Largest Strong-Correlation Balanced Module (LSCBM)

• LSCBM is a maximum-size subgraph in statistically validated signed networks where all pairwise correlations are strong (|W| ≥ σ) and every triangle is structurally balanced.
• It enforces that each edge meets a minimum strength and that every triad's sign-product is positive, ensuring both density and balance in the module.
• Detection methods like MaxBalanceCore and correlation clustering efficiently extract LSCBM, offering insights for financial risk assessment and network modularity.

Searching arXiv for the cited papers and closely related structural-balance work. Largest strong-correlation balanced module (LSCBM) denotes a maximum-size subset of vertices in a signed correlation network such that all pairwise relations inside the subset are both sufficiently strong and structurally balanced. In the stock-network formulation, an LSCBM is the largest subset of stocks for which every retained pairwise correlation is statistically validated, every such correlation exceeds a strength threshold in magnitude, and every triangle has positive sign-product (Qing et al., 7 Aug 2025). In a separate correlation-clustering operationalization, the term is used for the largest module extracted from the complete set of optimal structurally balanced partitions of a signed graph, with ties resolved by a sign-consistent module-strength score proposed in the synthesis built on exact optimal-solution enumeration (Arinik et al., 2023).

1. Formal definition

In the statistically validated signed-network framework, let $G=(V,E,W,S)$ be a weighted signed graph where $W_{ij}=\hat C_{ij}\in[-1,1]$ and $S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$. A subset $U\subseteq V$ with $|U|\ge 3$ is a strong-correlation balanced module (SCBM) if two conditions hold: every pair $i\ne j$ in $U$ has a nonzero strong edge, meaning $W_{ij}\ne 0$ and $|W_{ij}|\ge \sigma$, and every distinct triple satisfies the structural-balance condition $S_{ij}S_{jk}S_{ki}=+1$ (Qing et al., 7 Aug 2025).

The largest strong-correlation balanced module is then defined by the optimization problem

$W_{ij}=\hat C_{ij}\in[-1,1]$0

This definition combines a complete strong-edge requirement with a triadic sign-consistency requirement. The first condition enforces dense, high-magnitude dependence inside the candidate module; the second excludes frustrated signed triangles.

An equivalent form uses a bipartition. The subset $W_{ij}=\hat C_{ij}\in[-1,1]$1 is balanced if it can be decomposed as $W_{ij}=\hat C_{ij}\in[-1,1]$2, $W_{ij}=\hat C_{ij}\in[-1,1]$3, such that all edges within $W_{ij}=\hat C_{ij}\in[-1,1]$4 are positive, all edges within $W_{ij}=\hat C_{ij}\in[-1,1]$5 are positive, and all edges between $W_{ij}=\hat C_{ij}\in[-1,1]$6 and $W_{ij}=\hat C_{ij}\in[-1,1]$7 are negative, while every edge inside $W_{ij}=\hat C_{ij}\in[-1,1]$8 remains strong. This representation is the one used in the theoretical analysis of random signed graphs (Qing et al., 7 Aug 2025).

2. Statistically validated signed correlation networks

The stock-network construction begins with daily log returns

$W_{ij}=\hat C_{ij}\in[-1,1]$9

where $S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$0 is the closing price of stock $S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$1 at date $S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$2. Over a window of $S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$3 trading days, the Pearson coefficient between stocks $S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$4 and $S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$5 is

$S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$6

Statistical validation is performed by testing $S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$7 against the two-sided alternative $S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$8 using

$S_{ij}=\operatorname{sign}(W_{ij})\in\{-1,+1\}$9

with $U\subseteq V$0 degrees of freedom, and

$U\subseteq V$1

The decision rule used in the paper tests at significance level $U\subseteq V$2 and rejects $U\subseteq V$3 if $U\subseteq V$4 (Qing et al., 7 Aug 2025).

The statistically validated correlation matrix $U\subseteq V$5 is defined elementwise by retaining $U\subseteq V$6 when the null is rejected and setting the entry to $U\subseteq V$7 otherwise, with $U\subseteq V$8 for convenience. The resulting weighted signed network uses $U\subseteq V$9 and $|U|\ge 3$0, with non-edges corresponding to $|U|\ge 3$1.

“Strong” correlation is imposed after statistical validation. The paper retains an edge $|U|\ge 3$2 as strong if $|U|\ge 3$3 and $|U|\ge 3$4, with default $|U|\ge 3$5. The reported empirical analyses do not apply explicit multiple-testing correction, although Bonferroni and Benjamini–Hochberg are noted as possible alternatives (Qing et al., 7 Aug 2025).

3. Structural balance and module semantics

Structural balance is defined at the triad level: for a signed undirected graph, every triangle $|U|\ge 3$6 in a balanced subgraph must satisfy

$|U|\ge 3$7

Equivalently, each triangle contains an even number of negative edges. Thus, either all three edges are positive, or exactly two are negative and one is positive. In the standard interpretation, this encodes consistent triadic sentiment, including the familiar “enemy of my enemy is my friend” pattern (Qing et al., 7 Aug 2025).

For complete signed graphs, the Cartwright–Harary characterization states that structural balance holds if and only if the vertex set can be partitioned into two disjoint subsets $|U|\ge 3$8 and $|U|\ge 3$9 with positive edges inside each subset and negative edges across subsets. In the LSCBM setting this bipartition is combined with the requirement that all retained edges inside the candidate subset are strong. The result is a particularly restrictive object: a dense signed core with full pairwise support and exact balance.

The economic interpretation given in the stock-network formulation follows directly from this geometry. Positive within-cluster edges capture co-movement. Negative cross-cluster edges, when present, support hedging logic, since long exposure to one faction and short exposure to the other can offset shocks. In the Chinese A-share empirical study, however, LSCBMs contained only positive edges, reflecting the near absence of statistically significant negative correlations; under those conditions, the module acts as a cohesive risk unit rather than a natural hedging structure (Qing et al., 7 Aug 2025).

4. Detection via MaxBalanceCore

The direct detection algorithm proposed for statistically validated stock networks is MaxBalanceCore, a heuristic designed for sparse signed graphs. Its design principles are to prioritize high-impact seeds, enforce the bipartition constraints early, and exploit sparsity induced by statistical validation and the strength threshold (Qing et al., 7 Aug 2025).

The procedure begins by constructing a signed adjacency matrix $i\ne j$0 from $i\ne j$1: if $i\ne j$2 and $i\ne j$3, then $i\ne j$4; otherwise $i\ne j$5. For each node, one computes

$i\ne j$6

sorts nodes by decreasing impact, and considers up to the 100 highest-impact seeds, that is, $i\ne j$7 seeds. For a chosen seed, the candidate factions are initialized as $i\ne j$8 and $i\ne j$9.

Pruning is then applied in two stages. Intra-faction pruning removes any vertex in $U$0 that is not positively connected to every other current member of $U$1, and similarly for $U$2. Inter-faction pruning removes nodes violating the requirement that all edges between $U$3 and $U$4 be negative. After forming $U$5, the algorithm attempts expansion: an external node $U$6 can be added to $U$7 if it has strong edges to all current module members, is positive to all vertices in $U$8, and negative to all vertices in $U$9; symmetrically, it can be added to $W_{ij}\ne 0$0 if its signs are reversed with respect to the two factions. The output is the largest module found across all seeds (Qing et al., 7 Aug 2025).

The worst-case complexity is $W_{ij}\ne 0$1 in time and $W_{ij}\ne 0$2 in space. The paper reports practical scalability to networks with up to $W_{ij}\ne 0$3 nodes within tens of seconds, approximately $W_{ij}\ne 0$4 seconds in synthetic tests. The underlying optimization remains NP-hard, so MaxBalanceCore provides no exact optimality guarantee. Nevertheless, in synthetic networks with embedded ground-truth LSCBMs, it recovered the exact LSCBM in $W_{ij}\ne 0$5 of trials across the tested settings (Qing et al., 7 Aug 2025).

5. Correlation-clustering operationalization

A different route to LSCBM arises from correlation clustering (CC) on signed graphs. In this framework, one starts from a signed graph $W_{ij}\ne 0$6 or signed adjacency matrix $W_{ij}\ne 0$7, with $W_{ij}\ne 0$8 depending on whether an observed edge is positive, negative, or absent. A partition $W_{ij}\ne 0$9 is structurally balanced when positive edges fall inside modules and negative edges fall between modules. For general graphs the objective is to minimize disagreement, that is, positive edges cut by the partition plus negative edges placed inside modules (Arinik et al., 2023).

Using a membership vector $|W_{ij}|\ge \sigma$0, the standard unweighted disagreement count is

$|W_{ij}|\ge \sigma$1

The paper’s imbalance functional is

$|W_{ij}|\ge \sigma$2

which is equivalent to the disagreement count. The exact ILP formulation introduces binary variables

$|W_{ij}|\ge \sigma$3

and minimizes

$|W_{ij}|\ge \sigma$4

subject to the triangle inequalities

$|W_{ij}|\ge \sigma$5

The paper’s main contribution is efficient exact enumeration of all optimal CC partitions. It combines a strengthened ILP with branch-and-cut for the first optimum, recurrent neighborhood search (RNS) around each discovered optimum, and “jump” steps implemented by an ILP with an optimality constraint

$|W_{ij}|\ge \sigma$6

together with no-good constraints

$|W_{ij}|\ge \sigma$7

for each previously found partition $|W_{ij}|\ge \sigma$8. Local enumeration in Complete Neighborhood Search (CoNS) is pruned using atomicity conditions and the MVMO property; for unweighted graphs, a key strengthening is

$|W_{ij}|\ge \sigma$9

for each moving vertex in an atomic edit between optimal solutions (Arinik et al., 2023).

LSCBM is not a native term in the CC paper. The synthesis grounded in that paper proposes a CC-consistent module score

$S_{ij}S_{jk}S_{ki}=+1$0

which in unweighted graphs becomes

$S_{ij}S_{jk}S_{ki}=+1$1

Within a given optimal partition, the LSCBM is the module maximizing $S_{ij}S_{jk}S_{ki}=+1$2, with ties broken by $S_{ij}S_{jk}S_{ki}=+1$3. Across the complete optimal set $S_{ij}S_{jk}S_{ki}=+1$4, the proposed rule is

$S_{ij}S_{jk}S_{ki}=+1$5

understood lexicographically. This suggests a bridge between global balance optimization and local module extraction: the CC objective supplies globally optimal balanced partitions, while the LSCBM criterion extracts the largest sign-consistent module from that solution space.

6. Theory, empirical behavior, and limitations

For random signed graphs $S_{ij}S_{jk}S_{ki}=+1$6, where each unordered pair independently receives a positive edge with probability $S_{ij}S_{jk}S_{ki}=+1$7, a negative edge with probability $S_{ij}S_{jk}S_{ki}=+1$8, and no edge otherwise, the LSCBM framework admits asymptotic results. If $S_{ij}S_{jk}S_{ki}=+1$9, $W_{ij}=\hat C_{ij}\in[-1,1]$00, and $W_{ij}=\hat C_{ij}\in[-1,1]$01, then $W_{ij}=\hat C_{ij}\in[-1,1]$02 as $W_{ij}=\hat C_{ij}\in[-1,1]$03. In the general regime with fixed $W_{ij}=\hat C_{ij}\in[-1,1]$04, the expected size satisfies

$W_{ij}=\hat C_{ij}\in[-1,1]$05

where

$W_{ij}=\hat C_{ij}\in[-1,1]$06

Multiplicity also appears asymptotically, with $W_{ij}=\hat C_{ij}\in[-1,1]$07. In the dense regime $W_{ij}=\hat C_{ij}\in[-1,1]$08 and $W_{ij}=\hat C_{ij}\in[-1,1]$09 for $W_{ij}=\hat C_{ij}\in[-1,1]$10, one has $W_{ij}=\hat C_{ij}\in[-1,1]$11 and the LSCBM is with high probability all-positive. In the negative-dominated regime $W_{ij}=\hat C_{ij}\in[-1,1]$12 and $W_{ij}=\hat C_{ij}\in[-1,1]$13, the expected size becomes $W_{ij}=\hat C_{ij}\in[-1,1]$14 (Qing et al., 7 Aug 2025).

Simulation studies reported in the same work support both the algorithmic and asymptotic claims. In synthetic statistically validated networks with embedded balanced factions, MaxBalanceCore recovered the true LSCBM in $W_{ij}=\hat C_{ij}\in[-1,1]$15 of trials for $W_{ij}=\hat C_{ij}\in[-1,1]$16, with runtime remaining at most $W_{ij}=\hat C_{ij}\in[-1,1]$17 seconds at $W_{ij}=\hat C_{ij}\in[-1,1]$18. In random signed graphs under the general regime $W_{ij}=\hat C_{ij}\in[-1,1]$19, the dense regime $W_{ij}=\hat C_{ij}\in[-1,1]$20 with $W_{ij}=\hat C_{ij}\in[-1,1]$21, and the negative-dominated regime $W_{ij}=\hat C_{ij}\in[-1,1]$22, the observed LSCBM size divided by the corresponding theoretical prediction converged to $W_{ij}=\hat C_{ij}\in[-1,1]$23 (Qing et al., 7 Aug 2025).

The empirical analysis on Chinese A-shares from 2013 through 2024 uses annual windows of daily returns and $W_{ij}=\hat C_{ij}\in[-1,1]$24. The proportion of validated positive edges dominates every year, reaching approximately $W_{ij}=\hat C_{ij}\in[-1,1]$25 in 2015 and falling to approximately $W_{ij}=\hat C_{ij}\in[-1,1]$26 in 2021, while the validated negative-edge proportion remains at most $W_{ij}=\hat C_{ij}\in[-1,1]$27. Average positive strength peaks at $W_{ij}=\hat C_{ij}\in[-1,1]$28 in 2015 and remains elevated at $W_{ij}=\hat C_{ij}\in[-1,1]$29 in 2016; average negative strength stays around $W_{ij}=\hat C_{ij}\in[-1,1]$30 to $W_{ij}=\hat C_{ij}\in[-1,1]$31. LSCBM size increases sharply in stress periods and contracts in fragmented regimes: $W_{ij}=\hat C_{ij}\in[-1,1]$32 in the 2015 crash, $W_{ij}=\hat C_{ij}\in[-1,1]$33 in 2016, $W_{ij}=\hat C_{ij}\in[-1,1]$34 in 2021, and $W_{ij}=\hat C_{ij}\in[-1,1]$35 in 2024. Dominant industries rotate over time—Energy in 2013–2014, Industrials in 2015, 2018, and 2024, IT in 2016–2017 and 2023, Financials in 2019–2020 and 2022, and Materials in 2021. Across all years, the detected LSCBMs are all-positive, which the paper interprets as evidence of cohesive co-movement rather than intrinsic long-short hedging structure. Sensitivity analysis over $W_{ij}=\hat C_{ij}\in[-1,1]$36 shows a monotone decline in module share as the threshold tightens, with sharp drops beyond approximately $W_{ij}=\hat C_{ij}\in[-1,1]$37 (Qing et al., 7 Aug 2025).

Several limitations follow directly from the formulations. The stock-network pipeline assumes within-window stationarity of correlations and uses per-pair tests at $W_{ij}=\hat C_{ij}\in[-1,1]$38 without explicit multiple-testing correction in the reported experiments. Structural balance is exact and may exclude near-balanced yet economically meaningful subgraphs. The random-graph theory assumes edge independence and does not model empirical weight structure. In the CC setting, exact enumeration remains expensive because the underlying problem is NP-hard, and large optimal-solution spaces can produce substantial variation in the largest module across optima. A plausible implication is that practical use of LSCBM benefits from stability analysis—whether via repeated appearance across optimal CC partitions or by sensitivity of the detected module to the strength threshold $W_{ij}=\hat C_{ij}\in[-1,1]$39—rather than relying solely on a single maximum-cardinality output.