Predictive models and the Italian Crisis Code: an empirical validation on Italian SMEs

Predictive models and the Italian Crisis Code: an empirical validation on Italian SMEs



Abstract

This study tests whether a small set of financial ratios can anticipate, twelve months ahead, an Italian SME’s entry into an insolvency procedure, consistent with the forward-looking horizon of Articles 2 and 3 of the Italian Crisis and Insolvency Code (CCII). Three models, logistic regression, decision tree and random forest, were estimated on 25,577 SMEs (201,923 annual filings, 2015-2022) from the AIDA database using five explanatory variables. On a size-matched test set kept separate from training the models reach an AUC-ROC between 0.968 and 0.978. Operating profitability, balance-sheet solidity and asset coverage are the dominant predictors in all three. Altman’s and Bottani’s scores retain high recall but lower precision. Logistic regression is proposed as the preferred operational model on grounds of interpretability and ease of implementation.

1. Introduction

The Italian Crisis and Insolvency Code (CCII, Legislative Decree 14/2019) defines crisis as the debtor’s condition that makes insolvency likely, evidenced by prospective cash flows inadequate to meet obligations over the following twelve months (Art. 2(1)(a)), and requires directors to maintain organisational arrangements able to detect imbalances within the same horizon (Art. 3). The Code prescribes no statistical model, but the forward-looking diagnosis it requires can be supported, especially in smaller companies, by quantitative tools similar to those used in bank credit risk management [1, ch. 1]. Building on the insolvency-prediction literature from Beaver (1966) and Altman (1968) to the Italian contributions of Bottani, Cipriani and Serao (2004) and Altman, Danovi and Falini (2013), the study estimates and validates twelve-month predictive models on Italian SME accounts, compares econometric and machine-learning techniques under severe class imbalance, and assesses their integration into the CCII’s organisational requirements.

2. Data and methodology

Data come from the AIDA database (Bureau van Dijk) for financial years 2015-2022, a window spanning recovery, expansion, the pandemic and the 2022 inflation and rate shock. Filtering 2.3 million companies for age (founded before 2010), EU SME size criteria excluding micro firms, complete accounts and outliers yields 25,577 companies and 201,923 annual observations [1, sec. 3.1]. The dependent variable equals 1 if the company enters an insolvency procedure in year t+1, conditional on year-t accounts: 678 events (3% of companies, 0.34% of observations), of which 527 compositions with creditors (concordato preventivo), 102 debt restructuring agreements and 49 other procedures [1, sec. 3.1.4]. Five explanatory variables are used: retained earnings over total liabilities, personnel cost over production value, net working capital over total assets, ROA, and log total assets.

The training set includes all original observations; class imbalance is addressed on it with SMOTE (k = 5), which generates synthetic observations of the minority class only, giving 402,490 training observations. The test set is built by targeted matching rather than random sampling: each of the 678 companies with an event in the following year is matched with its two nearest non-distressed peers by total assets (2,034 observations), so that models cannot discriminate on size alone. Models are estimated with five-fold cross-validation and AUC-based hyperparameter search on the balanced training set, with standardisation fitted on training folds only to avoid leakage; the best configuration is retrained on the full training set and evaluated on the test set, which is kept separate from training [1, sec. 3.3-3.5].

3. Results

The logistic equation is logit(p) = −3.8334 − 2.8023 · Solidity + 1.3793 · PersonnelCost − 2.5807 · AssetCoverage − 13.2248 · ROA + 0.377 · log(Assets); all coefficients are significant (p < 0.001) with expected signs [1, sec. 4.1.1]. On the test set, AUC-ROC is 0.968 for logistic regression, 0.972 for the decision tree and 0.978 for the random forest; at the F1-optimal threshold, precision and recall are 0.934/0.898, 0.926/0.919 and 0.908/0.950 respectively [1, tab. 25]. A worked example from the thesis defence shows how the equation reads: for a company with solidity 0.078, personnel cost ratio 0.176, asset coverage −0.063, ROA −0.094 and log assets 10.71 from its 2023 accounts, the logit is 1.63 and the estimated probability of an event in the following year is 84%; the company entered composizione negoziata in 2024 and later a composition with creditors [11]. Variable rankings are stable across models: ROA, solidity and asset coverage dominate; in the random forest, solidity accounts for 48% of feature importance and ROA for 30%, confirmed by SHAP values [1, tab. 23, 26]. On the same test set with a uniform threshold, Altman’s Z’ scores 0.99 recall and 0.50 precision, Bottani’s score 0.97 and 0.73, against 0.90-0.94 on both measures for the estimated models [1, tab. 27]; binarising the classic scores is a simplification adopted for comparison only.

4. Discussion

The results answer the research question clearly. Five ratios computable from any filed set of accounts discriminate with high reliability, twelve months ahead, between SMEs that will enter an insolvency procedure and those that will not: AUC-ROC ranges from 0.968 to 0.978 and, at the optimal threshold, the models capture 90-95% of events while keeping false alarms between 3% and 5% of healthy firms [1, ch. 4]. This is achieved on 25,577 companies across eight years that include the pandemic and the 2022 inflation shock, and on a test set deliberately built to be demanding, where each distressed company is compared with peers of the same size; discriminating power therefore does not depend on scale or on a single economic cycle.

The second finding is convergence across techniques. Logistic regression, decision tree and random forest identify the same hierarchy of factors: operating profitability, balance-sheet solidity and asset coverage. That stability is evidence in itself, showing that the signals of deterioration are structural rather than an artefact of method, and provides large-scale quantitative support for what practice and the literature have long indicated qualitatively [1, sec. 6.1].

The third finding concerns the classic scores. On the same test set, Altman’s Z’ and Bottani’s score retain high recall but markedly lower precision (0.50 and 0.73 against 0.90-0.92 for the estimated models), capturing distressed firms at the cost of flagging many healthy ones [1, tab. 27]. Models calibrated on recent Italian data use the available information more efficiently and allow the trade-off between false positives and false negatives to be tuned to the user’s needs.

Operationally, the thesis proposes logistic regression as the working model. The AUC gap to the random forest is about 0.010, while the logistic equation fits on one page, runs in a spreadsheet and can be explained to owners and statutory auditors in terms of each variable’s effect on the probability of distress [1, sec. 4.5, 6.2]. That is a substantive requirement for a tool meant to form part of the organisational arrangements required by Art. 3 CCII and to withstand scrutiny by auditors and lenders. The random forest retains a periodic validation role. For companies the model is a component of internal early-warning arrangements, complementary to the forward-looking cash-flow test; for lenders and investors it identifies the variables to monitor before arrears and reports appear.

5. Scope and further work

The research design reflects explicit choices that define its scope. The 1:2 test-set ratio was chosen to test discriminating power under harsher conditions than real ones; the metrics measure the ability to separate the two groups at equal size, while deployment on a population where events are rare requires calibrating the threshold to the user’s objectives, which the thesis documents for three alternative thresholds [1, sec. 4.1.3, 6.3]. The exclusive use of annual accounting data is a choice of accessibility: the model works on public information available to any company, adviser or lender. The fixed twelve-month horizon mirrors the statutory one. The developments the thesis identifies build on this base: dynamic models treating accounts as trajectories; integration of non-accounting information such as central credit register data and the warning signals the CCII itself lists; and, most significantly, deployment within composizione negoziata, Italy’s out-of-court negotiated settlement procedure, and the Art. 3 arrangements, as part of periodic reporting shared by directors, experts and banks [1, sec. 6.4]. The thesis concludes that the distance between quantitative models and the law of business crisis is smaller than it appears: the Code asks for tools that make timely detection of crisis probable, and models validated on a large sample of Italian SMEs show they can perform that function [1, sec. 6.5].

References

  1. Zorzetto, M. (2025). Modelli predittivi e crisi d’impresa: applicazioni nel Codice della crisi, analisi dei casi di accesso a procedure concorsuali e validazione empirica su PMI italiane. Tesi di Master di II livello in Crisi di impresa e ristrutturazioni aziendali, Università degli Studi di Bergamo, relatore A. Danovi, correlatore S. Birolini, a.a. 2024/2025, versione finale 12 ottobre 2025.
  2. Alberici, A. (1975). Analisi dei bilanci e previsione delle insolvenze. Milano: Isedi.
  3. Altman, E. I. (1968). Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy. Journal of Finance, 23(4), 589-609.
  4. Altman, E. I. (2000). Predicting Financial Distress of Companies: Revisiting the Z-Score and ZETA Models.
  5. Altman, E. I., Danovi, A., Falini, A. (2013). Z-Score Models’ Application to Italian Companies Subject to Extraordinary Administration. Bancaria, n. 4, aprile 2013, 24-37.
  6. Beaver, W. H. (1966). Financial Ratios as Predictors of Failure. Journal of Accounting Research, 4(3), 71-111.
  7. Bottani, E., Cipriani, A., Serao, A. (2004). La previsione delle insolvenze aziendali: un’applicazione del modello Z-Score alle PMI italiane.
  8. Chawla, N. V. et al. (2002). SMOTE: Synthetic Minority Over-sampling Technique.
  9. Ohlson, J. A. (1980). Financial Ratios and the Probabilistic Prediction of Bankruptcy. Journal of Accounting Research, 18(1), 109-131.
  10. Zmijewski, M. E. (1984). Methodological Issues Related to the Estimation of Financial Distress Prediction Models. Journal of Accounting Research, 22 (suppl.), 59-82.
  11. Zorzetto, M. (2025). Presentazione della tesi, Master di II livello in Crisi di impresa e ristrutturazioni aziendali, Università degli Studi di Bergamo, 18 dicembre 2025, slide 8.

Edited by:
Marco Zorzetto




20/11/2026