|
on Computational Economics |
| By: | Vegard H. Larsen; Leif Anders Thorsrud |
| Abstract: | Building on recent advances in Natural Language Processing and modeling of sequences, we study how a multimodal Transformer-based deep learning architecture can be used for measurement and structural narrative attribution in macroeconomics. The framework we propose combines (news) text and (macroeconomic) time series information using cross-attention mechanisms, easily incorporates differences in data frequencies and reporting delays, and can be used together with Reinforcement Learning to produce structurally coherent summaries of high-frequency news flows. Applied and tested on both simulated and real-world data out-of-sample, the results we obtain are encouraging. |
| Keywords: | multimodal transformer, structural decomposition, text analytics, macroeconomic nowcasting |
| JEL: | C45 C55 E32 E37 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_12454 |
| By: | Szymon Lis; Robert \'Slepaczuk; Pawe{\l} Sakowski |
| Abstract: | This paper investigates whether short-term market overreactions can be systematically predicted and monetized as momentum signals using high-frequency emotional information and modern machine learning methods. Focusing on Apple Inc. (AAPL), we construct a comprehensive intraday dataset that combines volatility normalized returns with transformer-based emotion features extracted from Twitter messages. Overreactions are defined as extreme return realizations relative to contemporaneous volatility and transaction costs and are modeled as a three-class prediction problem. We evaluate the performance of several nonlinear classifiers, including XGBoost, Random Forests, Deep Neural Networks, and Bidirectional LSTMs, across multiple intraday frequencies (1, 5, 10, and 15 minute data). Model outputs are translated into trading strategies and assessed using risk-adjusted performance measures and formal statistical tests. The results show that machine learning models significantly outperform benchmark overreaction rules at ultra short horizons, while classical behavioral momentum effects dominate at intermediate frequencies, particularly around 10 minutes. Explainability analysis based on SHAP reveals that volatility and negative emotions, especially fear and sadness, play a central role in driving predicted overreactions. Overall, the findings demonstrate that emotion-driven overreactions contain a predictable structure that can be exploited by machine learning models, offering new insights into the behavioral origins of intraday momentum and the interaction between sentiment, volatility, and algorithmic trading. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.18912 |
| By: | Wentao Zhang; Mingxuan Zhao; Jincheng Gao; Jieshun You; Huaiyu Jia; Yilei Zhao; Bo An; Shuo Sun |
| Abstract: | The rapid advancement of Large Language Models (LLMs) has led to a surge of financial benchmarks, evolving from static knowledge tests to interactive trading simulations. However, current evaluations of real-time trading performance overlook a critical failure mode: severe behavioral instability in sequential decision-making under uncertainty. We empirically show that LLM-based trading agents exhibit extreme run-to-run variance, inconsistent action sequences even under deterministic decoding, and irrational action flipping across adjacent time steps. These issues stem from stateless autoregressive architectures lacking persistent action memory, as well as sensitivity to continuous-to-discrete action mappings in portfolio allocation. As a result, many existing financial trading benchmarks produce unreliable, non-reproducible, and uninformative evaluations. To address these limitations, we propose AlphaForgeBench, a principled framework that reframes LLMs as quantitative researchers rather than execution agents. Instead of emitting trading actions, LLMs generate executable alpha factors and factor-based strategies grounded in financial reasoning. This design decouples reasoning from execution, enabling fully deterministic and reproducible evaluation while aligning with real-world quantitative research workflows. Experiments across multiple state-of-the-art LLMs show that AlphaForgeBench eliminates execution-induced instability and provides a rigorous benchmark for assessing financial reasoning, strategy formulation, and alpha discovery. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.18481 |
| By: | Fausch, Jürg; Frigg, Moreno; Ruenzi, Stefan; Weigert, Florian |
| Abstract: | We present improved out-of-sample predictability of future fund flows using state-of-the-art machine learning methods. Nonlinear machine learning models significantly outperform linear models in terms of out-of-sample R-squared. Using interpretable ML methods, we identify past flows and the Morningstar rating as the most important predictors for net- flows, while other past performance variables are of minor importance. We find that the importance of Morningstar ratings and expenses has increased over time. In addition, the interaction effect of past flows with the Morningstar rating has a substantial impact on future flows. Furthermore, our results demonstrate that machine learning-based fund flow predictions can be used to ex-ante differentiate between high and low-performing mutual funds. Finally, funds whose flow predictions can be improved the most using ML reveal the worst performance, consistent with the idea that liquidity management is particularly challenging for these funds. |
| Keywords: | Machine learning, fund flow prediction, big data, interpretable machine learning |
| JEL: | C45 C52 C53 C55 G10 G11 G12 G17 G23 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:cfrwps:337467 |
| By: | Wenxi Geng; Dingyuan Liu; Liya Li; Yiqing Wang |
| Abstract: | Post-hoc explainability is central to credit risk model governance, yet widely used tools such as coefficient-based attributions and SHapley Additive exPlanations (SHAP) often produce numerical outputs that are difficult to communicate to non-technical stakeholders. This paper investigates whether large language models (LLMs) can serve as post-hoc explainability tools for credit risk predictions through in-context learning, focusing on two roles: translators and autonomous explainers. Using a personal lending dataset from LendingClub, we evaluate three commercial LLMs, including GPT-4-turbo, Claude Sonnet 4, and Gemini-2.0-Flash. Results provide strong evidence for the translator role. In contrast, autonomous explanations show low alignment with model-based attributions. Few-shot prompting improves feature overlap for logistic regression but does not consistently benefit XGBoost, suggesting that LLMs have limited capacity to recover non-linear, interaction-driven reasoning from prompt cues alone. Our findings position LLMs as effective narrative interfaces grounded in auditable model attributions, rather than as substitutes for post-hoc explainers in credit risk model governance. Practitioners should leverage LLMs to bridge the communication gap between complex model outputs and regulatory or business stakeholders, while preserving the rigor and traceability required by credit risk governance frameworks. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.18895 |
| By: | Ronald Richman; Mario V. W\"uthrich |
| Abstract: | The chain-ladder (CL) method is the most widely used claims reserving technique in non-life insurance. This manuscript introduces a novel approach to computing the CL reserves based on a fundamental restructuring of the data utilization for the CL prediction procedure. Instead of rolling forward the cumulative claims with estimated CL factors, we estimate multi-period factors that project the latest observations directly to the ultimate claims. This alternative perspective on CL reserving creates a natural pathway for the application of machine learning techniques to individual claims reserving. As a proof of concept, we present a small-scale real data application employing neural networks for individual claims reserving. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.15385 |
| By: | Douglas K.G. Araujo (Banco Central do Brasil); Harald Uhlig (University of Chicago, CEPR and NBER) |
| Abstract: | As Large Language Models (LLMs) are increasingly tasked with autonomous decisionmaking, understanding their behavior in strategic settings is crucial. We investigate the choices of various LLMs in the Ultimatum Game, a setting where human behavior notably deviates from theoretical rationality. We conduct experiments varying the stake size and the nature of the opponent (Human vs. AI) across both Proposer and Responder roles. Three key results emerge. First, LLM behavior is heterogeneous but predictable when conditioning on stake size and player types. Second, while some models approximate the rational benchmark and others mimic human social preferences, a distinct “altruistic†mode emerges where LLMs propose hyper-fair distributions (greater than 50%). Third, LLM Proposers forgo a large share of total payoff, and an even larger share when the Responder is human. These findings highlight the need for careful testing before deploying AI agents in economic settings. |
| Keywords: | Ultimatum Game, LLM, AI Agents, Behavioral Economics, Algorithmic Decision Making |
| JEL: | C70 C90 D91 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:bfi:wpaper:2026-29 |
| By: | Rilke, Rainer (WHU - Otto Beisheim School of Management); Sliwka, Dirk (University of Cologne) |
| Abstract: | A large body of research across management, psychology, accounting, and economics shows that subjective performance evaluations are systematically biased: ratings cluster near the midpoint of scales and are often excessively lenient. As organizations increasingly adopt large language models (LLMs) for evaluative tasks, little is known about how these systems perform when assessing human performance. We document that, in the absence of clear objective standards and when individuals are rated independently, LLMs reproduce the familiar patterns of human raters. However, LLMs generate greater dispersion and accuracy when evaluating multiple individuals simultaneously. With noisy but objective performance signals, LLMs provide substantially more accurate evaluations than human raters, as they (i) are less subject to biases arising from concern for the evaluated employee and (ii) make fewer mistakes in information processing closely approximating rational Bayesian benchmarks. |
| Keywords: | performance evaluation, large language models, signal objectivity, algorithmic judgment, Gen-AI |
| JEL: | J24 J28 M12 M53 |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:iza:izadps:dp18371 |
| By: | da Silva, Lucas Paulo (Trinity College Dublin) |
| Abstract: | Measuring media ideology is essential for researching media bias, media effects, and various important topics in political science, communication, and other social sciences. However, given journalistic norms of objectivity and the complexity of ideology, measuring media ideology accurately is uniquely challenging. Large language models (LLMs) have become valuable tools in this endeavor. Based on media communication theories, I argue that media ideology is expressed via different cues -- the topic, argument, framing, criticism, and sources of the media content -- and that LLMs often miss these. Standard methods of LLM classification also offer little control, flexibility, and data granularity to researchers. Drawing on insights about computational and quantitative measurement methodologies, I introduce the "Multi-Cue Classification" (MQ-Class) approach. With MQ-Class, an LLM classifies the different ideological cues separately and researchers then apply pre-specified weights and thresholds to combine them into one label per text. I compare standard LLM and MQ-Class methods using two example tasks -- classifying the economic and cultural ideologies of a novel sample of online media articles. Across multiple tests, MQ-Class is more accurate and puts researchers "back in the driver's seat." I conclude by discussing how MQ-Class could be implemented for other classification tasks and data. |
| Date: | 2026–02–20 |
| URL: | https://d.repec.org/n?u=RePEc:osf:socarx:zmtqp_v1 |
| By: | Paola Ganum; Tohid Atashbar |
| Abstract: | This paper empirically evaluates the ability of current Large Language Models (LLMs) to analyze macrofinancial coverage in IMF Article IV staff reports, using human economists' assessments as a benchmark. We test several GPT models on reports from 2016-2024, assessing their performance on both qualitative ratings and binary questions. Our findings indicate that the latest models can meaningfully assist economists, achieving an average accuracy of 71-75% on ratings and an average exact match rate of 76-81% on binary questions in 2024 across advanced GPT models. However, we find that LLMs tend to assign higher, less-dispersed ratings than human experts and struggle with open-ended questions that require deep contextual judgment. The paper provides quantitative evidence on current LLM accuracy in this domain, explores the drivers of its performance, and discusses key limitations such as optimistic bias. |
| Keywords: | AI; Large Language Model; Textual Analysis; Macrofinancial Surveillance; IMF Staff Reports; Human-AI Comparison |
| Date: | 2026–02–27 |
| URL: | https://d.repec.org/n?u=RePEc:imf:imfwpa:2026/035 |
| By: | Aidan Vyas |
| Abstract: | We introduce LemonadeBench v0.5, a minimal benchmark for evaluating economic intuition, long-term planning, and decision-making under uncertainty in large language models (LLMs) through a simulated lemonade stand business. Models must manage inventory with expiring goods, set prices, choose operating hours, and maximize profit over a 30-day period-tasks that any small business owner faces daily. All models demonstrate meaningful economic agency by achieving profitability, with performance scaling dramatically by sophistication-from basic models earning minimal profits to frontier models capturing 70% of theoretical optimal, a greater than 10x improvement. Yet our decomposition of business efficiency across six dimensions reveals a consistent pattern: models achieve local rather than global optimization, excelling in select areas while exhibiting surprising blind spots elsewhere. |
| Date: | 2026–01 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.13209 |
| By: | Klimm, Johanna; Müller, Kathrin; Rößle, Felix |
| Abstract: | Diese Arbeit analysiert die Einsatzfähigkeit von Large Language Models (LLMs) wie Google Gemini und Microsoft Copilot bei der Bearbeitung komplexer Controlling-Fallstudien. Praxisnahe Themen wie Kapitalflussrechnung, Selbstkostenberechnung und Eigen- vs. Fremdfertigung werden in Deutsch und Englisch anhand von fünf Kriterien (Korrektheit, Vollständigkeit, Plausibilität, Anwendbarkeit, Verständlichkeit) geprüft und von und ChatGPT ein zweites Mal geprüft. Die Ergebnisse zeigen, dass LLMs im ersten Schritt strukturierte und plausible Lösungen liefern, jedoch ihre Leistung in der Selbsteinschätzung stark überbewerten. Signifikanztests bestätigen, dass die Selbsteinschätzungen deutlich positiver ausfallen als die Bewertungen durch eine neutrale Künstliche Intelligenz (KI) und den Vergleich mit Musterlösungen. Entgegen der Erwartung können die LLMs englischsprachige Aufgaben nicht besser lösen als deutschsprachige. Auch dieser Zusammenhang erweist sich als statistisch signifikant. Die Ergebnisse verdeutlichen, dass menschliche Expertise und Kontrolle für strategische Entscheidungen weiterhin unerlässlich sind, während LLMs derzeit vor allem als Unterstützungstools geeignet sind. |
| Keywords: | Controlling, Künstliche Intelligenz, Large Language Models |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:rpaebs:337476 |
| By: | Pavel Koptev; Vishnu Kumar; Konstantin Malkov; George Shapiro; Yury Vikhanov |
| Abstract: | Invoice or payment dilution is the gap between the approved invoice amount and the actual collection is a significant source of non credit risk and margin loss in supply chain finance. Traditionally, this risk is managed through the buyer's irrevocable payment undertaking (IPU), which commits to full payment without deductions. However, IPUs can hinder supply chain finance adoption, particularly among sub-invested grade buyers. A newer, data-driven methods use real-time dynamic credit limits, projecting dilution for each buyer-supplier pair in real-time. This paper introduces an AI, machine learning framework and evaluates how that can supplement a deterministic algorithm to predict invoice dilution using extensive production dataset across nine key transaction fields. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.15248 |
| By: | Bricongne, Jean-Charles; Meunier, Baptiste; Macalos, Joao; Milis, Julia; Pical, Thomas |
| Abstract: | We investigate whether satellite observations of nitrogen dioxide (NO₂) – a short-lived pollutant primarily emitted by fossil fuel combustion – can improve the forecasting of oil demand. After retrieving, cleaning, and aggregating daily satellite data, we integrate NO₂ into a range of forecasting models. Across a panel of advanced and emerging economies, we find that including NO₂ significantly enhances nowcasting accuracy relative to benchmark models based on autoregressive terms and traditional predictors such as industrial activity, prices, weather, and vehicle registrations. Accuracy gains are particularly strong during crisis episodes but remain present in more stable times. Non-linear models, especially neural networks, yield the largest improvements, highlighting the non-linear link between energy demand and pollution. By offering a timely, globally consistent, and freely available proxy, satellite-based NO₂ data provide a valuable new tool for real-time monitoring of oil dema JEL Classification: C51, C81, E23, E37 |
| Keywords: | big data, energy consumption, machine learning, nowcasting, satellite data |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:ecb:ecbwps:20263198 |
| By: | Ata Keskin |
| Abstract: | Factor Engine is a high-performance, open-source Python library designed for the systematic computation and analysis of financial factors. Built around a modular and extensible API that leverages Python decorators, Factor Engine enables users to define custom factors with ease and integrates seamlessly with the modern data science ecosystem. To assess its practical effectiveness, we compare the mispricing factors computed by Factor Engine to those generated using a reference Stata implementation, finding that both approaches yield highly similar results and comparable performance in backtesting analyses. Furthermore, we experimentally apply these factors within machine learning workflows for trading strategy development, illustrating their practical utility and potential for quantitative finance research. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.14138 |
| By: | Weibels, Sebastian |
| Abstract: | Theories of limited attention predict that investors rely on typical patterns to navigate high-dimensional firm characteristics, making atypical firms hard to process. To quantify this difficulty, we propose a data-driven measure of firm atypicality using an autoencoder (ATYP). The model learns typical patterns that describe most firms, and our measure aggregates the deviations those patterns cannot explain. Unlike proxies based on disclosure or organizational complexity, this approach captures the processing difficulty of the characteristics themselves. Empirically, we document that atypicality strongly predicts future returns. A decile portfolio that sells high-ATYP firms and buys low-ATYP firms earns 1.47% per month (equal-weighted) and 0.82% (value-weighted). The effect strengthens where investor attention is low and arbi- trage is limited, suggesting mispricing as the explanation. |
| Keywords: | atypical firms, processing difficulty, return predictability, mispricing, machine learning |
| JEL: | G10 G11 G12 G14 C45 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:cfrwps:337469 |
| By: | Leogrande, Angelo; Arnone, Massimo; Drago, Carlo; Costantiello, Alberto; Anobile, Fabio |
| Abstract: | The paper aims to investigate the determinants of the Perceived Risk of Crime (PRC) in Italian regions for the period 2004-2022, with data provided by the ISTAT-BES framework. The analysis relies on a regional panel dataset, which is somewhat unbalanced, with an extensive set of socio-institutional, crime, and subjective well-being variables, such as social participation, trust in people, trust in the judiciary, pickpocketing, fear of crime, life satisfaction, pessimism about the future, and dissatisfaction with the regional landscape. The analysis combines classical panel data methodologies with machine learning techniques to check the robustness of the results and to detect regional latent patterns. In all models, namely, fixed effects, random effects, dynamic panel, and weighted least squares, it is confirmed that objective crime variables, as well as subjective ones, play a crucial role in determining PRC. In particular, it is confirmed that, among the variables, pickpocketing and fear of crime are the most important positive determinants of PRC, while trust in people and trust in the judiciary have a significant mitigating effect on PRC. Variables concerning pessimism about the future and environmental dissatisfaction are also confirmed to have a positive effect on PRC. Among several machine learning alternatives, the regularized linear regression model is selected as the best-performing predictive model, which provides an interpretable and accurate representation of the relationships between the variables. In addition, model-based clustering allows us to detect different regional profiles characterized by different combinations of crime, trust, well-being, and security perceptions. In conclusion, the results confirm that PRC in Italian regions depends on the complex interaction between actual crime, emotional reactions, trust, and quality of life, suggesting that effective policies to address PRC should be based on the integrated action of crime control strategies, trust-building, social cohesion, and quality of the regional landscape. |
| Date: | 2026–02–16 |
| URL: | https://d.repec.org/n?u=RePEc:osf:socarx:rd2sv_v1 |
| By: | Jiaqi Huang |
| Abstract: | Many economic models feature moment conditions that involve latent variables. When the latent variables are individual fixed effects in an auxiliary panel data regression, we construct orthogonal moments that eliminate first-order bias induced by estimating the fixed effects. Machine Learning methods and Empirical Bayes methods can be used to improve the estimate of the nuisance parameters in the orthogonal moments. We establish a central limit theorem based on the orthogonal moments without relying on exogeneity assumptions between panel data residuals and the cross-sectional moment functions. In a simulation study where the exogeneity assumption is violated, the estimator based on orthogonal moments has smaller bias compared with other estimators relying on that assumption. An empirical application on experimental site selection demonstrates how the method can be used for nonlinear moment conditions. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.08899 |
| By: | Jeremy McEntire |
| Abstract: | We introduce Leap+Verify, a framework that applies speculative execution -- predicting future model weights and validating predictions before acceptance -- to accelerate neural network training. Inspired by speculative decoding in language model inference and by the Automatically Scalable Computation (ASC) architecture for program execution, Leap+Verify decomposes training into three dynamically detected regimes (chaotic, transition, stable) using activation-space cosine similarity as a real-time Lyapunov proxy signal. Within each regime, analytic weight predictors (momentum, linear, quadratic extrapolation) attempt to forecast model parameters K training steps ahead; predictions are accepted only when validated against a held-out loss criterion. We evaluate Leap+Verify on GPT-2 124M and Qwen 2.5-1.5B trained on WikiText-103 across five random seeds, sweeping prediction depth K in {5, 10, 25, 50, 75, 100}. Momentum-based prediction (Adam moment extrapolation) fails catastrophically at both scales, with predicted losses exceeding actuals by 100-10, 000x -- a universal norm explosion in optimizer-state extrapolation. Finite-difference predictors (linear, quadratic) succeed where momentum fails: at 124M, they achieve 24% strict acceptance at K=5 in stable regimes; at 1.5B, they achieve 37% strict acceptance in transition regimes. The scale-dependent finding is in regime distribution: GPT-2 124M spends 34% of training in stable regime, while Qwen 1.5B spends 64% in chaotic regime and reaches stable in only 0-2 of 40 checkpoints. Larger models are more predictable when predictable, but less often predictable -- the practical bottleneck shifts from predictor accuracy to regime availability. Cross-seed results are highly consistent (less than 1% validation loss variance), and the three-regime framework produces identical phase boundaries (plus or minus 50 steps) across seeds. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.19580 |