Industry-Specific Product Space Analysis

• Industry-Specific Product Space is a network-based framework that maps product relatedness using data-driven analysis of co-production, co-export, or co-consumption patterns.
• It employs empirical metrics such as RCA and proximity calculations to identify core products, supply-chain risks, and transition pathways.
• Network analytics including centrality measures and community detection provide actionable insights for industrial policy, investment, and diversification strategies.

An industry-specific Product Space is a network-based quantitative framework that maps the relatedness among products or components within a specific industry by analyzing patterns of co-production, co-export, or co-consumption. It operationalizes the concept that productive diversification—and thus industrial upgrading or transformation—is strongly path-dependent, with a region’s or firm’s likelihood of moving into a new product governed by the similarity of its capability requirements to those already present. The framework extends the original country-level Product Space methodology by restricting analysis to a targeted subset of products (e.g., automotive components, pharmaceuticals, or retail SKUs) and employing industry-relevant data to illuminate diversification opportunities, core products, supply-chain risks, and transition pathways.

1. Theoretical Foundations and Capability-Based Logic

The industry-specific Product Space is rooted in formal capability-accumulation models, wherein each product $i$ requires a vector of binary capabilities $p_i\in\{0,1\}^m$, and each country or entity $c$ possesses a (latent) capability vector $c_c\in\{0,1\}^m$ (O'Clery et al., 2018). The probability that $c$ can competitively produce $i$ is exponentially suppressed by the number of missing capabilities:

$P(J_{c,i}=1)=q^{|p_i|-c_c\cdot p_i}$

where $q$ is the per-capability acquisition probability and $\Delta_{c,i}=|p_i|-c_c\cdot p_i$ is the capability gap. Inference of capability overlap between products $i$ and $j$ is carried out via the empirical log-likelihood ratio:

$$E_{i,j} = \log\left(\frac{P(M_{c,j}=1|J_{c,i}=1)}{P(M_{c,j}=1)}\right)$$

which is proportional to the true capability overlap $|p_i^j|$ up to constants, motivating a directed, weighted network representation.

Industry-specific Product Spaces therefore encode path-dependencies at fine sectoral resolution, identifying ecosystems of precursor products that statistically precede the emergence of each target product (O'Clery et al., 2018, Bartuska et al., 15 Dec 2025).

2. Data Assembly and Network Construction

The construction of an industry-specific Product Space involves several data-driven steps (Bartuska et al., 15 Dec 2025, Pachot et al., 2021, Zinoviev et al., 2015):

• Product Universe Selection: Curate a comprehensive set of products or components within the industry, mapped (for goods) to standardized classification schemes such as HS-6, NACE, or CPC codes.
• Revealed Comparative Advantage (RCA): For each region $r$ or firm $f$ and product $p$, calculate

$$\mathrm{RCA}_{r,p} = \frac{X_{r,p}/\sum_{p'} X_{r,p'}}{\sum_{r'} X_{r',p}/\sum_{r',p'} X_{r',p'}}$$

yielding a binary matrix $M_{r,p} = 1$ if $\mathrm{RCA}_{r,p}\geq 1$.

• Proximity Calculation: Estimate capability or know-how proximity between each pair of products $(p,p')$ using:

$$\phi_{p,p'} = \frac{\sum_r M_{r,p} M_{r,p'}}{\max(\sum_r M_{r,p},\ \sum_r M_{r,p'})}$$

or, equivalently, the minimum of conditional co-export frequencies. In directed variants, $E_{i,j}$ or its capped variant $\hat\phi_{ij}$ is used.

• Network Assembly: Construct a weighted network with products as nodes and edges weighted by $\phi_{p,p'}$ (undirected) or $E_{i,j}$ (directed). Prune weak links (typical thresholds $\phi \geq 0.1\text{--}0.45$) to enhance interpretability.

The table below summarizes core quantities:

Notation	Description	Formula
$M_{r,p}$	Specialization indicator	$M_{r,p} = 1$ if $\mathrm{RCA}_{r,p} \geq 1$, else $0$
$\phi_{p,p'}$	Proximity/capability-overlap	$$\sum_r M_{r,p} M_{r,p'} / \max(\sum_r M_{r,p},\sum_r M_{r,p'})$$
$E_{i,j}$	Empirical ecosystem link (directed)	$$\log\left(\frac{P(M_{c,j}=1|J_{c,i}=1)}{P(M_{c,j}=1)}\right)$$

3. Network Analytics and Structural Diagnostics

Several centrality and community analysis tools are applied to characterize the Product Space network (Bartuska et al., 15 Dec 2025, O'Clery et al., 2018, Zinoviev et al., 2015):

• Closeness Centrality: For node $p$, compute

$C_p = \frac{N-1}{\sum_{q\neq p} d(p,q)}$

where $d(p,q)$ is the shortest-path distance, restricted to competitively exported products in region-level subspaces.

• Core-Periphery Analysis: Employ $k$-core decomposition by iteratively pruning nodes with total degree $<k$. Core products require many inputs and occupy dense network interiors; periphery nodes correspond to niche or less-integrated products.
• Betweenness Centrality: The fraction of all shortest $j \rightarrow i$ paths that cross a node distinguishes "transition products," pivotal as stepping stones in development trajectories (O'Clery et al., 2018).
• Community Detection: Methods such as Louvain or $k$-clique percolation (for retail mini-categories) partition the network into clusters representing complementary or substitutable product sets (Zinoviev et al., 2015).

These diagnostics highlight industry landmarks (high in-degree), ecosystem contributors (nodes feeding many targets), transition products (high betweenness), and delineate distinct technological or capability clusters.

4. Predictive Modelling and Industrial Pathways

Industry-specific Product Spaces are used for out-of-sample predictions and policy analysis (Bartuska et al., 15 Dec 2025, O'Clery et al., 2018):

• Density/Closeness as Predictors: Logistic regression models relate the probability of gaining comparative advantage in product $p$ to its network closeness to the existing export basket. Empirical results show that in the EV sector, a $1\,\sigma$ increase in closeness centrality raises the probability of entering an EV market by $\approx 52\%$ (Bartuska et al., 15 Dec 2025).
• Sector-Level Multipliers: Sectoral averages of closeness (e.g., for "vehicles" or "aluminium products") can predict future diversification capacity. The appearance of new strengths is strongly path-dependent on pre-existing sectoral configuration.
• Risk Mapping: By combining closeness metrics with supply risk indicators (e.g., Herfindahl–Hirschman Index for import concentration), it is possible to map trade-offs between growth prospects and vulnerability to supply interruption.
• Regional or Firm Policy Recommendations: Analysis identifies which upstream capabilities to target, whether via R&D, subsidies, supply-chain restructuring, or industrial policy, to maximize the probability of successful industry entry or transformation (Bartuska et al., 15 Dec 2025, Pachot et al., 2021).

5. Adaptations and Applications Across Sectors

The methodology generalizes to any domain with sufficient production, export, or transaction data (Bartuska et al., 15 Dec 2025, O'Clery et al., 2018, Zinoviev et al., 2015, Han et al., 2021):

• Automotive Sector (EV Transition): Rich product spaces at both the firm-component and country-HS6 levels have revealed core EV clusters (batteries, power electronics, motor controllers) and persistent legacy clusters (ICE parts, transmission systems). Centralities in machinery and electrical equipment have predicted rapid capability diffusion in certain regions (Bartuska et al., 15 Dec 2025).
• Retail Consumption (Mini-Categories): Co-purchase networks of SKUs have been decomposed into small, structurally distinct "mini-categories" (cliques, stars, chains), providing a fine-grained, user-centric taxonomy for project or campaign design (Zinoviev et al., 2015).
• Manufacturing Flexibility and Resilience: Combination of product space with input–output tables enables facility-level recommendations for supplier substitution, productive jumps, and territorial resilience quantification based on proximity structure (Pachot et al., 2021).
• High-dimensional Design Industries: In font and design markets, convolutional neural networks embed unstructured attributes into Euclidean product spaces, enabling spatial competition and differentiation analysis (Han et al., 2021).

6. Domain-Specific Considerations and Limitations

• Data Requirements: For valid industry-level results, detailed, high-granularity transaction or export data (at HS6, firm-component, or SKU level) is crucial, as is alignment of classification systems and sufficient variation across regions or firms.
• Representation of Intangibles: Standard product spaces capture only observable, codified outputs; services, digital products, and organizational capabilities require methodological adaptation or additional data layers.
• Causality vs. Association: Proximity or network structure implies likelihood, not necessity; realized diversification depends on policy, factor endowments, infrastructure, and macroeconomic conditions that are exogenous to the model.
• Sectoral Specificity: Anchoring external nodes (e.g., chemical precursors for pharmaceutics) may be necessary in highly interconnected or value-chain dense industries.
• Overfitting Risks: Thresholds for link pruning, choice of proximity formula, and network resolution parameters must be validated for robustness across time windows and regional granularity (O'Clery et al., 2018, Bartuska et al., 15 Dec 2025).

7. Synthesis and Policy Implications

The industry-specific Product Space provides a rigorous, network-theoretic instrument for diagnosing diversification potential, technology cluster structure, and the sequential unfolding of new industrial capabilities. Empirical applications have demonstrated that central positions in the space are associated with higher rates of upgrading and diversification, while peripheral sectors face both opportunity and risk. In practice, this framework informs sectoral roadmaps, guides region- or firm-level investments, and supports supply-chain derisking by quantifying adjacent possibilities and latent productive potential (Bartuska et al., 15 Dec 2025, Pachot et al., 2021, O'Clery et al., 2018, Zinoviev et al., 2015).