📈 economics

Industry Aware Firm Level Network Reconstruction

This paper proposes an enhanced firm-level production network reconstruction method that integrates sectoral input-output constraints to achieve a near-perfect fit with macroeconomic data, significantly outperforming standard size-based approaches in replicating input-output structures.

Original authors: Mitja Devetak, Antoine Mandel

Published 2026-03-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mitja Devetak, Antoine Mandel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the economy as a massive, invisible web of connections. Every company is a node in this web, and every time one company sells something to another, a thread is pulled tight between them. These threads form a production network.

Understanding this web is crucial. If a major thread snaps (a big company goes bankrupt), it can cause a domino effect, shaking the entire economy. Economists call this "systemic risk."

The Problem: The Map is Missing
The trouble is, we don't have a complete map of this web. We know how much money each company makes (their sales) and how much they spend (their costs). We also know the big-picture flow between industries (e.g., how much the "Car Industry" buys from the "Steel Industry").

But we don't know the specific, tiny threads connecting Company A to Company B. It's like knowing the total traffic on a highway system and the number of cars in each city, but having no idea which specific car went from which garage to which destination.

The Solution: Reconstructing the Web
The authors of this paper are like digital detectives trying to rebuild the missing map. They use a method called Network Reconstruction. They take the known data (sales, costs, and industry totals) and use math to guess the missing connections.

Think of it like trying to guess the seating arrangement at a huge wedding. You know:

How many people are at each table (Industry totals).
Who is bringing a gift to whom (Sales/Strengths).
But you don't know exactly who is sitting next to whom.

The Old Way vs. The New Way

The Old Way (Standard Methods): Previous methods tried to guess the connections based mostly on company size. "Big companies probably trade with other big companies." It's a bit like guessing the wedding seating just by matching tall people with tall people. It's okay, but it often gets the "neighborhoods" wrong. It might put a Forestry company next to a Car Factory just because they are both big, even though they rarely talk to each other.
The New Way (Industry-Aware): The authors added a new rule: Respect the Neighborhoods. They used Input-Output Tables (which are like a census of industry-to-industry trade) to say, "Okay, the Car Industry buys a lot from Steel, so we need to make sure there are many threads connecting Car companies to Steel companies."

The Tools They Used
To build this map, they used two main tools:

The "Gravity" Model (for drawing the lines): Imagine companies are planets. The bigger they are, the stronger their gravity. The new method adds a twist: Planets in the same "solar system" (industry) have a special magnetic pull. This helps draw the lines (connections) in the right places first.
The "Balancing Act" (for filling in the numbers): Once the lines are drawn, they need to decide how much money flows through each line.
- Method A (Maximum Entropy): This is like trying to distribute water through a pipe system by guessing the most "random" way possible that still fits the total volume. The authors found this method was a bit too chaotic; the water pressure (weights) varied wildly and didn't match reality well.
- Method B (Iterative Proportional Fitting - IPF): This is like a meticulous accountant. They start with a guess, check the totals, adjust the numbers, check again, and tweak again until the numbers balance perfectly. The authors found this "accountant" method was much better at getting the specific amounts right, especially when they added the industry rules.

What They Found
They tested their new method using real data from Hungary.

The Good News: Their new method was a near-perfect fit for the big-picture data. If the Input-Output table said "Industry A sends $100 to Industry B," their reconstructed map showed exactly that. The old methods were way off, sometimes missing the mark by 100%.
The Bad News: While they got the big picture right, they still struggled with the fine details.
- They couldn't perfectly predict the shape of the network (e.g., how many friends a company has).
- They couldn't accurately identify the "Super-Connectors"—the specific small companies that, if they failed, would crash the system. The new map showed the right amount of traffic between industries, but it couldn't pinpoint exactly which small firms were the critical ones.

The Takeaway
This paper is a major step forward in mapping the economy. It proves that if you want to understand how money flows between companies, you must use industry-level data (Input-Output tables) to guide your guesses.

However, it's also a humble reminder that even with the best math and the most data, we still can't see the whole picture perfectly. We can build a map that shows the right highways and traffic volumes, but we still can't always tell you exactly which specific car is driving on which road. This means we need to keep refining our tools to better understand the hidden risks in our economic web.

1. Problem Statement

The paper addresses the challenge of reconstructing firm-to-firm production networks when direct relational data (who sells to whom) is unavailable. While macro-level data (Input-Output tables) and firm-level non-relational data (sales/revenues) are often available, the specific topology and weights of the network connecting individual firms are missing.

Existing reconstruction methods typically rely on:

Firm size (strengths): Using in-strength (total purchases) and out-strength (total sales) to infer connections.
Maximum Entropy (MaxEnt) or Iterative Proportional Fitting (IPF): To generate an ensemble of networks consistent with these constraints.

The Gap: Standard methods often ignore sectoral constraints provided by Input-Output (IO) tables. Consequently, reconstructed networks may satisfy firm-level totals but fail to reproduce the macro-level flow structure between industries (e.g., assigning high trade probabilities between sectors that rarely interact, such as forestry and car manufacturing). This limits the empirical applicability of these models for systemic risk analysis.

2. Methodology

The authors propose a two-step reconstruction framework that integrates meso-level constraints (sectoral IO flows) into the reconstruction process. The approach separates the reconstruction of the binary topology (who is connected to whom) from the weights (monetary flow values).

A. Binary Network Reconstruction (Topology)

The goal is to generate a random adjacency matrix $A$ where the probability of a link $p_{ij}$ exists.

Standard Approach (dcGM): The authors utilize the Density Corrected Gravity Model (dcGM). This assumes link probability depends on the product of firm $i$ $i$ 's out-strength and firm $j$ $j$ 's in-strength, calibrated to match the network's mean degree.
- Limitation: It ignores industry sectors.
Proposed Approach (dcIAGM): The authors introduce the Density Corrected Industry-Aware Gravity Model (dcIAGM).
- This extends the fitness ansatz to include a sector-pair factor ( $S_i, S_j$ ).
- The probability $p_{ij}$ is proportional to $z \cdot s_{S_i, S_j} \cdot s^{out}_i \cdot s^{in}_j$ , where $s_{S_i, S_j}$ is the total flow between the two sectors in the IO table.
- This ensures that links are more likely between firms in sectors that historically trade heavily.

B. Weighted Network Reconstruction

Once a binary topology $A$ is sampled, the weights $W$ are assigned. The paper compares two methods:

Iterative Proportional Fitting (IPF-industry):
- An extension of the classic IPF algorithm.
- It iteratively scales weights to satisfy three sets of constraints simultaneously:
  1. Firm out-strengths ( $s^{out}_i$ ).
  2. Firm in-strengths ( $s^{in}_j$ ).
  3. Sector-pair flows ( $s_{S,B}$ from the IO table).
- The weight formula is $w_{ij} = a_{ij} u_i v_j g_{S_i, S_j}$ , where $u, v$ are firm-level factors and $g$ is a sector-pair scaling factor.
CReM-B (Conditional Random Maximum Entropy):
- A MaxEnt approach that assigns weights based on an exponential distribution conditional on the edge existing.
- The authors extend CReM-B to include sector-pair constraints in the entropy maximization problem.
- Finding: The authors derive analytical expressions for the MaxEnt solution but find it suffers from high variance in sparse networks.

3. Key Contributions

Integration of IO Constraints: The paper formally derives analytical expressions for MaxEnt solutions that incorporate sectoral IO constraints for both binary and weighted reconstruction.
Algorithmic Extension:
- Proposes dcIAGM for topology generation.
- Proposes IPF-industry, an extension of IPF that enforces sectoral flow constraints alongside firm-level constraints.
Comparative Framework: Provides a rigorous numerical comparison between standard methods (dcGM + CReM-B/IPF) and the proposed industry-aware methods (dcIAGM + IPF/CReM-B) using real-world data.
Scalability: The proposed methods are computationally efficient enough to handle networks with ~200,000 firms, a significant improvement over previous configuration models limited to ~5,000 firms.

4. Results

The methods were tested using Hungarian firm-level VAT data (2021) harmonized with OECD Input-Output tables.

A. Flow Reconstruction Accuracy

Sectoral Flows: The dcIAGM + IPF-industry pipeline achieved an "almost perfect fit" to the Input-Output table data.
- Error Reduction: The total absolute error in IO table flows dropped from 110% (standard methods) to **4.1%** (dcIAGM + IPF).
- Standard methods (dcGM) failed to reproduce sectoral structures, resulting in massive deviations (100%+ error).
Firm Strengths: IPF-based methods significantly outperformed CReM-B in reconstructing the exact in/out flows of individual firms.

B. Structural Statistics (Micro-level)

Despite the macro-level success, all methods struggled to reproduce specific micro-level structural characteristics:

Degree Distributions: The tail exponents of in/out degrees were not accurately captured (reconstructed exponents were ~2.0, while empirical ranges are 1.4–1.6 for out-degree and 2–3 for in-degree).
Correlations: The correlation between strength and degree was flattened in reconstructed networks, failing to match empirical OLS slopes.
Clustering & Reciprocity: While some metrics (like reciprocity) were close to empirical ranges, others (like local clustering) varied significantly depending on the method.

C. Economic Systemic Risk Index (ESRI)

The authors evaluated the reconstructed networks using the ESRI, a metric identifying systemically important firms.

Overestimation: Reconstructed networks yielded significantly higher average ESRI scores ( $5 \times 10^{-4}$ ) compared to the empirical network ( $3 \times 10^{-5}$ ).
Top Firm Identification: All methods performed poorly at identifying the most systemically important firms. Only ~36-37% of the top 100 empirical firms were recovered, and 0% of the top 10 firms were consistently identified across realizations.
Conclusion: While adding IO constraints improves aggregate flow recovery, it does not resolve the difficulty in pinpointing specific critical nodes (systemic risk) without more granular data.

5. Significance and Conclusion

Necessity of Sectoral Constraints: The paper demonstrates that ignoring Input-Output constraints leads to structurally invalid production networks that cannot be used for reliable macro-economic modeling.
Methodological Preference: For weight reconstruction in the presence of IO constraints, IPF-industry is superior to MaxEnt-based CReM-B due to lower variance and better adherence to constraints.
Limitations: The study highlights a fundamental gap: while aggregate flows can be reconstructed accurately, micro-level heterogeneity (specific degree distributions and specific node importance) remains elusive with current public data and constraint sets.
Future Work: The authors suggest that future improvements require better data (e.g., supply-chain specific IO tables) or methods that can capture firm-level heterogeneity beyond simple strength and sector constraints.

In summary, this paper provides a robust, industry-aware framework for reconstructing production networks that successfully aligns with macro-economic sectoral flows, offering a significant improvement over standard gravity models, though challenges remain in accurately modeling the fine-grained topology required for precise systemic risk assessment.