Out-of-Block Diversification

• Out-of-Block Diversification is a measure that quantifies a firm's exports outside its core production block, emphasizing the exploration of new markets.
• It employs community-detection algorithms like BRIM on firm–product networks to identify modular production blocks, distinguishing exploration from exploitation activities.
• Empirical results show that higher out-of-block diversification is associated with increased growth (β = 0.016, P<0.001) compared to in-block diversification.

Out-of-Block Diversification is, in the firm-growth literature, a firm-level export diversification measure that distinguishes expansion outside a firm’s algorithmically identified core production area from diversification within that core. In its canonical formulation, a firm’s “block” is not an administrative sector but a data-driven cluster in a modular firm–product export network, and the central empirical claim is that these locations matter: firms with greater diversification outside their own block exhibit higher future growth, whereas firms with more diversification inside their own block exhibit lower future growth (Stefano et al., 29 Jul 2025).

1. Definition and formal structure

The concept is introduced in a setting where firm–product export networks are treated as modular rather than simply nested. Firms tend to export products concentrated in coherent clusters, called production blocks. A firm’s assigned block is interpreted as its core production area. Out-of-block diversification then measures how many products a firm significantly exports outside that core, while in-block diversification measures how many are exported inside it (Stefano et al., 29 Jul 2025).

The construction begins from annual firm–product export data $E_{fp}$. To remove size effects, exports are binarized using a firm-level revealed comparative advantage threshold, producing a binary incidence matrix $M$ with

$$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$

A second indicator,

$$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$

links the diversification counts to the modular partition. The paper’s exact definitions are

$$d^{in}_f=\sum_p M_{fp}\delta_{fp},\qquad d^{out}_f=\sum_p M_{fp}(1-\delta_{fp}),$$

with the decomposition

$d_f=d^{in}_f+d^{out}_f.$

Under this definition, $d^{in}_f$ is the number of products that firm $f$ significantly exports inside its own block, while $d^{out}_f$ is the number of products it significantly exports outside its own block. Conceptually, the distinction is between exploitation of an established capability base and exploration beyond it. The paper’s claim is not merely that “more diversification” matters, but that diversification has different implications depending on whether it remains inside a coherent production core or crosses into other parts of the production ecosystem (Stefano et al., 29 Jul 2025).

2. Detection of blocks and empirical setting

Production blocks are obtained by applying a community-detection algorithm to the bipartite network linking firms and products. The paper uses the BRIM algorithm (“Bipartite Recursively Induced Modules”; Barber, 2007), which maximizes bipartite modularity. In the appendix, modularity is written as

$$Q = \frac{1}{2m}\sum_{ij}\Biggl(A_{ij}-\frac{k_i k_j}{m}\Biggr)\,\delta(c_i,c_j),$$

with the usual meanings of adjacency, degree, total links, and community-membership indicator. In the application, BRIM adapts this logic to a bipartite graph whose nodes are firms and products (Stefano et al., 29 Jul 2025).

A practical issue is longitudinal consistency. Because the export network exists separately each year, year-by-year community detection would generate slightly changing partitions. To avoid this, the analysis restricts attention to firms present throughout 1993–2017, aggregates the firm–product graph across years, thresholds links using $M$0, and runs BRIM on a single aggregate bipartite graph. This produces a stable partition with seven blocks and modularity $M$1. The blocks are industrially interpretable: one large machinery/metal block, and others dominated by electrical machinery, paper/plastic, chemicals/minerals, textiles, food/animals, plus one mixed block. The baseline specification uses 4-digit communities, and the results are reported as robust to 6-digit communities as well (Stefano et al., 29 Jul 2025).

The underlying data combine ISTAT export records and ORBIS financial data. Products are observed at the 6-digit Harmonized System level, totaling 5,203 products, over 1993–2017. Restricting to firms that exported at least one product in every year leaves 18,597 firms. ORBIS financial information is available from 2013 onward for 12,852 of these firms, and that subset constitutes the regression sample. The financial variables include employment, operating revenues, net income, and profit margin (Stefano et al., 29 Jul 2025).

3. Predictive role for firm growth

The econometric design is explicitly predictive and long-run rather than contemporaneous. The panel is collapsed into a cross-section centered at $M$2, with explanatory variables averaged over the backward window $M$3 and outcomes over the forward window $M$4, using $M$5 as the baseline. The core regression is

$M$6

where $M$7 is either Growth or Profit per Employee, $M$8 is the vector of explanatory variables, and $M$9 are sector dummies. The main specification includes log operative revenue, log coherence, EXPY, log out-of-block diversification, log in-block diversification, plus 21 HS sector dummies; standard errors are HC1 robust (Stefano et al., 29 Jul 2025).

The central result is asymmetric across the two diversification measures. In the main OLS regression for Growth, the coefficient on log Out-of-block Diversification is

$$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$0

while the coefficient on log In-block Diversification is

$$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$1

Conditional on firm size, EXPY, coherence, and sector fixed effects, firms with greater diversification outside their core block exhibit higher future growth, whereas firms with more diversification inside their own block exhibit lower future growth. The pattern does not carry over to Profit per Employee: out-of-block diversification has coefficient $$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$2 with $$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$3, and in-block diversification has coefficient $$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$4 with $$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$5, so neither is significant there (Stefano et al., 29 Jul 2025).

The paper also reports nonparametric visual evidence. Average growth rises with the fraction of products exported outside the firm’s cluster until a plateau is reached, and the highest-growth region corresponds to firms with relatively low in-block diversification and high out-of-block diversification. That pattern is consistent with the regression signs (Stefano et al., 29 Jul 2025).

A key robustness result concerns the definition of the blocks themselves. The positive predictive role of out-of-block diversification appears only when blocks are identified through the BRIM community-detection algorithm. If “inside” and “outside” are instead defined using standard Harmonized System sections, the analogous measures are not significant: log out-of-section diversification has coefficient $$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$6 and log in-section diversification has coefficient $$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$7, both not significant. The BRIM-based results are also reported as robust to changing the horizon from $$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$8 to 3 or 5 years, and to using 6-digit rather than 4-digit communities (Stefano et al., 29 Jul 2025).

4. Capabilities, exploration, and the macro–micro contrast

The authors interpret the distinction between in-block and out-of-block diversification through the lens of capabilities, economic complexity, and exploration versus exploitation. In-block diversification is associated with exploiting a coherent local capability base inside the firm’s established production core. Out-of-block diversification is associated with exploration into products linked to other production blocks, and is interpreted as evidence of transferable or latent capabilities that are not fully revealed by the firm’s current product mix (Stefano et al., 29 Jul 2025).

This interpretation also structures the contrast between growth and efficiency. In-block diversification may raise efficiency and is consistent with why coherence predicts Profit per Employee, but it does not generate growth and may even be associated with lower growth if it reflects path dependence, bounded search, or excessive concentration in familiar activities. Out-of-block diversification, by contrast, is linked to future revenue growth rather than contemporaneous profitability per worker. A plausible implication is that exploration across blocks captures expansion opportunities that are not reducible to local specialization alone (Stefano et al., 29 Jul 2025).

The paper situates this finding against the macroeconomic complexity literature. At the country level, diversification is generally growth-enhancing because countries expand across a nested product space without the same strong modular boundaries. At the firm level, however, the product space is more modular, so the location of diversification matters. The distinction is framed explicitly as one between exploration capabilities, associated with diversification outside the core block and sustained growth, and exploitation capabilities, associated with coherent, core-centered activity and higher profit per employee rather than growth. Hence the broader claim that growth depends not just on how much a firm diversifies, but on where its products lie within the production ecosystem, at both local and global scales (Stefano et al., 29 Jul 2025).

5. Related block-based diversification in finance and networked risk

Analogous block-based logics appear in finance, but there the blocks are usually inferred from dependence structure rather than production ecosystems. In graph-theoretic portfolio partitioning, a market is represented as a graph with edge weights $$M_{fp}=1 \quad \text{if } RCA_{fp}>1,\qquad M_{fp}=0 \quad \text{if } RCA_{fp}\le 1.$$9, and normalized graph cuts are used to identify clusters of highly dependent assets. The proposed allocation logic is to distribute capital across weakly connected clusters instead of accumulating many redundant names inside one correlation block. In the reported experiment on the 100 most liquid S&P 500 stocks, equal budget across final clusters after normalized cuts reached Sharpe ratio $$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$0 at $$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$1, versus $$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$2 for equal-weighting and $$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$3 for minimum variance (Dees et al., 2019).

A second financial formulation uses a correlation blockmodel in which assets in the same cluster are highly correlated with each other and have the same correlations with all other assets. The defining dissimilarity is

$$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$4

and the paper’s representative-selection theorem states that, when one asset is selected from each block, the minimum-variance choice is the asset with smallest variance in that block. Empirically, portfolios built from only 15–25 such representatives outperform SPY, sector-ETF portfolios, and several alternative clustering or classification-based subsets over a long sample, which supports the principle that diversification comes from spanning distinct correlation blocks rather than multiplying names within one block (Tang et al., 2021).

Network contagion models complicate the intuition further by distinguishing exposure diversification from damage diversification. Splitting one’s own exposures across more neighbors can increase connectivity and create more contagion channels, whereas limiting the damage a failing node can impose on each neighbor can reduce systemic risk. In that literature, diversification across counterparties is therefore not automatically stabilizing; what matters is whether diversification reduces own concentration only, or also caps outgoing bilateral damage (Burkholz et al., 2015).

An even sharper caveat appears in extreme-risk settings. When losses are independent, positive, and sufficiently heavy-tailed, the diversified weighted average

$$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$5

can be stochastically larger than the one-basket comparator

$$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$6

where all exposure is placed on a single risk chosen at random with probabilities $$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$7. The paper proves that diversification always increases small-threshold exceedance probabilities, and under subscalability conditions this local effect becomes global first-order stochastic dominance. In that domain, “diversifying across blocks” can be worse than concentrating in one randomly selected block (Vincent, 22 Jul 2025).

6. Analogues in machine learning and system design

In machine learning, the phrase does not usually refer to product or correlation blocks, but similar ideas appear whenever predictive robustness depends on spreading reliance across distinct feature groups or model components. One line of work argues that ensemble-based models improve expected OOD generalization by using a more diverse set of spurious features. In the paper’s stylized theory, output-space or weight-space ensembles reduce OOD error because different models rely on different subsets of spurious cues, so no single brittle cue dominates. On the synthetic MultiColorMNIST benchmark with 32 spurious features, ensemble OOD accuracy exceeds that of either constituent model across multiple corruption levels (Lin et al., 2023).

A more skeptical literature shows that diversification alone is not sufficient for OOD generalization. In diversification methods that train multiple hypotheses using unlabeled data, performance is highly sensitive to the unlabeled distribution, the learning algorithm, and the interaction between them. The paper reports that performance can drop by up to 30 absolute points away from a method-specific sweet spot in the unlabeled spurious ratio, that using the second-best model can cause an up to 20\% absolute drop in accuracy, and that increasing the number of hypotheses does not fix these failures (Benoit et al., 2023).

OOD detection with auxiliary outliers yields a related lesson. The generalization bottleneck is not merely lack of outlier supervision but lack of sufficiently diverse outlier support. The proposed diverseMix method expands the effective support of the auxiliary outlier set through adaptive mixup, and the paper argues theoretically that more diverse auxiliary outliers reduce the distribution-shift term in a generalization bound. On CIFAR-10 and CIFAR-100, the method reports FPR95 values of $$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$8 and $$\delta_{fp} = \begin{cases} 1 & \text{if firm } f \text{ and product } p \text{ belong to the same block,}\ 0 & \text{otherwise,} \end{cases}$$9, respectively, outperforming previously listed baselines (Yao et al., 2024).

A systems-level analogue appears in federated learning. Fed-FBD decomposes a ResNet backbone into six functional blocks and maintains a warehouse of $$d^{in}_f=\sum_p M_{fp}\delta_{fp},\qquad d^{out}_f=\sum_p M_{fp}(1-\delta_{fp}),$$0 color variants assembled from independently tracked blocks. The diversification mechanism is architectural rather than statistical: clients can update only assigned block–color lineages, so contamination cannot spread outside them. Reported experiments show a modest IID accuracy gap on adequately sized datasets, attacks confined to the poisoned client’s own blocks with at most $$d^{in}_f=\sum_p M_{fp}\delta_{fp},\qquad d^{out}_f=\sum_p M_{fp}(1-\delta_{fp}),$$1 AUC drift on clean colors, membership-inference AUC essentially at chance, and surgical unlearning of a departed participant at sub-second cost (Chen et al., 10 Jun 2026).

Across these literatures, a common theme emerges. Out-of-block diversification is not equivalent to “having more items.” It is a relational property defined with respect to a block structure—production blocks, correlation clusters, network neighborhoods, feature groups, or functional modules. The substantive question is therefore not whether activity is diversified in the aggregate, but whether it crosses the relevant block boundary, how that boundary is inferred, and what form of performance or risk is being evaluated once it does.