Decentralized blockchain governance for explainable causal AI in macroeconomic forecasting

1 Introduction

Macroeconomic forecasting constitutes one of the central pillars of modern economic analysis and public policy design. Accurate forecasts of inflation, economic growth, unemployment, industrial production, exchange rates, and interest rates are essential for governments, central banks, financial institutions, and international organizations seeking to maintain macroeconomic stability and support sustainable economic development (Makridakis et al., 2020). Forecasting systems play a critical role in monetary policy formulation, inflation targeting strategies, financial risk management, and investment decision-making. Consequently, improving the reliability, adaptability, and interpretability of macroeconomic forecasting models remains a major challenge in both academic research and practical policymaking.

Despite significant advances in econometric modeling over recent decades, macroeconomic forecasting continues to face substantial limitations, particularly during periods characterized by structural instability, financial crises, and elevated economic uncertainty (Lenza and Primiceri, 2022). Traditional econometric frameworks such as Autoregressive Integrated Moving Average (ARIMA), Vector Autoregression (VAR), Structural VAR (SVAR), and Dynamic Stochastic General Equilibrium (DSGE) models are generally constructed under assumptions of parameter stability, linear relationships, and relatively stable economic structures over time. However, contemporary global economic environments are increasingly shaped by financial volatility, geopolitical tensions, technological disruptions, energy crises, and uncertainty shocks that continuously modify macroeconomic dynamics and weaken historical relationships among economic variables (Baker et al., 2020; International Monetary Fund, 2024).

Major global events such as the 2008 global financial crisis, the COVID-19 pandemic, and the 2022 inflationary shock demonstrated that conventional forecasting models frequently experience severe predictive deterioration during periods of structural disruption. These crises generated abrupt changes in inflation persistence, monetary policy transmission mechanisms, supply chain structures, financial market dynamics, and global trade patterns. As a result, forecasting systems calibrated under historically stable conditions became unable to adequately capture rapidly evolving macroeconomic environments. The increasing occurrence of structural breaks and regime shifts therefore represents one of the most important methodological challenges in modern macroeconomic forecasting research (Pinto and Marçal, 2023).

Structural instability refers to changes in the underlying statistical and economic relationships governing macroeconomic variables over time. Such changes may emerge from monetary policy interventions, financial crises, institutional reforms, technological innovation, geopolitical conflicts, or external economic shocks. When structural breaks occur, parameters estimated by traditional econometric models frequently become unstable, leading to biased estimations, weak statistical inference, and substantial forecasting errors (Pinto and Marçal, 2023). Consequently, forecasting systems relying exclusively on fixed-parameter linear specifications often fail to adapt to evolving macroeconomic conditions and uncertainty regimes.

The growing complexity of macroeconomic systems has encouraged researchers to increasingly integrate machine learning and artificial intelligence techniques into economic forecasting frameworks (Rohan et al., 2025). Machine learning algorithms such as Random Forests, Gradient Boosting Machines, XGBoost, Artificial Neural Networks, and Long Short-Term Memory (LSTM) networks possess strong predictive capabilities and can efficiently process high-dimensional datasets characterized by nonlinear interactions and dynamic dependencies (Gu et al., 2020; Annastasya et al., 2025). Compared with traditional econometric approaches, machine learning methods provide greater flexibility and adaptability when modeling structurally unstable economic environments (Tao, 2025; Mohammed et al., 2026).

Recent empirical evidence suggests that machine learning algorithms frequently outperform conventional econometric models in forecasting inflation dynamics, economic growth, financial volatility, and recession probabilities. Deep learning architectures, particularly LSTM networks, are especially effective in capturing temporal dependencies and evolving nonlinear relationships within macroeconomic time-series data. Nevertheless, despite their strong predictive performance, conventional machine learning models also present important limitations. Many machine learning algorithms operate as “black-box” systems characterized by limited transparency and weak economic interpretability. In macroeconomic policy analysis, forecasting accuracy alone is insufficient because policymakers and central banks require forecasting systems capable of identifying economically meaningful causal relationships and explaining the mechanisms underlying predicted outcomes.

In response to these limitations, recent research increasingly focuses on integrating causal inference methodologies with machine learning forecasting systems. Causal machine learning combines the predictive flexibility of artificial intelligence techniques with the interpretability and identification procedures of modern econometrics. Approaches such as Double Machine Learning (DML), Causal Forests, and heterogeneous treatment effect estimation frameworks allow researchers to distinguish causal relationships from spurious statistical correlations while simultaneously handling high-dimensional macroeconomic environments. The integration of causal inference into machine learning forecasting systems therefore represents a major methodological advancement in empirical macroeconomics and policy-oriented forecasting research.

At the same time, the rapid digitalization of financial systems and the emergence of blockchain technologies are transforming the architecture of modern economic infrastructures. Blockchain systems provide decentralized, immutable, transparent, and auditable environments designed to support data integrity, governance reliability, and institutional trust within digital financial ecosystems. Initially developed to support cryptocurrencies and decentralized financial transactions, blockchain technologies are increasingly being applied to broader economic and financial applications, including supply chain management, financial governance, digital identity systems, smart contracts, and decentralized data infrastructures.

The integration of blockchain technologies into macroeconomic forecasting systems creates new opportunities for improving forecasting transparency, reproducibility, and data governance. Traditional macroeconomic forecasting frameworks frequently rely on centralized databases and opaque data processing pipelines that may suffer from issues related to data manipulation, limited auditability, and weak reproducibility. In contrast, blockchain-enabled forecasting architectures can provide immutable timestamp verification, decentralized data validation, and transparent forecasting pipelines capable of enhancing institutional credibility and trust in AI-driven policy systems.

Furthermore, the growing importance of Explainable Artificial Intelligence (XAI) has reinforced the need for transparent and interpretable forecasting frameworks. Techniques such as SHAP (Shapley Additive Explanations) and feature importance analysis enable researchers and policymakers to better understand how macroeconomic variables influence forecasting outcomes across different economic regimes. Integrating XAI methodologies with blockchain infrastructures may therefore significantly improve both the interpretability and governance of machine learning forecasting systems.

Against this background, this study proposes a blockchain-enabled causal machine learning framework for macroeconomic forecasting under structural instability and economic uncertainty. The proposed framework integrates blockchain-based data governance mechanisms, structural break detection procedures, causal inference methodologies, advanced machine learning algorithms, and Explainable Artificial Intelligence tools within a unified forecasting architecture. More specifically, the empirical framework combines Bai–Perron structural break analysis, Markov-Switching regime identification, Double Machine Learning estimation, Causal Forest methodologies, and SHAP-based explainability techniques in order to improve forecasting robustness, transparency, and causal interpretability under evolving macroeconomic conditions.

Unlike conventional AI-enabled forecasting systems, blockchain is not treated in this study as a predictive engine or machine learning component. Instead, blockchain is operationalized as a decentralized governance infrastructure that ensures secure data provenance, immutable model lineage, auditable prediction records, and institutional coordination across forecasting stakeholders. More specifically, the blockchain layer records dataset versions, preprocessing transformations, model retraining events, hyperparameter updates, forecast outputs, and SHAP-based explanations through smart-contract execution. This architecture enables explainable causal AI systems to become transparent, reproducible, and tamper-resistant within decentralized macroeconomic forecasting environments.

The empirical analysis relies on a large macroeconomic dataset covering major global crisis episodes, including the 2008 financial crisis, the COVID-19 pandemic, and the 2022 inflationary shock period. The forecasting performance of the proposed framework is compared with conventional econometric models and standard machine learning algorithms across multiple forecasting horizons and uncertainty regimes. In addition, a blockchain-based governance layer is incorporated into the proposed architecture to support immutable data provenance, dataset version control, decentralized verification, and transparent model lifecycle management. The empirical evaluation presented in this study focuses on the forecasting performance of the causal machine learning models, whereas the blockchain layer is described as the governance infrastructure supporting reproducibility and auditability.

This study contributes to the existing literature in several important dimensions. First, it develops an integrated forecasting framework combining blockchain infrastructures, causal machine learning methodologies, structural break analysis, and explainable artificial intelligence within a unified empirical environment. Second, the research extends the growing literature on decentralized financial infrastructures and AI governance by exploring the role of blockchain technologies in improving macroeconomic forecasting transparency and institutional reliability. Third, the study provides empirical evidence regarding the comparative performance of traditional econometric models, machine learning algorithms, and causal machine learning frameworks under structurally unstable economic environments.

From a practical perspective, the findings of this study may provide valuable implications for central banks, financial regulators, policymakers, and international institutions seeking more adaptive, transparent, and trustworthy forecasting systems for modern digital financial ecosystems. The integration of blockchain technologies with explainable causal artificial intelligence may contribute to the development of decentralized and auditable macroeconomic forecasting architectures capable of supporting monetary policy analysis, financial stability monitoring, and digital economic governance under conditions characterized by elevated uncertainty and structural change.

The remainder of this paper is organized as follows. Section 2 reviews the theoretical and empirical literature related to macroeconomic forecasting, blockchain technologies, causal machine learning, structural instability, and Explainable Artificial Intelligence. Section 3 presents the methodological framework and empirical procedures employed in the analysis. Section 4 discusses the empirical forecasting and causal estimation results. Section 5 provides a discussion of the principal findings and policy implications. Finally, Section 6 concludes the study and proposes directions for future research.

2 Literature review

2.1 Macroeconomic forecasting under structural instability and economic uncertainty

Macroeconomic forecasting represents a fundamental component of economic analysis, monetary policy formulation, and financial decision-making. Governments, central banks, international organizations, and financial institutions rely extensively on macroeconomic forecasts to anticipate future developments in inflation, economic growth, unemployment, exchange rates, industrial production, and financial market conditions. Accurate forecasting systems are essential for maintaining price stability, supporting sustainable economic growth, improving financial risk management, and designing effective macroeconomic policies. Consequently, the development of reliable forecasting methodologies has remained a central objective of both theoretical and empirical macroeconomic research for several decades.

Traditional macroeconomic forecasting approaches have historically been dominated by econometric time-series models based on statistical and economic theory. Early forecasting frameworks primarily relied on univariate methods such as Autoregressive (AR), Moving Average (MA), and Autoregressive Integrated Moving Average (ARIMA) models developed within the Box–Jenkins methodology. These models achieved widespread popularity because of their relative simplicity and satisfactory short-term forecasting performance. However, univariate models remain limited in their ability to capture complex interdependencies among macroeconomic variables and often fail to adequately represent the structural mechanisms underlying economic fluctuations.

The introduction of multivariate econometric frameworks marked a major advancement in macroeconomic forecasting research. One of the most influential contributions was the Vector Autoregression (VAR) model proposed by Sims, (1980), which enabled the simultaneous modeling of dynamic relationships among multiple macroeconomic variables without imposing strong theoretical restrictions. VAR models rapidly became central tools in empirical macroeconomics because they provided flexible frameworks for analyzing interactions between inflation, output growth, unemployment, interest rates, and exchange rates. Subsequent extensions such as Structural Vector Autoregression (SVAR) and Bayesian VAR (BVAR) models further improved forecasting performance by incorporating structural identification procedures and Bayesian estimation techniques capable of reducing overparameterization problems in high-dimensional datasets.

Another important methodological development involved Dynamic Factor Models (DFMs), which became increasingly relevant in data-rich macroeconomic environments characterized by large numbers of economic indicators. DFMs extract latent common factors from high-dimensional datasets and thereby improve forecasting performance while reducing dimensionality constraints. Similarly, Dynamic Stochastic General Equilibrium (DSGE) models gained substantial importance due to their microeconomic foundations and theoretical consistency. DSGE frameworks provide structural representations of interactions among households, firms, governments, and monetary authorities under rational expectations assumptions and have become widely used for monetary policy analysis and macroeconomic forecasting by central banks and international institutions.

Despite these methodological advances, macroeconomic forecasting remains highly challenging under conditions characterized by structural instability and elevated economic uncertainty. Most traditional econometric forecasting models are constructed under assumptions of parameter stability, linearity, and relatively constant economic relationships over time. In practice, however, macroeconomic systems are continuously affected by financial crises, geopolitical tensions, technological transformations, institutional reforms, pandemics, and policy regime shifts that alter the behavior of economic agents and generate structural disruptions within economic systems.

Structural instability refers to changes in the statistical properties and economic relationships governing macroeconomic variables across time. Such instability may involve shifts in means, variances, covariances, or transmission mechanisms among macroeconomic variables. Perron, (1989) emphasized that failing to account for structural changes may lead to biased parameter estimation, misleading statistical inference, and severe forecasting deterioration. Consequently, structural breaks have become a central concern in modern macroeconomic forecasting literature.

The importance of structural instability became particularly evident following major global economic crises. The 2008 global financial crisis generated unprecedented disruptions within international financial systems and significantly weakened traditional monetary transmission mechanisms. More recently, the COVID-19 pandemic created simultaneous supply-side and demand-side shocks that profoundly altered inflation dynamics, labor markets, industrial production, and global supply chains. In addition, the 2022 inflationary shock period associated with energy crises, geopolitical fragmentation, and tightening monetary policies further intensified macroeconomic uncertainty and increased forecasting difficulties worldwide.

These episodes demonstrated that macroeconomic relationships frequently behave differently during periods of crisis compared with stable economic environments. Inflation persistence, exchange rate dynamics, financial market volatility, and monetary policy effectiveness often become highly nonlinear and regime-dependent under uncertain conditions. As a result, forecasting systems calibrated using historical relationships frequently experience substantial predictive deterioration during periods characterized by structural breaks and uncertainty shocks.

One of the most important contributions to the analysis of structural instability was provided by Hamilton, (1989) through the introduction of Markov-Switching models. Hamilton demonstrated that macroeconomic variables may evolve across multiple unobservable regimes characterized by distinct statistical properties, such as expansionary and recessionary phases. Markov-Switching frameworks became widely adopted in macroeconomic and financial research because they effectively capture nonlinear dynamics, asymmetric business cycles, and abrupt changes in economic conditions.

Another major advancement in structural break analysis was developed by Bai and Perron, (2003), who proposed econometric procedures capable of detecting multiple endogenous structural breaks within time-series data. The Bai–Perron methodology allows researchers to identify unknown break dates without imposing prior assumptions regarding the timing of structural changes. This approach has become one of the most widely applied techniques in empirical macroeconomics due to its statistical robustness and flexibility in modeling evolving economic relationships.

The growing frequency of structural disruptions and uncertainty shocks has significantly challenged the predictive capacity of traditional forecasting frameworks. Forecasting models based exclusively on fixed-parameter linear relationships often fail to adapt to rapidly changing macroeconomic environments and evolving policy regimes. Consequently, recent literature increasingly emphasizes the importance of developing adaptive forecasting systems capable of incorporating nonlinear dynamics, regime changes, and structural instability into macroeconomic analysis.

Economic uncertainty has also emerged as a major determinant of forecasting performance in modern macroeconomic environments. Uncertainty indicators such as the Economic Policy Uncertainty (EPU) index and financial market volatility measures increasingly influence macroeconomic dynamics through investment behavior, consumption decisions, precautionary savings, and financial market expectations. Elevated uncertainty levels may amplify macroeconomic volatility, weaken policy transmission mechanisms, and generate substantial forecasting errors within conventional econometric systems.

Recent empirical evidence suggests that forecasting deterioration becomes particularly severe during periods characterized by heightened uncertainty and financial instability. Traditional econometric models often experience large forecasting breakdowns because historical correlations among macroeconomic variables become unstable during crises. Consequently, researchers increasingly advocate for more flexible forecasting methodologies capable of adapting to structurally unstable economic environments and evolving uncertainty regimes.

Overall, the literature on macroeconomic forecasting under structural instability highlights the growing limitations of conventional econometric approaches in modern economic environments characterized by nonlinear dynamics, structural breaks, and elevated uncertainty. These challenges have motivated the increasing integration of machine learning techniques, causal inference methodologies, and adaptive forecasting architectures into macroeconomic research. Understanding the effects of structural instability and uncertainty therefore constitutes a critical foundation for developing more robust and resilient forecasting systems capable of operating effectively within increasingly complex global economic environments.

2.2 Machine learning and causal artificial intelligence in macroeconomic forecasting

The rapid digitalization of economic systems, the expansion of computational capacity, and the growing availability of large-scale macroeconomic and financial datasets have profoundly transformed forecasting methodologies in recent decades. Traditional econometric models, although theoretically grounded and widely used in policy analysis, frequently encounter important limitations when dealing with nonlinear relationships, structural instability, and high-dimensional datasets. As a result, machine learning and artificial intelligence techniques have increasingly emerged as powerful alternatives capable of improving forecasting accuracy and adaptability in complex macroeconomic environments.

Machine learning refers to a group of computational algorithms designed to identify patterns and predictive structures directly from data with limited reliance on strict theoretical assumptions. Unlike conventional econometric approaches, which generally focus on structural interpretation and parameter estimation, machine learning methods prioritize predictive performance and adaptive learning processes. This flexibility enables machine learning systems to capture nonlinear interactions, hidden dependencies, and evolving dynamic relationships that are often difficult to model using traditional statistical frameworks.

The growing integration of machine learning into macroeconomic forecasting is closely related to the increasing complexity and dimensionality of modern economic data. Contemporary macroeconomic systems generate massive quantities of information through financial markets, digital transactions, online activity, high-frequency indicators, and real-time economic monitoring technologies. Traditional econometric frameworks often struggle to efficiently process such high-dimensional environments due to problems related to multicollinearity, overparameterization, and computational constraints. In contrast, machine learning algorithms are specifically designed to process large datasets while automatically selecting the most relevant predictive features.

One of the earliest machine learning techniques applied to macroeconomic forecasting was the Artificial Neural Network (ANN). Inspired by biological neural systems, neural networks consist of interconnected computational layers capable of approximating highly nonlinear relationships among economic variables. Empirical studies demonstrated that neural networks could significantly improve forecasting performance for inflation, economic growth, exchange rates, and unemployment dynamics. Hybrid forecasting systems combining neural networks with traditional econometric models were later developed to simultaneously capture linear and nonlinear components within macroeconomic time series.

Another major advancement in forecasting literature involved the development of ensemble learning methods. Random Forest algorithms introduced by Breiman (2001) combine multiple decision trees in order to improve forecasting stability and reduce overfitting risks. These models became increasingly popular in macroeconomic forecasting because they efficiently handle high-dimensional datasets and nonlinear interactions among explanatory variables. Empirical evidence suggests that Random Forest models frequently outperform conventional econometric approaches in forecasting inflation dynamics, financial stress indicators, and recession probabilities.

Boosting algorithms also contributed substantially to the evolution of machine learning forecasting systems. Gradient Boosting Machines (GBM) and Extreme Gradient Boosting (XGBoost) iteratively improve forecasting performance by combining weak learners into more powerful predictive structures. These algorithms are particularly effective in identifying complex interactions and nonlinear patterns within macroeconomic and financial data. Recent empirical studies indicate that boosting models generate strong forecasting performance under conditions characterized by elevated uncertainty and structural instability.

More recently, deep learning architectures have become increasingly important in macroeconomic forecasting research. Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks are specifically designed to process sequential and temporal data structures. Unlike conventional feedforward neural networks, LSTM architectures incorporate memory mechanisms capable of capturing long-term dependencies and evolving temporal relationships within macroeconomic time-series data. Consequently, LSTM models have demonstrated strong predictive performance in forecasting business cycles, inflation persistence, financial volatility, and crisis dynamics.

Despite their predictive advantages, machine learning models also face important limitations when applied to macroeconomic policy analysis. One of the principal criticisms concerns the “black-box” nature of many machine learning algorithms. Although these systems frequently generate highly accurate predictions, their internal decision-making mechanisms often remain difficult to interpret. In macroeconomics, interpretability is critically important because policymakers and central banks require forecasting systems capable not only of generating accurate predictions but also of explaining the economic mechanisms underlying forecasting outcomes.

Another major limitation concerns the distinction between correlation and causality. Most standard machine learning algorithms are designed to maximize predictive accuracy rather than identify causal relationships. Consequently, machine learning systems may capture unstable or spurious statistical associations that fail to remain valid across different economic regimes or crisis periods. This issue becomes particularly problematic in policy-oriented forecasting environments where policymakers seek to evaluate the effects of monetary interventions, fiscal policies, uncertainty shocks, or financial disturbances on macroeconomic outcomes.

In response to these limitations, recent literature increasingly emphasizes the integration of causal inference methodologies with machine learning algorithms. Causal Machine Learning represents a rapidly expanding field situated at the intersection of econometrics, statistics, and artificial intelligence. The principal objective of causal machine learning is to combine the predictive flexibility of machine learning systems with the identification procedures and interpretability of causal econometric analysis.

One of the most influential developments in this area is the Double Machine Learning (DML) framework introduced by Chernozhukov et al. (2018). DML integrates machine learning algorithms into econometric estimation procedures while preserving valid statistical inference under high-dimensional settings. The framework relies on orthogonalization and sample-splitting procedures to estimate causal effects while controlling for large numbers of confounding variables. By reducing sensitivity to overfitting and omitted variable bias, DML substantially improves causal estimation in complex macroeconomic environments characterized by nonlinear interactions and structural instability.

Another important methodological advancement involves Causal Forest algorithms developed by Athey and Imbens, (2016). Causal Forests extend Random Forest methodologies to estimate heterogeneous treatment effects across observations and economic regimes. Unlike conventional forecasting models assuming homogeneous effects, Causal Forest frameworks recognize that macroeconomic relationships may vary across time periods, economic conditions, and institutional environments. These models are therefore particularly useful for analyzing regime-dependent macroeconomic dynamics and uncertainty transmission mechanisms.

The growing integration of causal inference into machine learning forecasting systems reflects a broader methodological transition within empirical economics toward more interpretable and policy-relevant artificial intelligence systems. Policymakers increasingly require forecasting frameworks capable of identifying economically meaningful transmission mechanisms rather than relying exclusively on predictive optimization. Consequently, causal machine learning methodologies are becoming increasingly important in macroeconomic policy evaluation, inflation forecasting, financial stability analysis, and uncertainty modeling.

At the same time, the emergence of Explainable Artificial Intelligence (XAI) has further strengthened efforts to improve transparency and interpretability within machine learning systems. XAI techniques seek to identify, quantify, and explain the contribution of explanatory variables to forecasting outcomes while preserving strong predictive performance. Among the most influential explainability methods is the SHAP (Shapley Additive Explanations) framework developed by Lundberg and Lee, (2017). SHAP values provide consistent measures of variable importance derived from cooperative game theory and allow researchers to decompose forecasting outcomes into individual feature contributions.

In macroeconomic forecasting applications, SHAP analysis enables researchers and policymakers to better understand how variables such as inflation, interest rates, oil prices, exchange rates, uncertainty indicators, and financial volatility influence forecasting results across different economic regimes. Explainability techniques therefore improve not only forecasting transparency but also policy credibility and institutional trust in AI-driven decision systems.

Recent literature increasingly supports the integration of machine learning, causal inference, and explainable AI into hybrid forecasting frameworks capable of operating effectively under structurally unstable economic environments. Such systems combine predictive flexibility, causal interpretability, and transparency while improving forecasting robustness during periods characterized by crisis dynamics and elevated uncertainty.

Overall, the growing adoption of machine learning and causal artificial intelligence represents a major methodological transformation in macroeconomic forecasting research. By combining advanced computational techniques with econometric identification procedures and explainability tools, modern forecasting systems provide more adaptive, interpretable, and resilient frameworks capable of addressing the increasing complexity of contemporary macroeconomic environments.

2.3 Blockchain technologies and decentralized economic data governance

The rapid digital transformation of financial systems and economic infrastructures has significantly increased the importance of secure, transparent, and reliable data governance mechanisms within modern macroeconomic environments (Bhumichai et al., 2024; Passarella et al., 2025). Contemporary forecasting systems increasingly depend on large-scale datasets originating from heterogeneous sources, including central banks, international institutions, financial markets, digital platforms, and high-frequency economic monitoring systems. Although the growing availability of economic data creates substantial opportunities for improving macroeconomic forecasting performance, it also raises important challenges related to data integrity, transparency, reproducibility, cybersecurity, and institutional trust (Potluri, 2025). In this context, blockchain technologies have emerged as promising infrastructures capable of improving decentralized data governance and enhancing the reliability of digital economic systems (Casino et al., 2020).

Blockchain technology refers to a decentralized distributed ledger system designed to record, validate, and store transactions or data entries in immutable and transparent structures (Nassar et al., 2020). Unlike conventional centralized databases controlled by single institutions or intermediaries, blockchain systems operate through distributed consensus mechanisms that ensure synchronization and verification across multiple network participants. Each validated block contains cryptographically secured information linked to previous blocks, thereby creating immutable chains resistant to tampering and unauthorized modifications. Consequently, blockchain infrastructures provide important advantages in terms of transparency, traceability, auditability, and security (Alzoubi, 2025).

Initially developed as the technological foundation of cryptocurrencies such as Bitcoin, blockchain technology has rapidly expanded into broader financial, economic, and institutional applications (Casino et al., 2020). The emergence of decentralized finance (DeFi), smart contract systems, digital assets, and Central Bank Digital Currencies (CBDCs) has significantly accelerated the integration of blockchain architectures into modern financial ecosystems (Celestin et al., 2025). Beyond digital payments and financial transactions, blockchain systems are increasingly applied to supply chain management, healthcare systems, digital identity verification, governance structures, and data management infrastructures (Bhumichai et al., 2024).

One of the most important characteristics of blockchain systems concerns their capacity to improve data integrity and trust within decentralized environments (Nassar et al., 2020). Traditional macroeconomic forecasting systems generally rely on centralized data infrastructures in which economic information is collected, processed, and distributed by a limited number of institutions. Although centralized systems provide administrative efficiency, they may also face challenges related to data manipulation risks, limited transparency, weak auditability, and institutional dependence. In contrast, blockchain-based data governance frameworks enable decentralized verification and immutable timestamping of economic information, thereby improving the credibility and reliability of forecasting pipelines (Potluri, 2025).

The increasing complexity of macroeconomic forecasting environments has reinforced the need for transparent and auditable forecasting systems. During periods of financial instability, geopolitical tensions, or elevated uncertainty, policymakers and financial institutions require forecasting architectures capable of ensuring both predictive reliability and institutional trust. Blockchain technologies may substantially contribute to these objectives by enabling decentralized storage, automated verification mechanisms, and secure sharing of macroeconomic data across multiple institutions and stakeholders (Passarella et al., 2025).

Another major innovation associated with blockchain systems is the development of smart contracts. Smart contracts are self-executing computational protocols operating on blockchain networks that automatically implement predefined rules and validation procedures without requiring centralized intermediaries (Casino et al., 2019). Smart contracts therefore provide programmable governance mechanisms capable of automating data validation, transaction execution, and forecasting workflows within decentralized economic systems.

In macroeconomic forecasting applications, smart contracts may be employed to automate the collection, verification, and updating of macroeconomic indicators originating from multiple data providers. Such mechanisms improve forecasting reproducibility and reduce operational risks associated with manual data processing procedures. In addition, blockchain-enabled validation systems may ensure immutable storage of forecasting results, thereby strengthening transparency and accountability within AI-driven policy infrastructures (Nassar et al., 2020).

The integration of blockchain systems with artificial intelligence and machine learning technologies has recently emerged as an important interdisciplinary research field (Bhumichai et al., 2024; Alzoubi, 2025). Artificial intelligence systems often require large quantities of reliable data for training, forecasting, and decision-making purposes. However, AI-driven forecasting architectures frequently suffer from problems related to data quality, limited transparency, cybersecurity vulnerabilities, and algorithmic trust. Blockchain technologies provide potential solutions to these limitations by enabling secure data sharing, decentralized validation procedures, and transparent governance mechanisms for AI systems (Potluri, 2025).

Recent literature increasingly emphasizes the complementarities between blockchain infrastructures and artificial intelligence frameworks (Issami, 2025). Blockchain systems can improve the reliability and traceability of datasets used within machine learning environments, while AI algorithms may enhance the efficiency and scalability of blockchain networks through predictive optimization and automated analytics. The convergence between blockchain and AI technologies therefore creates opportunities for developing trusted and explainable intelligent systems capable of supporting financial governance and macroeconomic policy analysis.

The growing emphasis on Explainable Artificial Intelligence (XAI) has further reinforced the relevance of blockchain-enabled forecasting systems (Nassar et al., 2020). Explainability has become a central concern in AI governance because policymakers and institutions increasingly require transparent decision-making systems capable of providing interpretable and accountable forecasting outcomes. Blockchain technologies may complement explainable AI methodologies by creating immutable records of forecasting inputs, model updates, parameter adjustments, and prediction outputs. Consequently, integrating blockchain architectures into AI-driven forecasting systems may substantially improve transparency, accountability, and reproducibility.

Another important area of development concerns decentralized oracle systems and blockchain-based economic data infrastructures. Decentralized oracles are mechanisms that enable blockchain networks to securely access external information sources, including macroeconomic indicators, financial market data, and real-world economic events. These systems are particularly relevant for decentralized financial ecosystems because they ensure reliable transmission of external economic information into blockchain environments. In macroeconomic forecasting contexts, decentralized oracle systems may support automated updating of forecasting models through real-time integration of economic indicators and financial variables.

Blockchain technologies also play an increasingly important role in the development of Central Bank Digital Currencies and digital monetary infrastructures (Celestin et al., 2025). Many central banks worldwide are currently exploring blockchain-based architectures for implementing digital payment systems and improving financial governance mechanisms. The expansion of CBDCs may significantly transform monetary transmission channels, financial intermediation processes, and macroeconomic monitoring systems. Consequently, integrating blockchain infrastructures into macroeconomic forecasting frameworks becomes increasingly relevant within modern digital financial ecosystems.

Despite their advantages, blockchain systems also face several important limitations and challenges. Public blockchain networks may experience scalability constraints, high energy consumption, latency issues, and governance difficulties. In addition, integrating blockchain infrastructures into large-scale macroeconomic forecasting systems requires substantial computational resources, institutional coordination, and regulatory adaptation. Concerns related to privacy protection, interoperability, cybersecurity, and legal governance also remain central issues within blockchain research and implementation (Alzoubi, 2025).

Nevertheless, recent advances in distributed ledger technologies, consensus mechanisms, and hybrid blockchain architectures continue to improve the scalability and efficiency of blockchain systems (Shi and Wang, 2025). Private and consortium blockchain models increasingly provide more flexible infrastructures suitable for institutional forecasting environments requiring controlled access and regulatory compliance.

Overall, blockchain technologies represent a major innovation in decentralized economic governance and digital financial infrastructures. By improving transparency, immutability, auditability, and trust within data management systems, blockchain architectures create important opportunities for developing more reliable and accountable macroeconomic forecasting frameworks (Casino et al., 2020). The integration of blockchain systems with causal machine learning and explainable artificial intelligence may therefore contribute to the emergence of decentralized forecasting architectures capable of operating effectively within increasingly complex and uncertain economic environments (Nassar et al., 2020; Bhumichai et al., 2024).

2.4 Research gap and development of a blockchain-enabled causal forecasting framework

The growing complexity of global economic systems, combined with recurrent financial crises, structural disruptions, and rapid technological transformation, has significantly intensified the methodological challenges associated with macroeconomic forecasting (Makridakis et al., 2020; Yu et al., 2025). Although substantial progress has been achieved in econometric modeling, machine learning, and digital financial technologies, important limitations remain regarding forecasting robustness, interpretability, transparency, and adaptability under structurally unstable economic environments (Pinto and Marçal, 2023). Consequently, recent literature increasingly emphasizes the necessity of developing integrated forecasting frameworks capable of combining predictive performance, causal interpretability, decentralized governance, and institutional reliability within unified analytical architectures (Gunay et al., 2025).

One of the principal limitations of existing macroeconomic forecasting literature concerns the persistent separation between traditional econometric approaches and modern machine learning methodologies (Rohan et al., 2025). Conventional econometric frameworks such as VAR, DSGE, and Dynamic Factor Models provide strong theoretical consistency and policy interpretability but frequently suffer from limited adaptability under nonlinear and unstable economic conditions. In contrast, machine learning algorithms demonstrate strong predictive capabilities in high-dimensional environments but often operate as black-box systems characterized by weak economic transparency and limited causal interpretation (Arsenault et al., 2025). As a result, forecasting research remains fragmented between structurally interpretable econometric models and highly predictive but insufficiently explainable artificial intelligence systems.

Another important limitation concerns the insufficient integration of causal inference methodologies into macroeconomic forecasting systems (Mahadeva et al., 2025). Most standard machine learning forecasting approaches primarily focus on statistical prediction and optimization rather than identifying economically meaningful causal relationships. Consequently, these models may capture unstable correlations and spurious associations that deteriorate during crisis periods or regime shifts. In macroeconomic policy analysis, however, understanding causal transmission mechanisms remains essential because policymakers require forecasting systems capable not only of generating accurate predictions but also of explaining how policy interventions, uncertainty shocks, and financial disturbances affect macroeconomic outcomes.

Recent developments in causal machine learning, particularly Double Machine Learning and Causal Forest methodologies, have provided important methodological advances capable of addressing some of these limitations. These approaches combine machine learning flexibility with econometric identification procedures, thereby improving both predictive robustness and causal interpretability. Nevertheless, existing applications of causal machine learning in macroeconomic forecasting remain relatively limited and often focus primarily on forecasting performance without sufficiently addressing issues related to governance transparency, reproducibility, and decentralized data validation.

At the same time, blockchain technologies have emerged as transformative infrastructures capable of improving transparency, immutability, and trust within digital financial ecosystems (Casino et al., 2020). However, current blockchain research remains largely concentrated on cryptocurrencies, decentralized finance (DeFi), digital assets, and smart contract systems. Comparatively limited attention has been devoted to the potential role of blockchain infrastructures in macroeconomic forecasting and economic data governance (Gunay et al., 2025). Existing studies exploring blockchain and artificial intelligence integration frequently focus on cybersecurity, healthcare, logistics, or decentralized transaction systems rather than policy-oriented forecasting architectures.

Moreover, the literature on blockchain-enabled AI systems rarely addresses macroeconomic forecasting under structural instability and uncertainty (Yu et al., 2025). Most existing frameworks fail to incorporate structural break detection procedures, regime-switching dynamics, causal inference methodologies, and explainable artificial intelligence tools within decentralized forecasting environments. Consequently, important research gaps remain regarding how blockchain technologies can support adaptive and transparent macroeconomic forecasting systems capable of operating effectively under evolving economic conditions.

Another significant limitation concerns forecasting transparency and reproducibility. Traditional forecasting systems generally rely on centralized databases and opaque data-processing infrastructures that may limit auditability, reproducibility, and institutional trust. During periods characterized by elevated uncertainty and policy sensitivity, such limitations may weaken confidence in forecasting systems and reduce the credibility of policy decisions derived from AI-driven analytical frameworks. Although Explainable Artificial Intelligence techniques such as SHAP and feature importance analysis have improved interpretability within machine learning systems, transparency regarding data governance, model updating procedures, and forecasting validation remains insufficiently addressed within existing literature (Perez-Bernabeu and Polat, 2025; Arsenault et al., 2025).

The growing importance of digital financial ecosystems and Central Bank Digital Currencies further reinforces the need for decentralized and trustworthy macroeconomic forecasting architectures. As financial systems become increasingly digitalized and interconnected, policymakers require forecasting systems capable of integrating real-time economic information, ensuring data integrity, and supporting transparent policy analysis within technologically complex environments. The convergence between blockchain infrastructures, artificial intelligence systems, and macroeconomic governance therefore represents an increasingly important area of interdisciplinary research.

Against this background, the present study seeks to address several important gaps within the existing literature by developing a blockchain-enabled causal machine learning framework for macroeconomic forecasting under structural instability and economic uncertainty. The proposed framework integrates four complementary methodological dimensions within a unified forecasting architecture.

First, the framework incorporates structural break detection and regime-switching methodologies in order to account for evolving macroeconomic dynamics and nonlinear economic environments (Pinto and Marçal, 2023). Second, the study integrates advanced machine learning algorithms and causal inference techniques, including Double Machine Learning and Causal Forest estimation procedures, to improve forecasting performance while preserving economically meaningful causal interpretation. Third, blockchain-based data governance mechanisms are introduced to enhance transparency, immutability, reproducibility, and decentralized verification of macroeconomic forecasting pipelines (Casino et al., 2019). Fourth, Explainable Artificial Intelligence techniques are incorporated to improve forecasting interpretability and strengthen institutional trust in AI-driven policy systems (Perez-Bernabeu and Polat, 2025).

The proposed framework therefore contributes to bridging the gap between econometric forecasting theory, causal artificial intelligence, and decentralized digital infrastructures. Unlike existing forecasting systems that frequently prioritize either predictive accuracy or interpretability, the proposed architecture simultaneously seeks to achieve predictive robustness, causal transparency, decentralized governance, and explainability under structurally unstable economic conditions.

From a theoretical perspective, this study contributes to the emerging interdisciplinary literature at the intersection of macroeconomics, artificial intelligence, and blockchain technologies. The research extends existing forecasting literature by demonstrating how blockchain-enabled infrastructures may improve the governance and transparency of AI-driven forecasting systems. In addition, the study contributes to causal machine learning research by applying advanced causal inference methodologies within decentralized forecasting environments characterized by structural instability and uncertainty shocks.

From an empirical perspective, the study provides comparative evidence regarding the performance of conventional econometric models, standard machine learning algorithms, and blockchain-enabled causal machine learning frameworks across multiple crisis periods and forecasting environments. The empirical analysis therefore offers new insights regarding the role of decentralized infrastructures and explainable causal AI in improving forecasting resilience under uncertain macroeconomic conditions.

Overall, the identified research gaps highlight the growing need for integrated forecasting architectures capable of combining machine learning adaptability, causal interpretability, explainable AI, and decentralized blockchain governance within unified macroeconomic forecasting systems. The present study seeks to address these challenges by proposing a transparent, auditable, and adaptive forecasting framework designed to support policy analysis and economic decision-making within increasingly complex digital financial environments.

3 Methodology

3.1 Decentralized data infrastructure and blockchain governance

The empirical analysis relies on a monthly macroeconomic dataset spanning January 2000 to March 2025, yielding 303 observations per variable. The dataset combines macroeconomic, monetary, and financial indicators collected from publicly accessible institutional databases, including the Federal Reserve Economic Data (FRED), International Monetary Fund (IMF), OECD, World Bank, and Economic Policy Uncertainty database.

All datasets used in this study originate from publicly accessible institutional repositories and can therefore be independently retrieved and verified. Detailed information regarding data availability is provided in the Data Availability Statement at the end of this manuscript.

The dependent variable is inflation, measured as year-over-year Consumer Price Index (CPI) growth. Explanatory variables include GDP growth, unemployment rate, short-term interest rate, exchange rate index, Brent oil prices, money supply growth (M2), financial market volatility (VIX), and Economic Policy Uncertainty (EPU). These variables were selected based on macroeconomic theory and prior evidence regarding inflation determinants under structurally unstable environments.

Missing observations represented less than 2.3% of the total dataset and were imputed using Kalman smoothing to preserve temporal continuity. Extreme outliers outside the 1st and 99th percentiles were winsorized to reduce sensitivity to crisis-induced anomalies. All input features were standardized using z-score normalization prior to model estimation.

Table 1 summarizes the macroeconomic variables used in the empirical analysis, including their definitions, data sources, observation frequency, and preprocessing transformations.

TABLE 1. Dataset summary.

Source: Compiled by the authors using data from the Federal Reserve Economic Data (FRED), International Monetary Fund (IMF), OECD, data explorer, World Bank Open Data, and the Economic Policy Uncertainty (EPU) database.

The final dataset contains 303 monthly observations after preprocessing. Missing values (<2.3%) were imputed using Kalman smoothing, while extreme observations were winsorized at the 1st and 99th percentiles.

To preserve temporal dependence and avoid look-ahead bias, the dataset was split chronologically into training (70%), validation (15%), and test (15%) subsets. Rolling-window cross-validation was employed for hyperparameter tuning and robustness assessment across forecasting horizons.

3.2 Explainable causal AI forecasting framework

This study develops a blockchain-enabled causal machine learning framework designed to improve macroeconomic forecasting performance under conditions characterized by structural instability, nonlinear dynamics, and elevated economic uncertainty. The proposed framework integrates blockchain-based data governance mechanisms, structural break detection procedures, econometric forecasting models, machine learning algorithms, causal inference methodologies, and Explainable Artificial Intelligence (XAI) techniques within a unified empirical architecture. The primary objective of the framework is to enhance forecasting robustness, transparency, causal interpretability, and reproducibility within unstable macroeconomic environments.

Figure 1 illustrates the decentralized blockchain governance architecture of the proposed framework. The system integrates six layers: data sourcing, preprocessing, blockchain validation, causal AI forecasting, explainability analysis, and policy decision support. Blockchain operates as an immutable governance layer ensuring traceability of datasets, model versions, retraining events, and prediction outputs across institutional nodes.

It is important to note that the blockchain layer does not participate directly in the forecasting process and is therefore not evaluated using predictive performance metrics. Instead, its role is to ensure secure data provenance, immutable model versioning, prediction traceability, and auditability throughout the forecasting workflow.

A blockchain-based validation layer is incorporated into the forecasting architecture in order to ensure secure storage, decentralized verification, immutable timestamping, and transparent governance of macroeconomic datasets. Macroeconomic indicators collected from institutions such as the IMF, OECD, World Bank, and FRED are validated using cryptographic hash verification procedures before entering the forecasting pipeline. The blockchain verification process is represented as:where D t denotes the macroeconomic dataset at time t, H . represents the cryptographic hash function, and h t denotes the immutable blockchain identifier. Smart contracts are additionally employed to automate data validation and forecasting update procedures within the decentralized forecasting environment.

H D t = h t (1)

To account for structural instability and evolving economic regimes, the study first applies structural break detection methodologies. The Bai and Perron, (2003) multiple structural break framework is employed to identify endogenous breakpoints associated with major economic disruptions and regime shifts. The structural break model is expressed as:for:where Y t denotes the macroeconomic dependent variable, X t represents explanatory variables, β j denotes regime-specific coefficients, T j represents structural break dates, and ε t denotes the error term.

Y t = X t β j + ε t (2)

t = T j − 1 + 1 , . . . , T j (3)

The analysis additionally employs Markov-Switching models introduced by Hamilton, (1989) in order to identify latent expansionary and recessionary regimes within macroeconomic dynamics. The Markov-Switching specification is represented as:where s t denotes the unobserved economic regime, μ s t represents regime-specific intercepts, and ϕ s t denotes regime-dependent autoregressive parameters.

Y t = μ s t + ϕ s t Y t − 1 + ε t (4)

The interaction between the structural break detection module and the forecasting engine follows a sequential pipeline. First, Bai–Perron breakpoint detection identifies major endogenous structural shifts associated with crisis episodes and policy regime changes. Second, the Markov-Switching model estimates latent macroeconomic regimes, distinguishing expansionary and recessionary states. These regime labels are then incorporated into the forecasting framework as additional state-dependent features.

The regime-adjusted dataset is subsequently used to train benchmark econometric models (VAR and TVP-VAR), machine learning models (Random Forest, XGBoost, and LSTM), and causal machine learning models (DML and Causal Forests). DML estimates average causal effects while controlling for high-dimensional confounders, whereas Causal Forests capture heterogeneous treatment effects across regimes. Finally, SHAP analysis is applied to the best-performing model to explain feature contributions at both global and observation levels.

The forecasting analysis compares conventional econometric frameworks with advanced machine learning and causal machine learning models. Benchmark econometric forecasting systems include the Vector Autoregression (VAR) model proposed by Sims, (1980), specified as:where Y t denotes the vector of endogenous macroeconomic variables, A p represents coefficient matrices, and ε t denotes innovations. To account for parameter instability, the study also estimates Time-Varying Parameter VAR (TVP-VAR) models expressed as:where A t denotes time-varying coefficient matrices.

Y t = A 1 Y t − 1 + A 2 Y t − 2 + . . . + A p Y t − p + ε t (5)

Y t = A t Y t − 1 + ε t (6)

The machine learning component of the framework includes Random Forest, XGBoost, and Long Short-Term Memory (LSTM) neural network models. The forecasting methodology follows established time-series forecasting principles (Box et al., 2015), while the XGBoost model is based on the gradient-boosting framework developed by Chen and Guestrin (2016). The LSTM architecture follows the recurrent neural network formulation introduced by Hochreiter and Schmidhuber (1997). The Random Forest predictor is represented as:where T b x denotes the b-th decision tree and B represents the number of trees. The XGBoost optimization function is expressed as:where l . denotes the forecasting loss function and Ω . represents regularization components. To capture temporal dependencies and nonlinear macroeconomic dynamics, the study additionally employs LSTM neural networks using the following memory update mechanism:where C t denotes the memory cell state, f t represents the forget gate, and i t denotes the input gate.

f ^ x = 1 B ∑ b = 1 B T b x (7)

L ϕ = ∑ i = 1 n l y i , y ^ i + ∑ k Ω f k (8)

C t = f t C t − 1 + i t C ∼ t (9)

The LSTM architecture consists of two hidden recurrent layers with 64 units each, followed by a dense output layer for one-step-ahead forecasting. A sequence length of 12 months was used to capture annual temporal dependencies in macroeconomic dynamics. Dropout regularization (0.20) was applied to reduce overfitting. The model was trained using the Adam optimizer with learning rate 0.001, batch size 32, and 200 training epochs.

To ensure reproducibility and robust model comparison, the machine learning models were trained using the hyperparameter configurations reported in Table 2. Parameter selection was based on rolling-window cross-validation and validation-set optimization.

TABLE 2. Hyperparameter Configuration and training settings for machine learning models.

Source: Authors’ own compilation based on the rolling-window cross-validation procedure used in this study.

Table 2 reports the optimal hyperparameter configuration selected for each forecasting model together with the corresponding validation strategy. Hyperparameter optimization was performed using rolling-window cross-validation to preserve the temporal ordering of the data. For the Double Machine Learning framework, five-fold block cross-fitting was employed to estimate nuisance functions while maintaining valid causal inference under serial dependence.

The central methodological contribution of the study lies in integrating causal inference methodologies into machine learning forecasting systems. The analysis employs the Double Machine Learning (DML) framework proposed by Chernozhukov et al. (2018) in order to estimate causal relationships while controlling for high-dimensional confounding variables. The partially linear structural model is represented as: where Y denotes the macroeconomic outcome variable, D represents the treatment variable, X denotes high-dimensional controls, and θ 0 represents the causal effect parameter. The study also employs Causal Forest methodologies to estimate heterogeneous treatment effects across economic regimes. The Conditional Average Treatment Effect (CATE) is defined as:

Y = D θ 0 + g 0 X + U (10)

D = m 0 X + V (11)

τ x = E Y 1 − Y 0 ∣ X = x (12)

Because macroeconomic time-series data exhibit serial dependence, volatility clustering, and regime persistence, standard iid assumptions underlying conventional DML implementations may be violated. To address this issue, the present study adopts block cross-fitting instead of random sample splitting, thereby preserving temporal dependence within each fold. Specifically, observations are partitioned into contiguous time blocks for nuisance parameter estimation and orthogonal score construction. In addition, statistical inference is conducted using heteroskedasticity and autocorrelation consistent (HAC) standard errors based on Newey–West corrections. These adjustments ensure robust causal inference under weak temporal dependence and structural instability.

To improve forecasting transparency and interpretability, the framework integrates Explainable Artificial Intelligence techniques using SHAP (Shapley Additive Explanations) values introduced by Lundberg and Lee, (2017). The SHAP decomposition is expressed as:where ϕ i denotes the contribution of explanatory feature i and M represents the total number of explanatory variables.

f x = ϕ 0 + ∑ i = 1 M ϕ i (13)

Forecasting performance is evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE), defined respectively as:

R M S E = 1 n ∑ i = 1 n y i − y ^ i 2 (14)

M A E = 1 n ∑ i = 1 n ∣ y i − y ^ i ∣ (15)

M A P E = 100 n ∑ i = 1 n ∣ y i − y ^ i y i ∣ (16)

Forecast comparisons are additionally evaluated using Diebold–Mariano tests to determine the statistical significance of forecasting differences across competing models.

3.3 Blockchain governance layer

Although the primary objective of this study is to evaluate the forecasting performance of econometric, machine learning, and causal machine learning models, the proposed framework additionally incorporates a blockchain governance layer designed to strengthen transparency, auditability, data provenance, and reproducibility throughout the forecasting lifecycle. Unlike the forecasting engine, the blockchain component is not intended to improve predictive accuracy directly and therefore is not evaluated using forecasting performance metrics. Instead, it functions as a decentralized governance infrastructure that records and verifies the complete history of datasets, model development, prediction generation, and explainability outputs.

The blockchain governance layer operates immediately after data preprocessing and before model estimation. Following the collection of macroeconomic indicators from publicly available institutional repositories, including FRED, the International Monetary Fund (IMF), the OECD, the World Bank, and the Economic Policy Uncertainty database, the data undergo standard preprocessing procedures comprising missing-value imputation, outlier treatment, feature engineering, and normalization. Once preprocessing is completed, each dataset version is assigned a unique cryptographic fingerprint generated using the SHA-256 hashing algorithm. Rather than storing the complete datasets on the blockchain, only the corresponding cryptographic hash together with descriptive metadata is recorded on the distributed ledger. This design preserves storage efficiency while ensuring that any modification of the original dataset can be immediately detected through hash verification.

The metadata associated with each blockchain transaction include the dataset identifier, acquisition date, preprocessing version, feature configuration, timestamp, and software environment used during data preparation. Because any modification to the underlying dataset generates a different cryptographic hash, the blockchain provides an immutable mechanism for guaranteeing data integrity and complete provenance throughout the forecasting pipeline.

In addition to dataset validation, the governance layer maintains an immutable record of the model development process. Each model training event is linked to the corresponding dataset version and includes metadata describing the forecasting algorithm, hyperparameter configuration, software version, training period, random seed, and model identifier. Consequently, every forecasting model can be uniquely associated with the exact dataset version from which it was generated, thereby facilitating independent verification and full reproducibility of the forecasting experiments.

The governance architecture also records prediction metadata after model estimation. Each forecasting output is associated with information describing the forecast horizon, prediction timestamp, model version, evaluation metrics, and execution identifier. Similarly, the explainability module records references to the SHAP explanations associated with each prediction, thereby establishing a transparent relationship between model outputs and their corresponding interpretations. Such traceability enables policymakers and researchers to verify not only the generated forecasts but also the underlying explanations supporting each prediction.

To coordinate these operations, the proposed architecture assumes the use of permissioned blockchain smart contracts responsible for validating metadata consistency before new records are appended to the distributed ledger. The smart contracts verify dataset identifiers, preprocessing versions, model configurations, timestamps, and prediction metadata, ensuring consistency across participating institutional nodes while preventing unauthorized modifications of the governance records. Since only metadata and cryptographic fingerprints are stored on-chain, whereas the original datasets remain within secure institutional repositories, the proposed architecture combines the transparency and immutability of distributed ledger technology with the scalability required for macroeconomic forecasting applications.

It should be emphasized that the blockchain governance layer constitutes an architectural component of the proposed forecasting framework rather than an additional predictive model. Accordingly, the empirical analysis presented in Section 4 focuses exclusively on comparing the forecasting performance and causal inference capabilities of the econometric and machine learning models. The contribution of the blockchain layer lies instead in improving governance-related properties of the forecasting system, including immutable data provenance, transparent model lifecycle management, prediction traceability, auditability, and reproducibility. By clearly separating predictive intelligence from governance functions, the proposed framework provides a trustworthy foundation for deploying explainable causal AI in institutional macroeconomic forecasting environments.

4 Results

The empirical analysis evaluates the forecasting performance of conventional econometric models, machine learning algorithms, and blockchain-enabled causal machine learning frameworks under conditions characterized by structural instability and economic uncertainty. Because the blockchain component serves as a governance infrastructure rather than a predictive algorithm, the empirical analysis reported in this section evaluates only the forecasting performance of the econometric, machine learning, and causal machine learning models. The blockchain layer supports data integrity, traceability, and reproducibility but is not benchmarked as a forecasting model.

4.1 Forecasting performance comparison

The first stage of the empirical analysis focuses on comparing out-of-sample forecasting accuracy across benchmark econometric, machine learning, and causal AI models. This comparison provides an initial assessment of the predictive gains associated with integrating causal machine learning under structural instability. Table 3 reports the corresponding forecasting performance metrics.

TABLE 3. Comparative out-of-sample forecasting performance across benchmark models.

Source: Authors’ own calculations based on the empirical forecasting experiments.

The results reveal that the Double Machine Learning framework generates the lowest forecasting errors across all evaluation metrics, outperforming both traditional econometric models and standard machine learning algorithms.

Figure 2 Out-of-sample forecasting performance comparison across benchmark models. Lower values indicate better predictive performance. The proposed DML-based causal AI framework achieves the lowest RMSE, MAE, and MAPE, demonstrating superior forecasting accuracy relative to conventional econometric and machine learning approaches.

4.2 Forecasting robustness during major crisis regimes

To assess model robustness under extreme macroeconomic instability, we next evaluate forecasting performance during major crisis regimes characterized by severe structural breaks and elevated uncertainty. Specifically, the analysis focuses on three major stress episodes: the 2008 global financial crisis, the COVID-19 shock, and the 2022 inflationary surge. Table 4 reports the corresponding crisis-period forecasting errors across benchmark models.

TABLE 4. Forecasting performance across major macroeconomic crisis regimes.

Source: Authors’ own calculations based on crisis-period forecasting experiments.

Table 4 reports forecasting performance during three major macroeconomic stress episodes: the 2008 global financial crisis, the COVID-19 shock, and the 2022 inflationary surge. Forecasting errors increase substantially across all models during crisis regimes, confirming the destabilizing impact of structural breaks on macroeconomic prediction. The COVID-19 period exhibits the highest forecasting error, reflecting unprecedented volatility and nonlinear disruptions. Nevertheless, the proposed DML-based causal AI framework remains the most robust model, achieving the lowest average crisis-period RMSE and demonstrating stronger resilience to regime shifts than benchmark econometric and machine learning models.

Figure 3 further illustrates the forecasting performance of the competing models across 1-, 3-, 6-, and 12-month forecasting horizons. The results show that forecasting errors generally increase as the forecasting horizon extends, reflecting the greater uncertainty associated with longer-term macroeconomic predictions. Nevertheless, the proposed DML-based causal AI framework consistently achieves the lowest RMSE across all forecasting horizons, demonstrating its superior predictive accuracy and robustness relative to the benchmark econometric and machine learning models.

4.3 Causal machine learning results

While the previous sections focused on predictive accuracy and forecasting robustness, effective macroeconomic policy design also requires understanding the causal mechanisms driving inflation dynamics. To identify the key determinants of inflation under structural instability, we estimate causal effects using the Double Machine Learning (DML) framework. Table 5 reports the estimated causal effects along with standard errors, confidence intervals, and significance levels.

TABLE 5. Estimated causal effects of macroeconomic drivers on inflation using DML.

Source: Authors’ own estimates obtained using the Double Machine Learning framework.

Table 5 reports the causal effects estimated using the Double Machine Learning (DML) framework. The results indicate that oil prices exert the strongest positive causal effect on inflation, with a one-unit increase associated with a 0.52 increase in inflation, holding other factors constant. This finding confirms the critical role of energy price transmission in inflation dynamics, particularly during supply-side shocks and crisis periods.

Money supply growth also exhibits a strong and statistically significant positive effect (0.41), suggesting that expansionary liquidity conditions amplify inflationary pressures. Exchange rate depreciation (0.31) significantly increases inflation, reflecting imported inflation effects in open macroeconomic systems. Economic Policy Uncertainty (0.24) and financial volatility (0.18) also contribute positively to inflation, indicating that uncertainty shocks and financial stress worsen price instability.

In contrast, interest rate shocks show a significant negative causal effect (−0.38), confirming the disinflationary role of monetary tightening. This result is consistent with conventional monetary transmission theory, where higher policy rates reduce aggregate demand and inflationary pressures. Overall, the causal estimates reveal that inflation dynamics are driven by a combination of monetary, external, and uncertainty-related channels, with substantial heterogeneity across structural regimes.

Figure 4 Estimated causal effects of macroeconomic variables on inflation using Double Machine Learning (DML). Dots represent estimated treatment effects, while horizontal bars denote 95% confidence intervals. Oil prices and money supply growth exhibit the strongest positive causal effects on inflation, whereas interest rate shocks significantly reduce inflation. All estimated effects remain statistically significant at the 5% level.

While Table 4 and Figure 4 quantify average causal effects, the SHAP analysis in the next section further reveals how the importance of these variables varies across regimes and forecast observations.

4.4 Explainable artificial intelligence (XAI) analysis

To improve interpretability of the forecasting framework, SHAP (Shapley Additive Explanations) analysis was applied to quantify the contribution of each macroeconomic variable to model predictions. Unlike global causal coefficients, SHAP values provide observation-level explanations and reveal how feature importance changes across forecasting regimes and crisis episodes.

Figure 5 SHAP summary plot of global feature importance in inflation forecasting. Each point represents the contribution of a feature to individual model predictions. Higher absolute SHAP values indicate stronger influence on forecasting outcomes. Oil prices emerge as the dominant predictor, followed by interest rates and exchange rates, confirming the importance of energy shocks, monetary policy transmission, and imported inflation channels.

The SHAP distribution further shows strong regime dependence. During stable economic periods, monetary variables dominate prediction dynamics. During crisis regimes, however, Economic Policy Uncertainty and financial volatility become substantially more influential, indicating that uncertainty shocks amplify nonlinear inflation responses.

These findings have important policy implications. Central banks should adopt regime-sensitive forecasting systems, emphasizing inflation anchoring during stable periods while increasing attention to uncertainty management and external shocks during crisis regimes.

From a governance perspective, blockchain enhances the auditability of explainability outputs by ensuring immutable storage of SHAP explanations alongside model predictions. This enables regulators and institutions to verify not only forecasting outcomes but also the reasoning behind AI-driven decisions, thereby strengthening transparency and institutional trust.

4.5 Governance benefits and implementation trade-offs of blockchain integration

Since blockchain is not intended to directly improve forecasting accuracy, its contribution must be evaluated through governance-related dimensions. The following comparative assessment examines how blockchain integration enhances transparency, reproducibility, traceability, and decentralized coordination relative to conventional forecasting infrastructures.

Table 6 compares the governance characteristics of conventional centralized forecasting pipelines with the proposed blockchain-governed architecture. Unlike traditional infrastructures, where data revisions, preprocessing steps, and model updates may remain partially undocumented, the blockchain-based framework ensures complete traceability and immutable verification of all critical forecasting operations.

TABLE 6. Comparative governance assessment of conventional and blockchain-based forecasting pipelines.

Source: Authors’ own calculations.

The comparative analysis shows that blockchain does not directly improve forecasting accuracy but substantially enhances governance-related dimensions. The most important gains are observed in auditability, traceability, reproducibility, and tamper resistance. Through cryptographic hashing and smart-contract execution, each dataset version, preprocessing transformation, hyperparameter update, retraining event, forecasting output, and SHAP explanation is permanently recorded in the distributed ledger. This creates a fully auditable AI forecasting pipeline and reduces risks associated with silent data modification or undocumented model recalibration.

Blockchain integration also improves coordination among decentralized institutional stakeholders such as central banks, ministries of finance, statistical agencies, and international organizations. By relying on a shared ledger and common validation rules, participating institutions can verify data integrity and forecasting updates using a transparent consensus protocol, thereby reducing information asymmetry and strengthening trust in AI-generated policy intelligence.

Despite these governance benefits, blockchain adoption introduces important implementation trade-offs. Frequent model retraining and high-frequency data updates increase computational overhead due to cryptographic hashing, smart-contract execution, and network synchronization. Scalability challenges may emerge when multiple nodes participate in real-time validation. Furthermore, consortium governance requires carefully designed rules for access control, validation rights, consensus mechanisms, and smart-contract management. These findings suggest that blockchain should primarily be viewed as a governance infrastructure that strengthens transparency, accountability, and institutional trust rather than as a mechanism for improving predictive performance.

4.6 Robustness and sensitivity analysis

To assess the stability of the empirical findings, additional robustness checks were performed under alternative data splits and sample specifications. Table 7 reports the forecasting performance of the DML framework across multiple sensitivity scenarios.

TABLE 7. Robustness and sensitivity analysis of the DML forecasting framework.

Source: Developed by the authors.

The results indicate strong stability of the proposed framework. Changing the training-validation-test split from 70/15/15 to 60/20/20 leads to only a marginal increase in RMSE. Similarly, excluding the COVID-19 period or the 2022 inflation shock does not materially alter model performance or causal interpretation. A reduced feature specification produces slightly higher forecasting errors, confirming the importance of the full macroeconomic feature set.

Overall, the robustness analysis confirms that the superior forecasting performance and causal estimates of the proposed framework are not driven by a specific sample period, crisis episode, or model configuration.

5 Discussion

The discussion highlights that the proposed blockchain-enabled causal machine learning framework significantly improves macroeconomic forecasting performance under conditions characterized by structural instability and elevated economic uncertainty. The empirical results show that traditional econometric models remain highly vulnerable to structural breaks and nonlinear economic dynamics, whereas machine learning algorithms provide greater forecasting flexibility and adaptability. However, standard machine learning systems still suffer from limited interpretability and weak causal transparency.

The findings demonstrate that integrating causal inference methodologies, particularly Double Machine Learning (DML), substantially enhances forecasting robustness and causal interpretability across different economic regimes. The analysis further reveals that oil prices, uncertainty indicators, financial volatility, and monetary policy variables exert strong regime-dependent effects on inflation dynamics, especially during crisis periods.

The study also emphasizes the importance of blockchain technologies in improving forecasting transparency, auditability, reproducibility, and decentralized data governance. By integrating blockchain-based validation mechanisms with Explainable Artificial Intelligence (XAI) techniques such as SHAP analysis, the proposed framework enhances institutional trust and policy reliability within AI-driven forecasting systems.

Overall, the results suggest that combining blockchain infrastructures, causal machine learning, and explainable AI contributes to the development of more adaptive, transparent, and trustworthy macroeconomic forecasting architectures capable of supporting policy analysis and financial governance in increasingly complex digital economic environments.

Although the proposed blockchain governance layer enhances the conceptual transparency and traceability of the forecasting workflow, the present study does not empirically evaluate blockchain-specific performance indicators such as transaction latency, throughput, storage overhead, consensus efficiency, or execution costs. Accordingly, the governance-related benefits discussed in this study should be interpreted as architectural capabilities of the proposed framework rather than experimentally validated performance improvements. Future research will focus on implementing the governance layer within a permissioned blockchain platform to quantitatively assess its operational performance under real institutional forecasting environments.

6 Conclusion

This study developed a blockchain-enabled causal machine learning framework for macroeconomic forecasting under conditions characterized by structural instability, nonlinear dynamics, and elevated economic uncertainty. By integrating blockchain-based data governance mechanisms, structural break detection procedures, causal inference methodologies, advanced machine learning algorithms, and Explainable Artificial Intelligence (XAI) techniques within a unified empirical architecture, the research aimed to improve forecasting robustness, transparency, interpretability, and institutional reliability in modern macroeconomic environments.

The empirical findings demonstrate that traditional econometric forecasting models such as VAR and TVP-VAR remain highly vulnerable to structural breaks, uncertainty shocks, and evolving macroeconomic relationships. Although machine learning algorithms including Random Forest, XGBoost, and LSTM networks significantly improve forecasting performance by capturing nonlinear interactions and high-dimensional dynamics, conventional predictive AI systems continue to face important limitations related to weak causal interpretability and limited transparency.

The results further reveal that integrating causal inference methodologies into machine learning forecasting systems substantially improves forecasting accuracy and robustness across different economic regimes and crisis periods. In particular, the Double Machine Learning framework consistently generated the strongest forecasting performance under structurally unstable environments, demonstrating superior adaptability to uncertainty shocks and regime-dependent macroeconomic dynamics. The analysis additionally confirmed that oil prices, Economic Policy Uncertainty indicators, financial volatility measures, and monetary policy variables constitute major drivers of inflation dynamics, especially during recessionary and high-uncertainty regimes.

An important contribution of this study lies in the incorporation of blockchain technologies into macroeconomic forecasting systems. The proposed blockchain-enabled governance architecture improves data integrity, forecasting reproducibility, decentralized verification, and transparency of AI-driven forecasting pipelines. By integrating immutable timestamp verification and decentralized validation mechanisms, blockchain infrastructures strengthen institutional trust and accountability within macroeconomic forecasting environments increasingly characterized by digitalization and technological complexity.

The study also demonstrates the growing importance of Explainable Artificial Intelligence in policy-oriented forecasting systems. SHAP-based explainability analysis revealed substantial regime-dependent variation in macroeconomic forecasting drivers and significantly improved the interpretability of machine learning outputs. The integration of explainable AI techniques with blockchain-enabled infrastructures therefore contributes to the development of more transparent, auditable, and policy-relevant forecasting architectures.

From a theoretical perspective, this research contributes to the emerging interdisciplinary literature at the intersection of macroeconomics, artificial intelligence, and blockchain technologies. The study extends existing forecasting literature by proposing a decentralized and explainable causal forecasting framework capable of simultaneously addressing predictive performance, causal interpretation, and governance transparency under structurally unstable economic conditions.

From a practical perspective, the findings provide important implications for central banks, financial regulators, policymakers, and international institutions seeking more adaptive and trustworthy forecasting systems within increasingly digital financial ecosystems. The proposed framework may support inflation-targeting strategies, financial stability monitoring, uncertainty analysis, and digital monetary governance through more resilient and transparent forecasting infrastructures.

Despite these contributions, certain limitations remain. The implementation of blockchain-enabled forecasting architectures may involve substantial computational costs, scalability constraints, and institutional coordination challenges. In addition, the complexity of advanced machine learning systems may still generate interpretability limitations under highly nonlinear forecasting environments. Consequently, future research may further explore the integration of real-time decentralized oracle systems, federated learning architectures, transformer-based deep learning models, and Central Bank Digital Currency (CBDC) infrastructures within blockchain-enabled macroeconomic forecasting systems.

The proposed blockchain governance layer provides an architectural foundation for improving transparency, auditability, and reproducibility in AI-assisted macroeconomic forecasting. Future work will focus on implementing and empirically evaluating this governance layer within operational blockchain environments.

Overall, the findings of this study suggest that the convergence between blockchain technologies, causal machine learning, and explainable artificial intelligence represents a promising methodological direction for the future of macroeconomic forecasting. By improving transparency, adaptability, causal interpretability, and decentralized governance, blockchain-enabled causal forecasting frameworks may contribute to the development of more reliable and resilient economic decision-making systems capable of operating effectively within increasingly uncertain and digitally interconnected global economies.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Bakkiyaraj Kanthimathi Malamuthu, Morgan Stanley Financial Services, United States

Reviewed by: Shanmugam Muthu, Kalasalingam Academy of Research and Education, India

Vinoth Kumar Arumugam, Kalasalingam Academy of Research and Education, India