Artificial Intelligence for Predicting Insolvency in the Construction Industry—A Systematic Review and Empirical Feature Derivation

Abstract

The construction sector is particularly prone to financial instability, with insolvencies occurring more frequently among micro- and small-scale firms. The current study explores the application of artificial intelligence (AI) and machine learning (ML) models for predicting insolvency within this sector. The research combined a structured literature review with empirical analysis of construction sector-level insolvency data spanning the recent decade. A critical review of studies highlighted a clear shift from traditional statistical methods to AI/ML-driven approaches, with ensemble learning, neural networks, and hybrid learning models demonstrating superior predictive accuracy and robustness. While current predictive models mostly rely on financial ratio-based inputs, this research complements this foundation by introducing additional sector-specific variables. Empirical analysis reveals persistent patterns of distress, with micro- and small-sized construction businesses accounting for approximately 92% to 96% of insolvency cases each year in the Australian construction sector. Key risk signals such as firm size, cash flow risks, governance breaches and capital adequacy issues were translated into practical features that may enhance the predictive sensitivity of the existing models. The study also emphasises the need for digital self-assessment tools to support micro- and small-scale contractors in evaluating their financial health. By transforming predictive insights into accessible, real-time evaluations, such tools can facilitate early interventions and reduce the risk of insolvency among vulnerable construction firms. The current study combines insights from the review of AI/ML insolvency prediction models with sector-specific feature derivation, potentially providing a foundation for future research and practical adaptation in the construction context.

1. Introduction

The construction industry is one of the most financially vulnerable sectors in the global economy. Characterised by project-based revenue, high operational leverage, and subcontracting complexities, firms, especially micro- and small-scale contractors, face elevated risks of business failure [1,2,3]. Construction projects are often capital-intensive, irregular, and delayed due to factors such as weather, regulatory approvals, or client-side disruptions [4,5]. The industry also operates on thin profit margins, and contractors frequently rely on external debt or delayed client payments to manage cash flow [6,7]. The tiered supply chain structure of the construction sector, where subcontractors are financially reliant on upstream principal contractors, creates a fragile ecosystem in which payment disruptions at any level can trigger cascading defaults throughout the project network [8,9]. In many jurisdictions, including Australia and the UK, the sector also struggles with high rates of insolvency relative to other industries [10,11]. The aftermath of COVID-19 and recent macroeconomic shocks, including material cost inflation, interest rate hikes, and skilled labour shortages, have further amplified these structural weaknesses [12,13,14]. Yet, despite the prevalence of business failure in the sector, early detection remains challenging, and predictive mechanisms are often underutilised or inapplicable to sector-specific nuances [15].

Predicting insolvency has been an area of sustained research, beginning with early financial ratio models such as Altman’s Z-score, and evolving into more sophisticated multivariate and AI-based approaches [15,16]. Recent research shows that while these traditional models remain valuable, combining them with advanced statistical and machine learning techniques can significantly improve prediction accuracy [17,18]. Furthermore, recent studies increasingly advocate for AI and ML approaches due to their ability to handle non-linear relationships, process large volumes of data, and adapt to complex variable interactions [19,20]. However, most existing models rely heavily on generic financial ratios and lack the ability to capture governance failures, compliance lapses, or behavioural risks specific to construction operations.

There is a growing academic emphasis on developing context-sensitive models for insolvency prediction aligned to reflect the operational realities of specific sectors such as construction. This is especially relevant given that many small construction firms operate with limited financial disclosures and informal risk management structures, making conventional ratio-based models inadequate [21]. Alongside academic developments, regulatory bodies and industry insurers have also shown growing interest in leveraging predictive analytics for early risk identification. However, the translation of these insights into actionable tools for at-risk firms remains limited, particularly for micro and small contractors, who often lack access to specialised financial advisory services or benchmarking instruments. As a result, these firms are frequently left to rely solely on their internal accounting systems, basic ledgers, or the advice of internal and external accountants, which may not provide the depth of insight needed to detect early signs of financial distress. Recent studies emphasise the value of simplified early warning systems that can function with minimal data input, making them suitable for smaller construction firms that lack formal accounting infrastructure [22].

This study integrates three interconnected components, which together represent its key scholarly contributions. First, it presents a structured synthesis of AI/ML-based insolvency prediction models relevant to the construction sector (Section 2 and Section 4). Second, it offers critical insights on methodological trends, performance advantages, and sector-specific modelling gaps identified from prior studies (Section 4.2). Thirdly, it provides sector-informed feature guidance through the empirical derivation of construction-specific predictive features from Australian insolvency data covering the period 2014–2024 (Section 5). Collectively, these contributions form a cohesive analysis that links modelling approaches with contextually grounded, practical recommendations. To deliver these components, the present study employs a dual-method approach comprising a structured literature review and empirical analysis. It first undertakes a structured critical review to synthesise prevailing AI/ML approaches in insolvency prediction in the construction sector. Then, using construction insolvency data spanning a recent decade, it derives sector-specific empirical features that may enhance model sensitivity. This research attempts to support the advancement of data-driven financial risk management across the construction sector, with practical value for industry stakeholders by providing sector-specific early warning indicators and actionable insights to help micro and small firms, the most vulnerable segment in mitigating insolvency risks.

2. Background: Predictive Models in Financial Bankruptcy and Insolvency

Within the financial risk literature, terms such as insolvency, bankruptcy, and financial distress are often used interchangeably, though they may denote distinct stages of corporate decline. Financial distress typically signifies a state where a firm struggles to meet its obligations but has not yet defaulted [23]. Insolvency denotes a more formal condition where liabilities exceed assets or where a firm is unable to meet its debts in the normal course of business [24,25]. Bankruptcy, on the other hand, is a legal process triggered by insolvency, often used strategically to reorganise or liquidate operations [26]. Platt and Platt (2006) [27] argued that financial distress is not simply a precursor to bankruptcy but may involve different explanatory variables, requiring separate predictive considerations. Thus, this distinction is particularly relevant in construction, where firms may operate under prolonged financial distress without formally declaring insolvency or entering legal bankruptcy proceedings.

The evolution of predictive models from Altman’s Z-score to modern AI represents a significant transformation in financial forecasting, driven by advancements in methodology, computational power, and data availability. Figure 1 illustrates the critical shift points in this evolution, focusing on key developments that marked transitions in predictive modelling techniques. Each shift point reflects a change in approach and technology, contributing to the increased accuracy and applicability of predictive models in finance.

2.1. Altman’s Z Score

Edward Altman’s multiple discriminant analysis model (MDA), also known as the Z-score, introduced in 1968, was a pioneering statistical model for predicting corporate bankruptcy [28]. It combined five financial ratios (Working capital/Total assets, Retained earnings/Total assets, Earnings before interest and taxes/Total assets, Market value of equity/Book value of total liabilities, Sales/Total assets) into a single score using MDA [29]. This model achieved 94% accuracy for one-year predictions and 72% for two-year predictions, and Altman also included a table in his 1968 paper illustrating how the model’s accuracy declines beyond the initial two-year horizon [30]. The Z-Score categorises firms into “safe,” “grey,” or “distress” zones, offering early warning signals for financial instability and aiding investors, analysts, and decision-makers in risk assessment [31]. Studies have shown that the initial Z-score model may not fully capture the risk of bankruptcy in sectors such as non-manufacturing, banking or in emerging markets, where financial structures and accounting standards differ, necessitating local adjustments for reliable predictions [32]. The model also relies solely on historical financial ratios, ignoring qualitative factors such as management quality, market conditions, or macroeconomic shocks, which can be critical in assessing a firm’s true risk [28,33]. Despite these limitations, the Z-Score remains a valuable early warning tool, especially when adapted to specific contexts and supplemented with additional risk metrics.

2.2. Advanced Statistical Models

The 1980s saw the introduction of more sophisticated statistical methods, such as Ohlson O-score and logistic regression (Logit) and probit models. An alternative to the Z-score, the Ohlson O-score model utilises nine financial ratios to predict bankruptcy, offering a different perspective on credit risk assessment [34]. The O-Score has also been applied to assess the impact of ESG (Environmental, Social, Governance) factors on financial stability, with findings suggesting that strong governance practices are linked to lower bankruptcy risk [35]. In emerging markets and specific sectors, such as small and medium enterprises or insurance companies, the O-Score has demonstrated reasonable predictive power, though its accuracy may be limited by the quality of financial reporting [36,37]. Modifications to the O-Score, such as adjusting cut-off values or incorporating local institutional factors, can further enhance its predictive performance in different contexts [38].

Logit and probit models estimated the probability of default by analysing the relationship between a binary dependent variable and one or more independent variables, where the logit model uses the logistic function, while the probit model uses the cumulative normal distribution function [39]. Both models produce similar results overall, but probit tends to perform better with small, normally distributed samples, while logit may be more effective with larger or non-normally distributed data [40,41]. Both models are preferable to linear probability models, which can yield unrealistic predictions and violate key statistical assumptions [42,43].

2.3. Neural Networks

Neural networks (NNs) were first applied to financial predictive modelling in 1990, marking the initial shift toward AI-based approaches [44]. This development introduced ML to finance, enabling models to handle more complex patterns. However, early NNs were constrained by limited computational resources. NNs learn by adjusting the weights of connections between neurons, a process analogous to how biological brains adapt through synaptic changes [45]. It has become a prominent tool in bankruptcy prediction, consistently outperforming traditional statistical methods such as discriminant analysis and logistic regression in terms of predictive accuracy and robustness to sampling variations [46]. Odom and Sharda [47,48] explored NNs for bankruptcy prediction, laying the groundwork for future AI applications.

2.4. Advanced ML Models

With advancements in computing power and data availability, advanced ML techniques such as Random Forests (RF) and Support Vector Machines (SVM) began to influence predictive modelling in the early 2000s. Studies consistently show that RF tends to deliver higher or comparable accuracy, with some reporting RF achieving up to 98% accuracy, especially when a broad set of financial indicators is included, while SVM typically achieves slightly lower but still strong results, often in the 90–92% range [49,50,51,52]. RF is praised for its operational simplicity and speed, making it attractive for practical applications, whereas SVM can offer higher classification accuracy in certain contexts, particularly with smaller or more complex datasets [53,54,55].

2.5. Rise of Deep Learning and Advanced AI Models

In the 2010s, deep learning experienced a rapid rise in finance, driven by its superior performance over traditional ML models in tasks such as financial time series forecasting, stock price prediction, and sentiment analysis [56,57]. Deep learning models such as Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), Natural Language Processing (NLP), and Long Short-Term Memory (LSTMs) were increasingly adopted, showing improved performance, especially with large and complex datasets [58,59]. DNNs, with multiple hidden layers, enhanced the ability to model complex, non-linear relationships in financial data [60]. NLP models, including transformers, enabled the analysis of unstructured data such as news articles and earnings calls, providing additional insights into financial health [61,62].

Since 2017, predictive modelling has continued to evolve with hybrid models, such as CNN-LSTM and CNN-LSTM with an Attention Mechanism (AM), which outperform standalone models [63,64]. Recent research shows that graph neural networks (GNNs) are a leading approach for capturing the structure and features of financial and supply chain networks, enabling significantly better systemic risk prediction than traditional machine learning [65,66]. Studies report up to a 94% improvement in key metrics (e.g., Matthew’s Correlation Coefficient) when both network structure and edge features are incorporated [67]. Transformer-based NLP models such as Bidirectional Encoder Representations from Transformers (BERT) are highly effective for extracting risk indicators from unstructured textual data, outperforming traditional methods in accuracy and scalability. A study showed that in construction risk assessment, BERT-based entity recognition engines efficiently extracted predefined risk knowledge from unstructured safety data, enabling the creation of robust knowledge bases for risk management [68].

Furthermore, explainable AI (XAI) frameworks such as Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME) are increasingly used in AI insolvency models to enhance interpretability and stakeholder trust [69]. They are primarily used to interpret black-box models such as NNs, XGBoost, and RFs, which are common in insolvency and credit risk prediction tasks. These frameworks provide feature importance and local explanations, helping users understand which factors drive model decisions [70,71]. This period reflects ongoing innovation, with models addressing challenges such as class imbalance (utilising techniques like Synthetic Minority Over-sampling Technique-SMOTE) and feature extraction, which were significant challenges in bankruptcy datasets [72]. Hybrid models combine spatial and temporal data processing for superior performance [73].

3. Research Methodology

This study employs a dual-method research strategy (see Figure 2) comprising a structured and thematically refined literature review on predictive insolvency modelling in construction and an empirical analysis of Australian construction insolvency data to derive sector-specific features that can inform future model development.

3.1. The Literature Review Process

The literature review was initiated with a broad, systematic search using Scopus to map the landscape of insolvency risks in the global construction industry. The search query was designed to capture a comprehensive array of studies by combining sector-specific terms (e.g., “construction”, “contractor”, “residential sector”) with insolvency-related concepts (e.g., “bankruptcy”, “financial distress”, “insolvency”) and a range of risk factor terms (e.g., “phoenix activity”, “regulatory reform*”, “early warn*”, “predictor*”). This Boolean logic-based search yielded 466 records. After removing duplications, 461 unique records remained and were manually screened by title and abstract, focusing on English-language articles published between 2009 and 2024 that were relevant to the construction context. This process reduced the dataset to 178 relevant studies. To mitigate the limitation of relying on a single primary database, a snowball sampling approach was used, where the reference lists of relevant studies were screened to identify additional eligible sources. This process yielded 17 further studies, increasing the comprehensiveness of the final pool to 195 studies.

To further refine this study pool, topic modelling techniques were employed, revealing five dominant thematic clusters within the relevant literature. Among these, Theme 3 (financial modelling and bankruptcy analysis) was identified as the most aligned with the focus of this study. This theme covered a range of methodological approaches, including classical statistical models, AI and machine learning applications, and hybrid frameworks for insolvency prediction in the construction sector. A refined subset of 27 studies from this theme was selected for focused analysis. Each study involving AI/ML was evaluated for its methodological design, model architecture, feature engineering techniques (e.g., recursive feature elimination with cross-validation, expert screening), and validation metrics (e.g., accuracy, area under the curve, classification rates). This rigorous, tiered approach to the literature review ensured both thematic breadth and analytical depth.

3.2. Empirical Feature Derivation from Sector-Level Construction Insolvency Data

In parallel, the study conducted an empirical investigation using insolvency statistics from the Australian Securities and Investments Commission (ASIC) for the period 2014–2024. ASIC publishes several statistical series offering insights into the level and nature of corporate insolvencies in the construction sector [74]. All data used in this analysis were downloaded in Excel format from the official website of ASIC. The analysis began with the identification of broad empirical patterns in construction insolvency across the ten-year period. These patterns were examined across multiple dimensions, including firm size, primary causes of failure, types of director or officer misconduct, and financial deficiency at the point of collapse. This exploratory phase aimed at understanding recurring structural, operational, and governance-related vulnerabilities that may precede insolvency within the construction industry. This step provided a contextual baseline and sector-specific insights to inform subsequent model design. Building on these empirical observations, the study proceeded to derive a set of potential predictive features suitable for integration into insolvency risk models as supplements. This feature derivation process ensured that the resulting model inputs were not only empirically grounded but also shaped to the operational and financial realities of the construction sector. The outcome is a feature set designed to enhance the accuracy and contextual sensitivity of AI-based insolvency prediction models.

4. Analysis of Predictive Models for Bankruptcy in Construction

4.1. Overview of Predictive Model Studies

A review of 27 key research studies (see

Table 1. Research studies on bankruptcy prediction and financial distress in construction.

Study	Purpose	Model/Technique	Type	Country
[75]	Evaluate financial distress using logistic regression and financial ratios	Altman, Springate, Grover, Zmijewski with Logistic Regression	Statistical	Malaysia
[92]	Compare the Altman, Ohlson, and Zmijewski models in construction bankruptcy	Altman, Ohlson, Zmijewski	Statistical	Greece
[99]	Assess the bankruptcy risk of state-owned construction firms	Altman Z-score	Statistical	Vietnam
[93]	Compare the financial risk/efficiency of state-owned enterprises vs. private firms	Financial ratio analysis, DEA	Statistical	Indonesia
[76]	Predict distress using ensemble learning and non-financial variables	Ensemble Learning (Soft Voting), SMOTETomek, Recursive Feature Elimination with Cross-Validation (RFECV)	AI/ML	China
[98]	Analyse bankruptcy reasons in Slovak construction small and medium-sized enterprises (SMEs)	Descriptive statistics, Regional stratification	Statistical	Slovakia
[94]	Compare ML models for long-term distress in construction	Random Subspace (RS), ML, Ensemble Models	AI/ML	South Korea
[77]	Develop an insolvency model for creditworthiness in construction	Discriminant Analysis, Logistic Regression, Classification Trees	Statistical	Poland
[95]	Predict the business failure of construction contractors using financial, market, and macroeconomic data	LSTM RNN	AI/ML	US
[78]	Estimate bankruptcy probability in construction companies	Logit regression	Statistical	Russia
[79]	Compare methods for predicting construction company bankruptcy using financial and non-financial data	Logistic/probit regression, Classification trees, RF, NNs	AI/ML	Russia
[80]	Integrate SVM and market-based models to forecast contractor default in Taiwan	SVM, Artificial Neural Network, Hybrid models	AI/ML	Taiwan
[81]	Assess the stability of bankruptcy predictors over time for the Czech construction and manufacturing sectors	Univariate (T-test, F-test) and multivariate methods (Boosted trees)	AI/ML	Czech Republic
[82]	Evaluate dynamic indicators for bankruptcy prediction in Czech construction firms	Dynamic financial indicators, statistical testing	Statistical	Czech Republic
[83]	Predict financial loss in building projects using insurance claim data	Multiple regression analysis	Statistical	South Korea
[96]	Measure and analyse construction firm efficiency pre-/post-financial crisis	DEA, Bootstrap regression	Statistical	Spain
[87]	Predict construction firm default using financial ratios and market factors	Logit regression	Statistical	Taiwan
[100]	Predict financial loss for large construction companies	Logistic regression	Statistical	US
[86]	Predict the financial distress of Chinese construction and real estate companies	Back propagation NN (BPNN), AdaBoost, Bagging ensembles	AI/ML	China
[85]	Predict the bankruptcy of construction companies using financial ratios	Fisher’s Linear Discriminant Analysis	Statistical	China
[84]	Predict the bankruptcy probability of construction companies with business risk indicators	Fuzzy logic model, Decision-support tool	Fuzzy logic	US
[101]	Compare financial trends of construction firms using bankruptcy prediction models	Altman Z-score (multivariate discriminant analysis)	Statistical	Korea, Japan, US
[88]	Identify and rank critical factors for predicting insolvency in construction firms	Systematic review, Statistical analysis	Statistical/Review	UK
[91]	Develop a bankruptcy prediction model for construction companies using sector-specific data	Classification and Regression Trees (CART)	AI/ML	Czech Republic
[97]	Review and critique methodological approaches in construction business failure prediction studies	Systematic review, meta-analysis	Review	Global
[89]	Improve construction business failure prediction using entropy-based discriminant analysis	Entropy measure, Discriminant analysis	Statistical	Taiwan
[90]	Predict insolvency using Bidirectional LSTM (BiLSTM) autoencoder and financial/macro indicators	BiLSTM Autoencoder	AI/ML	Australia

) demonstrates the evolution and diversity of approaches to predicting bankruptcy and financial distress within the construction industry. Most prominently, the majority of the works are dedicated to the direct prediction of bankruptcy, insolvency or financial distress among construction firms [75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91]. Other research focused on the comparative analysis or development of new predictive models focused specifically on the construction or real estate sector [92,93,94,95,96,97]. Furthermore, several investigations analysed the impact of macroeconomic factors and financial crises on sectoral stability [98,99,100,101], while some works explore the influence of managerial, market, or technical efficiency dimensions on the financial risks of the firms [82,83,87].

The methodologies employed in these studies reflected both historical and contemporary advances in predictive analytics. Statistical models such as logistic regression, discriminant analysis, and probit analysis remain foundational due to their interpretability and regulatory familiarity [102,103,104,105]. Classic bankruptcy prediction models, notably Altman’s Z-score, Ohlson O-score, and Zmijewski, are frequently used as benchmarks [75,92,99]. Data Envelopment Analysis (DEA) was used in research examining technical efficiency and operational risk [93,96]. Qualitative methodologies combined with statistical models, including systematic reviews and expert opinion, provide conceptual syntheses and meta-analytical perspectives [88,97,98]. While statistical methods continue to dominate, there is a shift toward the integration of AI and ML techniques in recent research work. These include the application of NNs ensemble learning methods, SVMs, RFs, and deep learning architectures, offering enhanced accuracy, particularly when handling large or unbalanced datasets or incorporating non-financial predictors. This trend is addressed in detail in Section 4.2.

The studies reviewed provide a global perspective, with research conducted in diverse national contexts, including China, Taiwan, South Korea, Vietnam, Malaysia, Indonesia, the Czech Republic, Russia, Spain, Poland, Slovakia, Greece, the UK, Australia, the US, Egypt, the United Arab Emirates, and Uganda. This diversity not only emphasises the universal nature of financial distress challenges in construction but also highlights the necessity of context-sensitive modelling to accommodate varying financial systems, accounting standards, and market environments.

Key trends emerging from this analysis show that early studies primarily adapted generic bankruptcy prediction models from manufacturing and general corporate finance, often without accounting for construction-specific risks. Meanwhile, more recent research emphasises the development of sector-specific models, acknowledging the unique financial structures and operational characteristics of construction firms, such as project-based revenue, cyclicality, and high leverage. Additionally, the adoption of AI/ML approaches is accelerating, often resulting in superior predictive performance and the ability to synthesise a wider array of indicators.

4.2. AI/ML-Based Predictive Model Studies

A focused analysis of the AI/ML-driven studies (see

Table 2. Research studies using AI/ML for bankruptcy prediction in construction.

Study	AI/ML Function	Data Type/Scope	Variable Selection	Model Evaluation	Key Results
[76]	Predict financial distress 1 to 2 years in advance	Chinese listed construction firms, Financial and non-financial variables	RFECV	Compared against single classifiers	Soft voting ensemble model outperformed all single models
[94]	Predict medium-to-long-term (3, 5, 7 years) financial distress	Korean construction firms, multiple-year forecasts	Used financial ratios relevant for long-term predictions	AUC, Friedman test for model comparison	The RS model achieved the best performance over a medium-to-long-term horizon
[95]	Predict business failure 1–3 years in advance	US construction contractors	Compared models using only accounting, only market, only macro, and all combined	Compared the LSTM RNN performance for variable sets and prediction windows	Adding market and macro variables increased prediction accuracy by 2–4% over using accounting alone
[79]	Predict bankruptcy within 1 year	Russian construction firms, Public financial and non-financial data (2011–2017)	Included financial ratios (profitability, liquidity, stability, activity) and non-financial factors (firm size, age)	Compared logistic/probit regression, classification trees, RF, and NN, Used AUC/Gini as quality metrics	Neural networks had the highest predictive power, and Logistic regression with discretisation also performed well
[80]	Forecast contractor defaults using hybrid AI	Taiwanese construction firms, Financial and stock data	Integrated from the literature and expert screening	Accuracy and model comparison	Hybrid models outperformed individual algorithms
[81]	Explore predictor stability over time and sectors	Czech data (2003 to 2013), 34,533 firms comprise	Statistical significance testing	Cross-sectoral and temporal model robustness	Predictor importance varies by sector and time frame
[86]	Predict financial distress 1 to 3 years in advance	Chinese listed construction and real estate companies	Three financial ratio datasets were constructed from public stock exchange data	Compared BPNN-AdaBoost and BPNN-Bagging ensembles to a single BP neural network and Z3-score model	Both ensemble models outperformed single BPNN and Z3-score, AdaBoost best for 1–2 years, Bagging best for 3-year prediction
[91]	Predict the bankruptcy of construction companies	Czech construction firms (period: 2011 to 2014), 29 financial indicators	Selection based on relevance from accounting/financial data	Evaluated using correct classification rate, Type I/II errors, Receiver Operating Characteristic (ROC) curve (AUC)	The CART model for construction firms had the highest discrimination ability and outperformed generic models
[90]	Predict insolvency using reconstruction error	Australian construction firms (2000 to 2020), 180 records post-balancing	17 financial, operational, growth, and macroeconomic indicators	Accuracy, Precision, Recall, F1-score	Achieved 97.3% accuracy, outperforming CNN-BiLSTM and LSTM models

) reveals the rapidly expanding role of AI and ML in the prediction of bankruptcy and financial distress within the construction industry. These models predict insolvency by processing financial, operational, and contextual data through algorithms capable of identifying complex, non-linear patterns linked to business failure. In this review, each study is examined in terms of its data inputs, feature selection process, modelling technique, and evaluation metrics, enabling a clear understanding of the predictive process. To further clarify how AI operates in this context, Figure 3 illustrates a typical AI/ML workflow for insolvency prediction, beginning with data acquisition, followed by feature engineering and model training, and leading to prediction and evaluation using performance metrics and XAI tools. This diagram provides a concise visual summary of the process, complementing the detailed discussion presented in this section.

These works span a variety of ML techniques, including classical NNs, ensemble learning, SVM, RFs, and, most recently, deep learning architectures such as LSTM and BiLSTM autoencoders. The AI/ML functions explored across studies are primarily aimed at forecasting financial distress or bankruptcy risk one or more years in advance, by enabling early intervention for at-risk firms (for example, [76,79,86,95]). AI/ML models enable earlier and more reliable identification of at-risk firms, supporting proactive risk management for investors, regulators, and company management [106,107]. The data types and scope are diverse, ranging from financial statement data of national construction sectors (e.g., China, Korea, Australia, the Czech Republic, Russia) to the inclusion of market, macroeconomic, and non-financial indicators. Notably, one study showed that incorporating both financial and non-financial variables yields marginally higher predictive accuracy; here, the one-year area under the curve (AUC) increased from 0.937 to 0.946, a 0.9% improvement over models using only financial data [76]. Moreover, several studies employed large, multi-year datasets, sometimes with public and private firm data, to improve robustness and generalisability [81,90].

Variable selection is a crucial step in AI/ML, aiming to improve model performance, interpretability, and efficiency [108]. There are several strategies, each with strengths and trade-offs. No single method is universally best, and the optimal approach depends on data size, structure, and the specific application. In this analysis, some studies apply RFE, or expert-based selection, to identify the most informative predictors. Others test combinations of accounting, market, and macroeconomic features to assess incremental predictive value. RFE typically starts with all features and repeatedly builds a model, ranks features by importance and removes the least important ones in each iteration. This process continues until the desired number of features is reached or performance stops improving [109,110]. Model evaluation is rigorous, with most studies benchmarking their AI/ML approaches against traditional statistical models (such as logistic regression or discriminant analysis) and reporting on metrics such as accuracy, AUC, and classification rates [79,91,94]. AUC measures the ability of a classifier to rank positive instances higher than negative ones across all possible thresholds. It is especially valuable for imbalanced datasets, as it is less affected by class prevalence and provides a consistent evaluation across different scenarios [111,112,113]. Classification rates represent the percentage of correct predictions. While easy to interpret, accuracy can be misleading in imbalanced datasets, as it may remain high even if the model fails to identify minority class instances [114,115].

The outcomes show that the AI/ML models (particularly ensemble and deep learning models) consistently outperform classical methods. Notably, Jeong and Kim (2022) [94] showed that the random RS ensemble model achieved the best performance for medium-to-long-term predictions (3–7 years) in Korean construction firms, with AUCs remaining above 0.7 even for seven-year horizons, substantially higher than logistic regression and traditional models. Multiple studies report improvements of several percentage points in predictive accuracy when using soft voting ensembles, LSTM networks, or boosted neural networks compared to single-model or traditional statistical baselines. For example, in a recent study on Chinese construction firms, a soft voting ensemble model achieved an AUC of 0.945 for one-year-ahead predictions, outperforming all single classifiers, while the inclusion of non-financial variables further boosted predictive performance [76]. Further, Jang et al. (2020) [95] reported LSTM RNN models improving long-term (three-year) failure prediction accuracy by 2–4% when market and macroeconomic features were added. The enhanced predictive power is especially evident in settings that require addressing class imbalance, integrating diverse variable sets, or capturing non-linear relationships within the data. Hybrid model architecture and advanced deep learning frameworks have demonstrated the capacity to synthesise both temporal and cross-sectional information, further strengthening early detection capabilities. The AI/ML-based predictive models in the key selected studies highlight the transformative potential of intelligent modelling approaches for identifying financially vulnerable construction firms. The transition toward more sophisticated, context-aware AI models signifies a pivotal advancement for risk assessment and early warning systems in the construction industry, offering substantial benefits for financial stability and proactive risk management.

5. Empirical Analysis and Feature Derivation from Australian Construction Insolvency Data

This section presents an empirical case analysis of construction sector insolvency in Australia, drawing on detailed data from the ASIC over the past decade (2014–2024) [74]. In this analysis, insolvency refers specifically to ASIC-reported financial insolvency events in the construction sector, in line with statutory definitions under the Australian Corporations Act 2001 [116]. The analysis provides contextual evidence on patterns and drivers of insolvency, which are then used to inform the selection and development of potential predictive features for AI-based risk modelling.

5.1. Patterns and Trends in Australian Construction Insolvency (2014–2024)

Over the decade from 2014 to 2024, the annual number of construction insolvencies in Australia exhibited substantial volatility (see Figure 4a), reflecting both sector-specific challenges and broader economic disruptions. Insolvency cases ranged from a low of 919 in 2021–2022 to a high of 1964 in 2015–2016. The years immediately preceding the COVID-19 pandemic (2014–2019) were marked by consistently high insolvency counts, averaging approximately 1700 cases per year. However, the pandemic period saw a sharp and unprecedented decline, with cases falling to 953 in 2020–2021 and reaching their lowest point at 919 in 2021–2022, likely due to temporary government support measures, industry hibernation, or delayed insolvency impact [117]. This trend quickly reversed as support measures ceased, insolvencies surged to 1541 in 2022–2023 and nearly doubled to 1942 in 2023–2024, approaching the previous peak levels of the decade. This dynamic pattern emphasises heightened sensitivity of the construction sector to both macroeconomic cycles and policy interventions, highlighting the importance of incorporating temporal trends and external shocks into insolvency risk modelling.

A consistent and striking trend throughout the 2014–2024 period is the major concentration of insolvencies among micro and small construction firms (see Figure 4b), as measured by full-time equivalent (FTE) employees. On average, micro-sized companies with fewer than 5 FTEs accounted for approximately 77% to 80% of annual insolvency cases, peaking at 80.2% in 2020–2021. Small firms (5–19 FTEs) comprised the next largest group, typically representing between 14% and 18% of insolvencies each year. In contrast, medium-sized firms (20–199 FTEs) made up only 4% to 7% of cases annually, and large firms (200+ FTEs) were almost absent from insolvency statistics, consistently accounting for less than 0.2% of cases across all years. This persistent pattern reveals the acute vulnerability of micro and small enterprises in the construction sector, as groups with limited financial buffers and risk management capacity are disproportionately affected, while insolvency among medium and large firms remains comparatively lower.

Analysis of the leading causes of construction insolvency from 2014 to 2024 highlights a persistent set of internal, operational risk factors (Figure 5). Throughout the decade, inadequate cash flow or high cash use consistently ranked as the most common cause, accounting for approximately 16% to 20% of cases each year (peaking at 20.2% in 2019–2020). Poor strategic management was the next most cited cause, representing 13% to 18% of insolvencies annually, while poor financial control, including lack of records, was present in 11% to 15% of cases. Trading losses remained a significant contributor, comprising roughly 13% to 15% of insolvencies each year. These figures demonstrate that financial mismanagement, weak strategic direction, and recurring cash flow issues have remained the predominant drivers of failure in the sector, with their relative importance showing only modest fluctuation over the observed period.

Due to a change in ASIC data reporting, detailed information on possible misconduct is analysed only for the most recent four years (2020–2024), as shown in Figure 6. Across this period, insolvent trading (Section 588G) was the most frequently identified misconduct, cited in 32.5% to 36.0% of cases annually. Breaches of directors’ duties (Sections 180–184) were nearly as prevalent, accounting for roughly 29.2% to 31.2% of cases each year. Failures relating to keeping proper financial records (Sections 286 and 344) comprised around 16.5% to 17.4% of misconduct incidents. Other notable, though less common, misconducts included failure to provide company books to the liquidator (around 5%), officers not assisting the liquidator (4–5.5%), and reporting deficiencies or offences by officers and employees (1–4%). Categories such as agreements to avoid employee entitlements or offences under other legislation were rarely reported, typically making up less than 1% of cases. These findings reinforce the significance of both behavioural and compliance-related risks, emphasising that governance failures, particularly insolvent trading and breaches of duty, remain central to the pattern of construction insolvency.

A detailed examination of net deficiency (see Figure 7), the gap between liabilities and assets at the point of insolvency, reveals the considerable scale of financial distress among failed construction firms. Over the decade, most insolvencies occurred in the lower to mid-deficiency brackets. For example, in 2023–2024, 444 companies (nearly 23% of cases) had deficiencies between $50,001 and $250,000, while 327 firms (around 17%) were in the $250,001 to $499,999 range. Notably, a substantial portion of companies each year exhibited even larger shortfalls as seen by 517 cases (26.5%) in 2023–2024, involving deficiencies between $1 million and $5 million, and the number of firms with a deficiency exceeding $10 million has grown steadily, reaching 122 in the latest year, an all-time high for the decade. While the smallest bracket (<$50,000) accounted for only 9% of cases in 2023–2024, higher-value brackets have generally seen increasing counts since 2020. This pattern highlights that insolvency in the sector is often preceded by the accumulation of significant financial shortfalls. These trends emphasise the importance of continuous monitoring and early intervention based on changes in net deficiency as a leading indicator of severe financial stress.

5.2. Variable Derivation and Model Feature Framework

Building on the empirical findings of Section 5.1, this section identifies a set of potential supplementary features, derived from ASIC construction insolvency data, which may enhance the predictive power of established insolvency risk models. These variables are not intended to replace well-validated predictors in the literature, but to serve as practical additions where data is available and relevant. Further, these features are intended as a flexible framework for future model development. They are not prescriptive and should select, adapt, or supplement features based on their specific data, context, and research objectives. The analysis of Australian construction insolvency highlights several company, financial, and behavioural factors that could be valuable supplements to core predictors such as financial ratios, liquidity measures, profitability, and firm age. For instance, the acute vulnerability of micro and small firms, recurring issues of financial control and capital adequacy, patterns of director or officer misconduct, and net deficiency at the point of collapse all provide contextually rich, sector-specific signals.

Table 3. Potential supplementary features for insolvency predictive models.

Empirical Pattern	Potential Supplementary Feature	Definition/Rationale
High risk among micro/small firms	Company size	Number of employees, higher risk for micro/small firms
Poor financial control and record-keeping	Financial control flag	An indicator of management quality, supplementing ratio-based analysis
Under-capitalisation	Capital adequacy flag	Adds qualitative depth to capital structure metrics
Chronic cash flow issues	Liquidity/cash flow issues	Enhances standard liquidity ratios with behavioural evidence
Trading losses	Trading losses flag	Flag recent performance trends that are not always visible in annual ratios
Company or officer misconduct	Misconduct indicator	Provides behavioural context to quantitative analysis
Persistent negative equity	Net deficiency bracket	Categorical supplement to net worth calculation
Large debts to creditors	Outstanding debt	Context for leverage or creditor concentration
Employee/tax arrears	Compliance arrears flag	Captures regulatory and late-stage distress signals
Number of unsecured creditors	Number of unsecured creditors	Enhances the detail of the creditor risk assessment

summarises how these empirically observed variables can be mapped to standardised, industry-neutral features, ready for integration alongside existing model inputs.

Beyond these principal features, the ASIC dataset allows for the inclusion of further supplementary variables, such as unpaid employee entitlements, outstanding taxes, amount and distribution of secured and unsecured debts, estimated dividend to creditors, and debts incurred post-insolvency [74]. When available, these variables can further refine risk assessment models, especially for late-stage distress detection or for supporting early warning frameworks. These sector-informed variables are presented as optional supplements, meant to enhance, rather than replace, the core predictors already widely used in insolvency risk modelling. Their inclusion should be guided by data availability, model purpose, and sector context. This section supports future research, policy, and industry practice in developing more sensitive and interpretable models for construction insolvency risk.

6. Conclusions, Limitations, and Future Agenda

This study explored the predictive landscape of bankruptcy and financial distress in the construction sector, with a particular focus on the role of AI and ML methodologies. Employing a dual-method approach, the study conducted a thematically refined literature review alongside an empirical analysis of official insolvency data from the ASIC. The review revealed a growing preference for advanced AI/ML models, particularly ensemble learning and deep learning techniques, due to their capacity to handle complex and high-dimensional financial data. Complementing this, the empirical analysis highlighted recurring sector-specific vulnerabilities, especially among micro- and small firms, which inform both model development and feature engineering. From this integrated approach, a set of potential predictive features was derived, drawing on quantitative, operational, and behavioural indicators observed in ASIC insolvency records. These variables include company size, financial management quality, compliance breaches, and liquidity stress markers. Together, they offer a robust foundation for the development of more context-sensitive and responsive insolvency prediction systems in the construction industry.

Nonetheless, this study has some limitations. The literature review was based primarily on Scopus, while the use of a snowball sampling approach helped to identify additional studies outside the initial search results. Nevertheless, some relevant studies indexed only in other databases may not have been captured, representing a potential limitation of the review. The empirical feature derivation relies on aggregated ASIC data, which, although credible, does not capture real-time or firm-level financial dynamics. Moreover, this research stops short of developing or validating a predictive model, considering its aim was to inform the structure and content of such models rather than to implement them directly.

A particularly critical future agenda item is the creation of self-assessment tools tailored for micro- and small-scale contractors or builders. These firms, which represent the majority of insolvency cases, often lack access to financial consultants or risk assessment services. Developing intuitive digital platforms that allow such businesses to input basic financial data and receive automated risk scores could empower early detection and corrective action. Such tools could be delivered through mobile applications or web dashboards, integrating predictive algorithms with simplified financial health indicators to enhance usability and impact. In conclusion, while the technical promise of AI in financial distress prediction is well established, the real opportunity lies in contextualising these models to sector realities and scaling them for inclusive, practical use, particularly among vulnerable actors in the construction industry.

This combined approach, integrating a critical synthesis of AI/ML insolvency prediction models with empirically derived, construction-specific predictive features, offers a sector-relevant knowledge base that may assist both researchers and practitioners. By linking methodological insights with operational realities, the study provides guidance that can inform the design of practical early warning systems, particularly for micro and small construction firms, with the potential to support improved financial resilience and sector stability. The findings also have implications for regulators and policymakers, who may leverage these insights to strengthen oversight mechanisms, enhance disclosure requirements, and design targeted interventions that reduce systemic vulnerabilities in the construction supply chain. For industry stakeholders, the study highlights the importance of embedding AI-enabled monitoring into routine financial management practices, ensuring that risk signals are detected early and acted upon proactively. Future research could build on these findings by implementing and testing these sector-specific features within AI/ML models, validating their performance across diverse datasets, and working with industry partners to create accessible, user-friendly diagnostic tools for early insolvency risk detection. Ultimately, the integration of sector-informed features into predictive models can advance both academic knowledge and industry practice, supporting a more resilient and transparent construction sector.