Synthetic Firm-Level Dataset with Realistic Credit Risk and Financial Structures

This repository contains a fully synthetic firm-level dataset comprising approximately 20,000 companies, constructed to mirror the statistical and structural characteristics of proprietary datasets used by commercial information providers, such as credit bureaus and business intelligence firms like Coface, Equifax, B&D, among others. While no real firms or identifiers are included, the data reflect realistic distributions and dependence patterns typically observed in firm-level financial, credit, and risk-related variables.

The dataset was generated using generative modeling techniques calibrated on a confidential reference dataset that cannot be shared. The objective of the generation process is not to replicate individual records, but to preserve the joint statistical structure of the data, including marginal distributions and cross-variable dependencies that are relevant for modeling and empirical analysis.

To assess fidelity, the synthetic data were systematically compared to the real reference dataset using multiple validation criteria. These include comparisons of descriptive statistics, distributional similarity, and dependence structures. Overall, the synthetic data reproduce the main structural patterns observed in the real data, while exhibiting deviations that are explicitly measured and reported.

In particular, the average absolute difference in pairwise correlations between real and synthetic data is 0.189, and several variables show statistically significant discrepancies according to Kolmogorov–Smirnov and Wasserstein distance metrics. These results reflect both the realism achieved by the generative process and the inherent approximation error associated with synthetic data generation.

The dataset is intended for research, methodological development, benchmarking, and reproducible experimentation in areas such as credit risk modeling, firm dynamics, machine learning, and applied econometrics. It must not be interpreted as representing real firms or actual economic populations, nor should it be used for operational or policy decision-making.

Variable Description

Each observation corresponds to a synthetic firm. The dataset includes the following variables:

Firm_ID: Unique synthetic identifier for each firm. This identifier is randomly generated and carries no real-world meaning.

Sector: Categorical variable indicating the firm’s primary sector of activity (e.g., Manufacturing, Services, Technology). Sectoral labels are synthetic but reflect realistic sectoral composition.

Region: Categorical variable capturing the broad geographic region in which the firm operates (e.g., USA, LatAm). This variable is intended to proxy for regional heterogeneity in firm environments.

Size: Continuous variable measuring firm size, expressed as a synthetic scale proxy (e.g., revenues or total assets). The distribution is right-skewed, consistent with empirical firm-size distributions.

Leverage: Continuous variable representing the firm’s leverage ratio, defined as the proportion of debt relative to total financing. Values are bounded between 0 and 1.

Profit_Margin: Continuous variable capturing firm profitability, defined as profits relative to revenues. This variable may take negative values, reflecting loss-making firms.

RD_Intensity: Continuous variable measuring research and development intensity, expressed as a ratio relative to firm size. Higher values indicate greater innovation effort.

Org_Complexity: Discrete variable capturing organizational complexity, proxied by an integer scale (e.g., number of operational layers or managerial depth).

base_hazard: Baseline hazard component used in the data-generating process for firm exit or failure. This variable reflects exogenous risk common to firms.

cov_effect: Composite covariate effect summarizing the contribution of firm-specific characteristics to the hazard rate in the underlying survival model.

hazard: Firm-specific hazard rate combining the baseline hazard and covariate effects. This variable governs the probability of firm exit over time.

Event_Time: Continuous variable representing the simulated time to event (e.g., exit, failure, or censoring), measured in arbitrary time units.

Status: Binary event indicator, equal to 1 if the firm experiences the event of interest and 0 if the observation is right-censored.