Paper: arXiv 2502.15726
Authors: Luiz Tavares, Jose Mazzon, Francisco Paletta, Fabio Barros
Abstract
The marketing departments of financial institutions strive to craft products and services that cater to the diverse needs of businesses of all sizes. However, it is evident upon analysis that larger corporations often receive a more substantial portion of available funds. This disparity arises from the relative ease of assessing the risk of default and bankruptcy in these more prominent companies. Historically, risk analysis studies have focused on data from publicly traded or stock exchange-listed companies, leaving a gap in knowledge about small and medium-sized enterprises (SMEs). Addressing this gap, this study introduces a method for evaluating SMEs by generating images for processing via a convolutional neural network (CNN). To this end, more than 10,000 images, one for each company in the sample, were created to identify scenarios in which the CNN can operate with higher assertiveness and reduced training error probability. The findings demonstrate a significant predictive capacity, achieving 97.8% accuracy, when a substantial number of images are utilized. Moreover, the image creation method paves the way for potential applications of this technique in various sectors and for different analytical purposes.
Complexity vs Empirical Score
- Math Complexity: 4.5/10
- Empirical Rigor: 6.0/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper applies a standard CNN architecture (low-to-moderate math complexity) but uses a large dataset (>10,000 images) and reports high accuracy (97.8%), indicating significant data processing and implementation focus. However, it lacks detailed backtesting or live trading results, placing it in the empirical rigor range.
Research Flowchart
flowchart TD A["Research Goal:<br>Assess SME Bankruptcy<br>Risk via Images"] --> B["Data Input:<br>10,000+ Images<br>Generated from SME Data"] B --> C["Methodology:<br>CNN Model Architecture"] C --> D["Computational Process:<br>Training & Learning"] D --> E["Key Outcome:<br>97.8% Accuracy<br>High Predictive Capacity"] E --> F["Impact:<br>Improved Credit Risk<br>Assessment for SMEs"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes generating image representations of SME financial data for analysis with a GoogLeNet-based convolutional neural network to predict default risk.
Source sentence
this study introduces a method for evaluating SMEs by generating images for processing via a convolutional neural network (CNN).
Assumes 2
- A sufficiently large number of images is required for the CNN to achieve high predictive performance. data
Source sentence
The quantity of images, therefore, emerges as a critical factor influencing the enhancement of CNN outcomes.
- Images constructed from accounting account data capture relevant information for default prediction. model
Source sentence
For Division 47, as elucidated in Section 4.2, the models constructed from images based on accounting account data yielded exceptionally reliable predictions
Data
- SME enterprise dataset fundamentals unknown
Source sentence
Within the broader dataset of 105,000 enterprises, a subset of 14,456 companies was extracted based on these categories.
- Monthly data vectors for Division 47 fundamentals unknown
Source sentence
a total of 416,710 monthly data vectors were compiled for Division 47
- Monthly data vectors for Division 86 fundamentals unknown
Source sentence
an additional 116,487 vectors for Division 86
Costs
Not applicable
Finds 5
- The CNN achieved 97.8% accuracy when trained on a large set of SME images. predictive-accuracy
Source sentence
The findings demonstrate a significant predictive capacity, achieving 97.8% accuracy, when a substantial number of images are utilized.
- Models using accounting‑based images for Division 47 reached accuracy exceeding 95% and high test accuracy. predictive-accuracy
Source sentence
... the models constructed from images based on accounting account data yielded exceptionally reliable predictions, as evidenced by their high accuracy and test accuracy rates.
- Combining Division 86 with Division 47 data raised accounting‑image model accuracy to 96.4% and F1 Score to 97.9%. predictive-accuracy
Source sentence
In the case of the model based on accounting account imagery, accuracy escalated to 96.4%, with F1 Score achieving 97.9% — the highest in all training iterations
- When tested on Division 86 alone, the combined model improved test F1 Score to 93.6% versus 71.5% for the single‑division model. predictive-accuracy
Source sentence
establishing the combined model as superior to the one trained exclusively on Division 86 data.
- Divisions with fewer than 4,610 images showed lower accuracy (minimum 92.2%) and higher loss rates. performance
Source sentence
beyond 4,610, the lowest recorded accuracy was 92.2%, although this did not uniformly translate to a low rate of model loss—as evidenced by a 24.2% loss rate in the third training iteration.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 27 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.