AI-Based Credit Assessment: A Myth to Debunk?
Leanus Analysis Reveals the Uniqueness of SMEs and consequent limits of the Algorithms
Leanus conducted for MF Milano Finanza a detailed analysis of the correlation between the main accounting data of over 6.220 Italian companies operating in the road freight transport sector (code ATECO 4941). Despite their apparent sectoral homogeneity, the study on the 2023 financial statements of the selected companies highlights the extreme heterogeneity of the corporate realities and the consequent difficulty of adopting evaluation criteria that ignore the in-depth analysis of individual cases.
The Main Findings
The analysis of the correlation matrix shows an almost total absence of significant statistical correlations between the accounting variables examined. This suggests that companies, although operating in the same sector, present a notable uniqueness, linked to specific and non-standardizable factors, such as company history, management and strategic choices. These results cast doubt on the reliability of automatic credit assessment processes based exclusively on algorithms or Artificial Intelligence (AI)
The correlation
Correlation measures the degree of dependence between two variables: a value of 1 indicates a perfect positive correlation (when one variable increases, the other also increases proportionally); on the contrary, a value of -1 indicates a perfect negative correlation (when one variable increases, the other decreases). Values equal to or lower than 0,5 indicate a weak or absent correlation, at least from a statistical point of view.
The sample taken into consideration
6.220 SMEs Italian companies (with 2023 Revenues between 500.000 Euro and 50 Million Euro), classified as operating in sector 4941 – Road freight transport. In the case of simultaneous presence of both the Ordinary and Consolidated Budget, the Ordinary Budget was selected (in order to avoid double counting). The overall value of Revenues is close to 27 Billion Euro (up 4,5%) with a EBITDA average of 5%. Average collection time from Customers or DSO equal to 100 days, average payment time of suppliers or DPO equal to 109. These data are widely used for sector statistics and to define the criteria for access to credit but, as argued below, are often misleading.
The Profit and Loss Account
The analysis of the correlation between the items of the Profit and Loss Account only confirms the legal relationship between the size, expressed by Revenues, and the main cost items (Services and Personnel). The connection with the other relevant variables is, however, not very significant. The correlation with Consumption (Purchases + Change in Inventories) is equal to just under 0,6, the correlation with EBITDA (0,4), Depreciation (0,5), EBIT (0,2) and profit (0,5) is absent or weak. In summary, the growth in expenditure for operating costs is proportional to the size but the greater size does not necessarily translate into better economic results and therefore not even into greater revenues for the state, given that the correlation with operating taxes does not reach the value of 0,5. The correlation between the value of EBITDA (and consequently of EBIT) with the value of Taxes is, however, significant, an aspect which confirms the propensity to optimize the tax profile for the majority of companies.
The Balance Sheet
The analysis of the balance sheet items confirms the propensity for overall indebtedness for most companies, the correlation between Total Debts and Revenues is in fact equal to 0,8; furthermore, they highlight the payment habits of the Italian system; the correlation between Revenues and Customer Credits is in fact equal to 0,8 as is the correlation between Revenues and Supplier Debts. Contrary to what would be logical to expect, considering that these are companies operating in the same sector, the correlation with inventories is almost non-existent (0,2)
The Cash Flow Statement
Available liquidity shows a slight correlation with the value of Net Equity (0,6), with the value of EBITDA (0,5); absent or weak correlation with other economic and patrimonial variables. Operating Cash Flow (the cash generated in the financial year by characteristic management) shows an element of attention that is worth reflecting on. In addition to being correlated with the variables on which it directly depends (EBITDA, Profit), shows an interesting, albeit weak (0,5), relationship with net fixed assets (i.e. the sum of tangible, intangible and financial assets net of depreciation funds), partially confirming the importance of investments for cash generation.
In short (Italian only)
The evidence emerging from the analysis of the correlation matrix allows us to confirm the value of the identity of the individual company, not as belonging to a category, a sector or a region, but rather as a unique entity, different from others, characterized by its history, by the people who manage it and work there, by its decisions and by many other factors.
Every company has its own DNA different from that of others and difficult to replicate and only partially, and to a minimal extent, comparable with that of other companies, even less so if identified through sectoral coding which, for obvious reasons of practicality, cannot take into account the unique factors of individual realities.
The immediate consequence of such evidence is that the evaluation processes of companies (whether aimed at accessing credit, M&A operations or otherwise) cannot ignore the search for such uniqueness. The use of standard metrics (high Ebitda, good ratio Pfn/EBITDA or other criteria) open the way to the risks of adverse selection, or rather to the concrete possibility of dedicating attention to undeserving companies and vice versa.
Given the low correlation between the main variables and the lack of the ability to find other information of better or similar quality, mathematical models, based on traditional algorithms or AI systems, are destined to offer distorted evaluation attempts, a risk clearly highlighted by the banking system's control bodies (EBA, Bankit) who expressed themselves in favour of an approach, albeit supported by technology, in which the human factor must continue to play a leading role.
Although technology, in all its current and future forms, is the best partner for organizations and Professionals, the challenge to decode the unique characteristics of companies cannot be delegated to machines alone; the variables are innumerable and often non-repetitive and as such they can hardly be the object of a process entirely industrialized or not supported by the skills and evaluation capabilities of human beings.
The same thesis has already been well argued by the governing bodies of the banking system which, in addition to defining the regulations of the European Banking Authority (EBA) on credit assessment, developed to ensure that financial institutions adopt robust, transparent and consistent processes in assessing credit risk, have defined the guiding principles for the growing use of tools based on artificial intelligence and algorithms. The regulation, in fact, aims to balance technological innovation with the need to preserve credit quality and financial stability.
Key Points of the Regulation EBA
1. Risk-Based Approach: The regulation requires banks to assess credit risk by considering not only traditional financial data but also the qualitative characteristics of companies, ensuring that the analysis takes into account the specificity of each debtor.
2. Transparency and Documentation: Institutions must ensure that assessment models are transparent, documented and regularly audited. This includes the requirement to understand and be able to explain the results produced by AI-based models.
3. Oversight of Artificial Intelligence: The EBA requires rigorous oversight of AI models, ensuring they are used responsibly. Banks must be able to justify credit decisions made by algorithms, avoiding bias or systematic errors.
4. Prudential Assessment: The regulation requires that the use of automated models be integrated with human judgment, especially in cases of derogations or assessments that go beyond standardized situations.
5. Stress Testing Capability: Banks must stress test their credit models, including AI-based ones, to assess resilience under adverse market conditions.
Implications for Banks
6. Strengthening Internal Controls: Banks must strengthen internal controls and governance over credit models, ensuring that risks arising from the use of AI are adequately managed.
7. Requirement for Multidisciplinary Skills: A combination of technical skills and expert judgment is required to correctly interpret model results.
8. Focus on the Uniqueness of Debtors: The regulation encourages an approach that recognizes the uniqueness of individual businesses, overcoming the limitations of standard sector classification.
In short, the legislation also EBA promotes the prudent and responsible use of artificial intelligence in credit assessment, integrating technological innovation with the rigor of human judgment to maintain the integrity of the credit process
Post your feedback on this topic.