The foundations of the old paradigm were Linear Mathematics and classical Economic Theory. The roots of the new paradigm are not only different but represent a liberation from the constraints imposed by those foundations. E.conometrics is made possible by three conceptual and technological “liberations”: the liberation from linearity (Non-Linear Mathematics), the liberation from analytical tractability (High Computational Power), and the liberation from homo economicus (Neural Sciences).

3.1.1. Non-Linear Mathematics and Econophysics

While classical econometrics sought to force inherently non-linear economic phenomena into the tractable framework of linear mathematics, the first root of e.conometrics embraces this complexity, using econophysics, an interdisciplinary field that applies theories and methods from statistical physics to economic systems.

The fundamental contribution of econophysics is a shift in perspective: the so-called “stylized facts” (heavy-tailed return distributions, volatility clustering, leverage effects) are no longer seen as anomalies or violations of assumptions to be “corrected” (as attempted by ARCH/GARCH models in the old paradigm). Instead, they are recognized as the universal and fundamental statistical properties of a complex system composed of interacting agents. Seminal works like those of Mantegna and Stanley and Bouchaud and Potters ([26]; [5]) established the empirical foundations for this approach, demonstrating that these statistical regularities are the equivalent of physical laws for economic systems.

3.1.2. High Computational Power

If the first root provides the language (non-linear mathematics), the second provides the engine. Classical econometrics was itself a product of the first computational revolution: the advent of mainframe computers in the 1960s made large macro-econometric models like the MIT-FED-PENN possible ([27]). We are now living through the second computational wave, defined by widespread access to High Performance Computing (HPC) and Graphics Processing Units (GPUs).

This new class of computing power is the root that makes the other components of the paradigm practicable. Many of the theories of complexity and non-linear models had existed for decades but remained largely computationally intractable. It is the exponential increase in computing power that directly enables data-driven approaches (requiring the optimization of millions of parameters) and simulation-based approaches (requiring the simulation of millions of interacting agents) ([18]).

3.1.3. Neural Sciences

The third root provides the liberation from the most restrictive assumption of classical economics: the rational, optimizing agent, which forms the microeconomic foundation of the DSGE models criticized in Section 2. Neural Sciences, and in particular Neuroeconomics, provide an empirical and biologically plausible micro-foundation for economic behaviour ([32]; [12]).

This field studies the neurobiological and computational bases of value-based decision-making, integrating methods from neuroscience, behavioural economics, and psychology. Neuroeconomic research conclusively demonstrates that human decision-making is not a unitary process of utility maximization, but the result of the interaction between multiple, often competing, systems that govern emotions, habits, heuristics, and cognitive biases. This understanding of “bounded rationality” provides the intellectual justification and behavioural foundations for building bottom-up, agent-based models, as we will see in the Left Trunk.

This first trunk represents the generative and simulative approach to e.conometrics. Instead of starting from top-down aggregate equations (like the CLRM), this approach builds the system bottom-up. It defines the behavioural rules of heterogeneous agents (based on insights from Neural Sciences) and their modes of interaction; the macroscopic behaviour of the system (cycles, crashes, patterns) then emerges spontaneously from these micro-interactions.

3.2.1. Branch 1: Agent Based Models (ABMs)

ABMs are computational simulations that model a system as a collection of autonomous and heterogeneous agents (be they individuals, households, firms, or banks). These agents interact with each other and their environment according to adaptive, often sub-optimal, behavioural rules.

The primary function of ABMs is to act as “digital laboratories” to explore the dynamics of complex systems. They excel at explaining emergent phenomena—such as speculative bubbles, sudden crashes, and economic cycles—that classical equilibrium models, by their very nature, cannot predict ([11]; [37]; [2]). They also allow researchers and policymakers to explore “counterfactual narratives” (e.g., “What would have happened if Lehman Brothers had been saved?”) and to test the impact of policy interventions (like a fiscal stimulus or financial regulation) in a complex environment, capturing non-linear effects and distributional consequences. Recent reviews confirm that ABMs are “coming of age,” moving from theoretical constructs to practical tools for policy and economic analysis.

3.2.2. Branch 2: Networks Models (NMs)

This branch complements ABMs by shifting the focus from the agents themselves to the connections between them. The modern economy, and particularly the financial system, is a complex graph of interdependencies: interbank networks, global supply chains, cross-credit exposures, and trade flows ([8]; [21]). The analysis of these network structures is fundamental to understanding the system’s resilience and fragility.

This methodological approach is the tool of choice for studying systemic risk, shock propagation, and financial contagion. It allows for the identification of “too connected to fail” institutions not based on their size, but on their centrality in the network. It is crucial to distinguish this branch (the theory and econometric analysis of networks) from its counterpart in the tree crown. Networks Models described here are the theoretical framework for describing and measuring the system’s topology; Graph Neural Networks (Section 3.4.1) are the deep learning tool that learns and predicts the dynamics on that same structure.

This second trunk represents the predictive and theory-agnostic approach. While classical econometrics (Section 2) starts from a strong theoretical hypothesis and uses data to estimate a pre-specified model, this approach inverts the process. Leveraging the High Computational Power root, it “lets the data speak”, using statistical-computational methods to discover complex patterns and relationships without imposing a rigid a priori structure.

3.3.1. Branch 3: Machine Learning (ML)

Machine Learning(ML) offers a set of non-parametric statistical tools for the analysis of complex systems, focusing primarily on predictive accuracy. The adoption of ML replaces or supplements many steps of traditional econometric analysis ([28]; [29]; [16]):

1) Variable Selection: Instead of relying on economic theory and (often problematic) p-values for variable selection, ML uses techniques based on regularization (like LASSO) or predictive importance, which effectively manage high dimensionality and collinearity.

2) Unstructured Data: Unlike econometric models, which require structured numerical data, ML models can natively ingest high-frequency alternative data, such as text (news articles, social media), through embedding techniques that transform qualitative information into dense numerical representations.

While classical econometrics is focused on parameter estimation and causal inference, ML is optimized for prediction.

3.3.2. Branch 4: Deep Learning (DL)

Deep Learning (DL) is a subclass of ML based on artificial neural networks with multiple layers (deep neural networks). These models have proven to be the most powerful tool for capturing extremely complex non-linear dependencies and long-term temporal patterns in data.

In e.conometric applications, architectures like Long Short-Term Memory (LSTM) ([19]) or Transformers([24]; [39]) are the direct successors to the time-series models (VAR, SARIMA) criticized in Section 2. However, unlike their predecessors, DL models do not require rigid assumptions of stationarity or linearity and can simultaneously handle a vast number of time series.

Furthermore, DL models benefit from so-called scaling laws ([22]): unlike classical econometric models, which have a fixed number of parameters and quickly reach a performance plateau, DL models improve their predictive accuracy as computational power and, especially, the volume of training data increase. This makes them the technology of choice in the era of big data.

The true frontier of e.conometrics lies not in the choice between simulation (Left Trunk) and data (Right Trunk), but in their synthesis. The tree crown represents this hybridization: the use of data-driven models to calibrate and inform simulations, and the use of theoretical and simulative structures to constrain and guide data-driven models.

3.4.1. Graph Neural Networks (GNNs)

Graph Neural Networks (GNNs) represent the direct fusion of the Left Trunk (Networks Models) and the Right Trunk (Deep Learning). They are a deep learning architecture designed specifically to operate on graph-structured data ([40]).

While classical network analysis (Section 3.2.2) can describe the structure of a financial network, a GNN can learn the dynamics of shock propagation on that same network. In a systemic risk context, a GNN is trained to recursively aggregate information on the state of neighbouring nodes (other banks) and predict how the default or stress of one institution will propagate through the system, capturing non-linear contagion effects. Comprehensive surveys, such as that by [38], have mapped the broad potential of these architectures for the non-Euclidean data typical of finance.

3.4.2. Physics-Informed Neural Networks (PINNs)

Physics-Informed Neural Networks (PINNs) fuse the Root (Non-Linear Mathematics/Econophysics) with the Right Trunk (Deep Learning) ([30]; [6]). These models directly address the main problem of purely data-driven approaches: the “black box.” A standard DL model, trained for pure predictive accuracy, might generate economically absurd or physically impossible results, such as violating the no-arbitrage principle.

PINNs solve this by incorporating theoretical knowledge—physical laws or, in our context, financial Partial Differential Equations (PDEs), like the Black-Scholes equation ([3])—directly into the loss function of the neural network. As detailed in the seminal work of Raissi, Perdikaris, and Karniadakis ([31]), the network is trained simultaneously for two objectives: i) to minimize the error on observed data (market prices) and ii) to minimize the “residual” of the PDE, i.e., how much the network’s solution violates the theoretical equation. This forces the neural network to learn solutions that are not only predictively accurate but also theoretically coherent.

3.4.3. Deep Hedging (DH)

Deep Hedging (DH) is perhaps the most sophisticated example of hybridization, blending simulation (Left Trunk) with Deep Reinforcement Learning (DRL), an advanced branch of Deep Learning (Right Trunk). This approach offers a radical departure from classical hedging models (like the delta-hedging derived from the Black-Scholes model).

The Black-Scholes model assumes perfect, complete, and frictionless markets. Deep Hedging, in contrast, does not seek an analytical solution. Instead, it trains a DRL “agent” to learn the optimal hedging strategy by interacting with a market simulator (often an ABM) that realistically incorporates real-world frictions: transaction costs, liquidity constraints, and market impact.

The main contribution on this was provided by [7]. The agent is not trained to replicate a theoretical price, but to directly optimize a risk measure of the final Profit & Loss(P&L)distribution, such as the Conditional Value-at-Risk (CVaR). This leads to more robust, realistic, and economical hedging strategies, designed for incomplete and illiquid markets.

This work has traced the evolution from classical econometrics to e.conometrics. The logical tree of econometrics (Figure 1), based on linear mathematics and strong top-down theoretical assumptions (CLRM, DSGE), has demonstrated its fragility in the face of a non-linear, unstable, and shock-prone global economy.

The new e.conometrics paradigm (Figure 2) is not a single replacement model but a flexible, bottom-up framework that rests on multidisciplinary foundations (Roots): Non-Linear Mathematics (Econophysics), High Computational Power, and Neural Sciences. This framework integrates two main approaches: Modelling Complexity (Left Trunk), which uses simulation (ABMs) and network models (NMs) to understand emergent behaviours; and Data-Driven Approaches (Right Trunk), which use ML and DL to extract predictive patterns from large volumes of data.

The frontier of this field (The Crown) lies in the hybridization of these approaches—GNNs, PINNs, and DH—to create models that are simultaneously predictively powerful, theoretically grounded, and robust to real-world frictions.

A crucial element that distinguishes the new paradigm is the use of alternative, high-frequency data, capable of capturing dynamics that traditional macroeconomic datasets, with their slowness and rigidity, cannot grasp ([9]). Digital transactions, geolocalized mobility, satellite imagery, energy consumption, and textual or visual information extracted from the web represent a new information infrastructure.

These seemingly heterogeneous and noisy data take on predictive value when processed by machine learning architectures (Right Trunk) and linked to the world of econophysics (Root), capable of selecting, weighing, and integrating them with the constraints of economic theory (Crown). The output is not a deterministic forecast, but a multidimensional nowcast, an ultra-high-resolution snapshot of economic reality.

The most innovative aspect is the ability to capture emergent behaviours: phenomena that arise from the interaction of millions of agents (Left Trunk) and cannot be deduced from the sum of individual historical data. Power laws in firm distribution, volatility clusters in financial markets, and consumption patterns that change in real-time following exogenous shocks are examples of emergent regularities that become visible only through the combination of fast data and advanced computational power.

The paradigm shift from econometrics to e.conometrics is not our mere theoretical proposal but is already being actively implemented in pioneering projects. A significant example is the ongoing research named “Econometric Accelerator,” conducted by the Italian International Economic Center ([20]) and the LUISS Quantum & AI Lab at Luiss Guido Carli University, with the computational support of CINECA.

This project embodies the principles of e.conometrics: it moves away from monolithic, theory-heavy models (like those in Section 2) to employ a dynamic, multi-model system that integrates Artificial Intelligence (Right Trunk) with econophysics principles (Root) and the use of a vast array of alternative and high-frequency data (Section 4.2).

The experimental setup can be summarised as follows. The target variable is the Italian real GDP quarter-on-quarter growth rate (year-over-year). The dataset comprises more than 50 variables at mixed frequencies (daily, monthly, and quarterly), sourced from official providers (ISTAT, OECD, Banca d’Italia, ECB), financial data vendors (Bloomberg, Refinitiv), and alternative data sources including electronic payment transactions, energy consumption, tourism flows, and sentiment indicators. The time series span from 1980Q1 to 2024Q4. Mixed-frequency data are aligned to quarterly frequency through a preprocessing pipeline that combines deterministic aggregation with Bayesian imputation. Model validation follows an Expanding Window Rolling Forecast strategy, ensuring strict temporal separation between training and test sets to prevent data leakage. The forecasting architecture is based on a dynamic ensemble combining time-series foundation models, gradient boosting regressors, and tabular foundation models, with model selection optimised independently for each forecast horizon (1 to 4 quarters ahead). The institutional forecasts shown in Figure 3 (grey lines) correspond to publicly available projections issued by major Italian and international institutions over the same period.

The effectiveness of this approach, particularly in tackling periods of extreme volatility, is illustrated by the model’s performance during the COVID-19 crisis. Figure 3 shows an out-of-sample forecast of the Italian GDP growth rate, comparing the IIEC’s AI-driven estimate (blue line) with the actual growth rate (orange line) and the forecasts of various institutions relying on traditional econometric models (grey lines) ([33]).

Figure 3. A comparison between the actual Italian GDP growth rate (orange line) and out-of-sample forecasts over the 2019-2024 period.

As is evident, the sudden and severe economic shock of 2020 caused a dramatic plunge in GDP, followed by a strong rebound in 2021. The traditional models (grey lines), calibrated on decades of relative stability and bound by linear assumptions, failed to capture the magnitude of this non-linear event.

In contrast, the IIEC model (blue line), by leveraging its ability to process real-time data and adapt to structural breaks (principles of the Right Trunk), tracked the real economic trajectory with notably greater precision. Figure 4 provides the quantitative validation of this observation.

If we observe the Median Absolute Percentage Errors (MdAPEs or Median MAPEs), Symmetric Median Absolute Percentage Errors (SMdAPEs or Median SMAPEs), and Root Median Squared Errors (RMdSEs or Median RMSEs), in each case, a lower dark bar signifies a smaller forecast error and thus higher accuracy of results.

At this stage, the aim is to demonstrate that the methodology is consistent with the e.conometric paradigm and capable of delivering more accurate economic predictions than the baselines described above.

Figure 4. Quantitative performance analysis of forecast errors.

This preliminary result provides compelling evidence that the assumptions of linearity and stationarity underlying classical econometrics (Section 2) are inadequate for today’s volatile world ([33]).

The statistical basis, while adequate for a preliminary validation, was more limited than the researchers would have desired, owing in part to difficulties in accessing proprietary data sources for research purposes—a challenge that would benefit from a clearer legal framework distinguishing research use from commercial exploitation.

The preliminary empirical validation of the AI Econometric Accelerator project demonstrates that e.conometrics can provide more resilient and reliable tools for navigating and understanding the modern economy.

The weaknesses of the old models have real costs. Fiscal policies, which use econometric forecasts to draft public budgets, often mask forecasting errors behind complex political needs. Monetary policies, on the other hand, are paying a high cost in the form of losing their forward guidance function, applying a logic of data-driven methods on past data, that arrives late and often relies on personal judgments rather than robust e.conometrics models.

In this context, the real economy suffers while virtual finance thrives, fueled by the same uncertainty that the old models cannot tame. Embracing complexity, harnessing diverse data sources, and integrating the advanced computational methods described in the e.conometrics tree (Figure 2) is not just an academic exercise but a necessity for policymakers and market operators.

While the e.conometrics framework represents a step forward in forecasting methodology, some limitations of the current work should be noted.

These constraints primarily relate to the present scope of the application, which serves as a preliminary proof of concept for the broader e.conometrics architecture. Future research will aim to extend this model to a wider array of macroeconomic indicators and diverse geographic regions, thereby enhancing the framework’s robustness across varied and heterogeneous economic landscapes. There is also significant room to further optimize the synergy between the different structural layers of the e.conometrics “crown”, particularly in refining how advanced computational power interacts with hybrid modeling frameworks to handle increasingly complex datasets. Additionally, while the current approach effectively integrates various information streams, subsequent iterations will focus on the continuous refinement of data acquisition techniques, especially concerning high-frequency and alternative sources that are essential for navigating periods of structural change and high uncertainty. Such developments will facilitate the evolution of the methodology into a more versatile and adaptive tool, capable of providing deeper insights for decision-makers and stakeholders operating within an ever-shifting global economic environment, ultimately bridging the gap between theoretical modeling and real-world operational needs.