paper gives

$\nu=T-2$3

If additionally

$\nu=T-2$4

then multiplicity still holds: $\nu=T-2$5

These results do not analyze MaxBalanceCore’s approximation ratio. Their role is different: they justify why searching for large balanced modules is meaningful, why the two-faction decomposition is structurally appropriate, and why positive-dominated empirical markets should tend to exhibit large mostly all-positive modules (Qing et al., 7 Aug 2025).

4. Empirical behavior in Chinese stock networks

The empirical application uses Chinese A-share data from 2013 to 2024. For each year $\nu=T-2$6, the paper constructs $\nu=T-2$7, runs MaxBalanceCore, and studies the resulting $\nu=T-2$8. It summarizes module scale by

$\nu=T-2$9

the proportion of the full market contained in the detected LSCBM (Qing et al., 7 Aug 2025).

The reported sizes vary sharply across years. Examples include:

• 2013: $|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$0, $|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$1, $|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$2;
• 2014: $|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$3;
• 2015: $|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$4, $|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$5;
• 2016: $|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$6, $|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$7;
• 2017: $|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$8;
• 2021: $|t_{i,j}|> t_{\nu}\!\left(\frac{\alpha}{2}\right).$9, $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$00;
• 2024: $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$01, $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$02.

The interpretation offered is regime-dependent. During market stress or crisis, LSCBM size increases, reflecting stronger synchronization; during stable or fragmented periods, it contracts. The 2015 Chinese stock-market crash is the clearest example. In that year,

• $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$03,
• average positive validated correlation $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$04,
• $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$05, and the paper interprets the large balanced core as a crisis-driven cohesive subsystem.

The pandemic period is described as two-phase. In 2020, elevated co-movement raised the LSCBM size to $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$06. In 2021, fragmentation dominated, $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$07 dropped sharply, and the LSCBM collapsed to size $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$08. The algorithm is therefore presented not merely as a static clustering tool but as a lens on shifting market cohesion.

Another reported feature is annual compositional turnover. The paper states that there is almost no overlap between consecutive years in several recent periods, so the core is not permanent. Dominant sectors rotate over time:

• 2013–2014: Energy,
• 2015: Industrials,
• 2016–2017: Information Technology,
• 2019–2020, 2022: Financials,
• 2021: Materials,
• 2024: Industrials / Building Products / Industrial Machinery.

Although the SCBM definition permits negative cross-faction ties, the Chinese data do not realize that possibility inside the detected LSCBMs. The paper states that all identified LSCBMs over 2013–2024 contain only positive edges, consistent with the very small proportion of significant negative edges in the full yearly networks: $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$09 The financial reading is correspondingly narrow: within these core modules there are no inherent hedging opportunities; rather, they behave as concentrated common-risk units.

5. Computational status and limitations

The paper claims that MaxBalanceCore has time complexity $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$10 and space complexity $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$11. The stated rationale is that constructing $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$12 is $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$13, pruning and expansion are $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$14 in the worst case for each seed, and the cap of at most $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$15 seeds is treated as a constant. Practical scalability is attributed to three choices: high-impact seed restriction, early structural-balance pruning, and sparsity induced by significance filtering and the strength threshold $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$16 (Qing et al., 7 Aug 2025).

Simulation evidence is reported in two synthetic settings. When varying network size with $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$17, $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$18, $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$19, and $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$20, the paper states that MaxBalanceCore consistently identifies the true LSCBM and that runtime for networks up to 10,000 nodes is within about 20 seconds. When varying module asymmetry with $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$21, $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$22, and $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$23, it states that MaxBalanceCore always recovers the true LSCBM exactly, even when $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$24 is empty. These are strong empirical claims, but they are claims about the reported planted examples rather than about worst-case approximation quality.

Several limitations are explicit. First, the algorithm is heuristic and has no approximation guarantee. Second, it explores only the top 100 seeds and uses greedy one-pass expansion, so globally optimal modules may be missed. Third, the paper notes that there are no prior algorithms specifically for LSCBM, and correspondingly it provides no direct algorithmic baseline comparison to clique methods, signed community detection, exact branch-and-bound, or $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$25-core-style methods.

The statistical front-end also has caveats. The network depends on Pearson correlation, a pairwise $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$26-test, and a significance level $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$27. The paper does not discuss multiple-testing correction across the $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$28 pairwise tests, robustness to heavy tails or nonlinear dependence, or alternatives to annual fixed windows. Output is also materially dependent on the strength threshold $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$29; the paper states that $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$30 decreases monotonically as $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$31 increases from $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$32 to $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$33.

Finally, the paper acknowledges notation issues. The Pearson and $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$34-statistic formulas contain minor typesetting omissions, and the pseudocode definition of strong_candidates mixes $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$35-thresholding with a ternary signed matrix $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$36. These inconsistencies do not obscure the intended logic, but they limit the precision with which the pseudocode can be treated as a drop-in implementation specification (Qing et al., 7 Aug 2025).

6. Related meanings of “balanced core” in adjacent literatures

Outside statistically validated stock networks, the exact name MaxBalanceCore does not generally appear, but several papers define structurally similar objects. These constructs illuminate the broader semantic field of “balanced core,” while remaining distinct from the stock-network algorithm.

Domain	Closest construct	Relation to MaxBalanceCore
Payment systems	Maximum-weight balanced payment subsystem $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$37	Largest internally settleable balanced subnetwork (Fleischman et al., 2020)
Payment-channel rebalancing	Maximum feasible balancing circulation $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$38	Core optimization for globally optimal rebalancing (Avarikioti et al., 2021)
Blockchain faucets	AMF / WAMF claim logic	Computes a user’s fair claimable balance under max-min fairness (Metin et al., 2021)
Cooperative game theory	Minimal balanced collections / balanced sets	Balancedness certificates for core nonemptiness and core stability (Mermoud et al., 8 Jul 2025)
Cooperative game theory	Closest balanced game and least square core	Projection onto balanced-game polyhedron to recover a nonempty core (García-Segador et al., 16 Jan 2026)

In “Balancing the Payment System”, the matrix

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$39

is proved to be a maximum-weight balanced payment subsystem, meaning the largest settleable-in-full internally balanced subnetwork that can be removed without changing the net positions of the remainder (Fleischman et al., 2020). In “HIDE & SEEK: Privacy-Preserving Rebalancing on Payment Channel Networks”, the closest analogue is the bounded-circulation optimization

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$40

subject to flow conservation at every node and edge-capacity constraints, i.e. a maximum feasible balancing circulation (Avarikioti et al., 2021). In “Max-min Fairness Based Faucet Design for Blockchains”, the analog is the AMF/WAMF claim computation, where the closest notion to a maximum fair balance is the current claimable amount

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$41

or, in the unweighted case, $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$42 (Metin et al., 2021).

Two cooperative-game papers push the connection in a different direction. “Minimal balanced collections and their applications to core stability and other topics of game theory” develops balancedness certificates based on

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$43

and, at a second level, balanced sets

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$44

using them to characterize core nonemptiness, exactness, extendability, feasible collections, and core stability (Mermoud et al., 8 Jul 2025). “On the closest balanced game” then defines the orthogonal projection of a game onto the polyhedron of balanced games and proposes the least square core, with the key reconstruction formula

$$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$45

where $$r_{i}(\tau)=\mathrm{log}\frac{ P_{i}(\tau)}{P_{i}(\tau-\triangle \tau)},$$46 is the closest balanced game (García-Segador et al., 16 Jan 2026).

This broader comparison suggests that “MaxBalanceCore” is best treated as a domain-specific label whose exact meaning depends on the ambient mathematical object: a structurally balanced stock subgraph, an internally settleable liabilities subnetwork, a maximum circulation, a max-min-fair claim function, or a balanced cooperative-game projection. In the strict sense, however, the exact identifier denotes the seed-and-prune heuristic for LSCBM detection in statistically validated stock networks (Qing et al., 7 Aug 2025).

7. Conceptual interpretation

MaxBalanceCore is most precisely understood as a search procedure for a maximum-size complete signed subgraph that satisfies two simultaneous constraints: strong validated dependence and structural balance. Because the stock-network paper uses validated Pearson correlations, the algorithm is neither a generic signed-clique solver nor a general community-detection method. Its target is narrower: a core market subsystem in which every surviving pair of stocks is strongly linked and every triangle has positive sign product (Qing et al., 7 Aug 2025).

This focus explains both its strengths and its limits. The method is well aligned with positive-dominated crisis regimes, where the theory predicts large all-positive modules and the Chinese data indeed exhibit large, synchronized cores. A plausible implication is that the algorithm is especially informative when market-wide co-movement is the dominant statistical feature. Conversely, when the market is fragmented, when negative edges are scarce, or when the meaningful latent structure is not structurally balanced in the Harary sense, MaxBalanceCore’s output should be interpreted as one particular balance-constrained core, not as an exhaustive description of dependency structure.

In the wider literature, the recurring idea behind MaxBalanceCore-like constructions is the extraction of a largest or closest balanced object under explicit resource, flow, or coalition constraints. What varies from field to field is the underlying conserved quantity: obligation weight in payment systems, circulation volume in payment-channel networks, claimable ether in blockchain faucets, or coalition worth in cooperative games. The stock-network version specializes that pattern to statistically validated signed dependence and makes the “balanced core” operational as the LSCBM (Qing et al., 7 Aug 2025).