|
on Artificial Intelligence |
| By: | Aaron P. Kaye; Kazimier Smith; Neil Thompson; Aaron Kaye |
| Abstract: | We characterize the lifecycles of open-weight AI models using a weekly panel of the 108, 074 most-downloaded models on Hugging Face, which together account for 99.5 percent of the platform's more than 53 billion downloads. We link these data to token usage from OpenRouter and performance rankings from Arena and Artificial Analysis. We establish key institutional details about the supply of and demand for AI models, and we organize our analysis into three phases: birth, life, and death. In the "birth" phase, we show that major developers often release models in batches, concurrently introducing models that are both vertically differentiated (e.g., by size, capability, and latency) and horizontally differentiated (e.g., by task specialization, hardware compatibility, and alignment). Because weights are open, developers can also build on one another's models. We document cumulative innovation through fine-tuned, quantized, merged and adapted descendants of source models. In the "life" phase, we quantify how demand evolves after release. Even within the sample of top models, demand is concentrated, and the concentration persists: the median model loses ∼68 percent of its release-week downloads within ten weeks, while the most popular models decline gradually; as a result, the top percentile's advantage over the median model is ∼600-fold at release, and ∼900-fold six months later. Further, we estimate spillover effects from new popular model releases on related models and find that the first popular third-party descendant is associated with a 76 to 99 percent increase in weekly downloads of its source model. In the "death" phase, we construct measures of technical obsolescence and usage obsolescence. We find that technical obsolescence is associated with a slow but persistent decrease in demand, while usage obsolescence coincides with an immediate but temporary drop in demand. Our findings characterize open-weight AI as a market of complementary model portfolios, one in which third-party developers building on a source model can raise demand for the original, and in which demand is slow to respond to technical obsolescence. |
| Keywords: | artificial intelligence, open-source, open-weights, technology diffusion, platform economics, cumulative innovation |
| JEL: | O33 L17 L86 O31 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_13004 |
| By: | Maxim Chupilkin |
| Abstract: | AI systems are becoming participants in the evaluation and use of scientific research. They encounter citation counts, download statistics and lists of popular articles developed around human readers, but the collective consequences of these signals for artificial readers remain uncertain. This paper adapts the Music Lab design to a market for academic attention. In the first experiment, 1, 000 AI agents choose papers from the titles and abstracts of all 114 regular research articles published in the American Economic Review in 2025. The experiment has five independent-choice communities and five social-influence communities, each with 100 sequential agents. Only agents in the social-influence condition observe earlier selections within their community. Agents may select any number of papers. Social-information communities select 17.2 percent fewer papers per agent, concentrate their choices more heavily, and collectively cover 73 papers, compared with 90 independently. Between-community variation is greater under social information. In a second experiment with 200 agents across twenty social communities, randomly assigning papers five initial selections raises their subsequent selection rate by 45.55 percentage points (95% CI: 41.20 to 49.90). Choices have modest correspondence with external citations and little correspondence with download counts. The results show how a simple information rule shapes the volume, breadth and distribution of scientific attention in an artificial population. |
| Date: | 2026–09 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2609.22408 |
| By: | Vinicius Ferraz; Leon Houf |
| Abstract: | People increasingly compete against AI agents rather than other human opponents. We distinguish two channels: an opponent effect and an information effect. These are different elements with different consequences: the opponent effect is specific to a given computational system, the information effect a property of the information environment that an organisation or policymaker can control. We separate them in a preregistered experiment (N = 1, 395) using a dynamic all-pay auction, a repeated contest in which escalation of commitment arises from the incentives. What participants are told about the opponent (human, an AI trained to imitate people, or an AI trained to compete well) is varied and crossed with who they actually face, in a deception-free design. What people are told influences escalation: the median price rises by 6.7 points when a human might be the opponent and falls by 8.8 when an optimising machine might be, a spread of about 15% of the prize value of the competition, produced by information alone. Competing against the AI agents lowers prices, yet reduces the chance that both sides finish with positive earnings, showing distinct effects of the opponent channel. The information effect is not explained by articulated strategy, or individual differences, and is consistent with a competitive response engaged when a human is a live possibility. This shows that describing an AI competitor is not behaviourally neutral. |
| Date: | 2026–09 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2609.21439 |
| By: | Davood Wadi; Yu Ma |
| Abstract: | Large language models (LLMs) now serve as conversational shopping assistants on platforms that also sell advertising. These AI agents face a conflict of duty. They advise consumers who rely on their judgment, yet are deployed by platforms that benefit when sponsored listings are chosen. Sponsorship disclosures, designed to allow consumers to penalize paid placements, now reach the AI agent rather than the consumer, and the agent's evaluation of them is hidden from the consumer. Drawing on the fiduciary concept of conflict of duty, we argue that an agent's evaluation of a sponsored listing should not depend on which party deployed it. In controlled choice experiments, we manipulate assigned roles in the system prompt to name either a traveler or a booking platform as the agent's principal. Platform delegation significantly attenuates the penalty that agents apply to sponsored listings and weakens the skepticism that disclosure triggers in their reasoning traces. We replicate out findings across LLMs and reasoning depths. A second study decomposes the disclosure label and shows that the divergence between the two delegates widens significantly when the paid placement is attributed to the platform. Stricter terminology ("Sponsored" instead of "Promoted") lowers choice of paid listings but does not close this gap when the platform is named. The findings show that disclosure mandates designed for human consumers cannot by themselves protect consumers in AI-mediated commerce. |
| Date: | 2026–09 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2609.17989 |
| By: | Alex Smolin; Bryan Wilder |
| Abstract: | Inferring intelligence from observable behavior is a foundational challenge in artificial intelligence. We develop a theory of Bayesian intelligence for agents such as language models. Each prompt induces a possibly imperfect internal experiment; the agent updates a full-support prior by Bayes' rule and faithfully reports its posterior over the possible answers to the question. Repetitions draw fresh, independent outcomes from the same unobserved experiment at one fixed state. We show that the agent's behavior admits this explanation if and only if its reports are not fully contradictory, i.e., some state remains possible under every report across all prompts. Report frequencies and the sizes of positive probabilities impose no further restrictions. We further propose and characterize the behavioral implications of an intelligence order that makes the behavior of two agents consistent with one agent having access to a more informative experiment: there should exist a coupling of report distributions such that the more informative agent's report excludes every answer excluded by its counterpart. Finally, we show the difficulty of aggregating coarse reports from intelligent agents: unless the agent reports a belief about the complete state of the world, the optimal aggregation can assign arbitrary weights to states that have not been excluded. These results provide a basis for understanding when agents' behavior is intelligent and highlight the difficulty of rejecting Bayesian rationality. |
| Date: | 2026–09 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2609.14724 |
| By: | Carol Luengo (University of Waikato); Steven Tucker (University of Waikato); Yilong Xu (Utrecht School of Economics, Utrecht University); Frank Scrimgeour (University of Waikato) |
| Abstract: | This paper investigates the causal impact of artificial intelligence (AI) advice on price discovery in a controlled asset-market experiment. Using a call market setting in which the asset's value follows a stochastic geometric random walk, we compare a baseline "No AI" condition against two treatments: "Good AI" (advice aligned with long-term fundamentals) and "Bad AI" (myopic advice anchored to current buyout prices). Our results show that access to AI advice significantly reduces mispricing relative to the baseline. In particular, Good AI market prices started low, then converged and closely followed the fundamentals in the second half of the market. Interestingly, we find no meaningful difference in aggregate mispricing between the Good and Bad AI treatments. Analysis of trader behavior reveals that while over half of the participants follow AI recommendations, a significant portion actively filters out erroneous advice, particularly in the Bad AI treatment. This selective adherence explains why market prices remain resilient to poor advice. |
| Keywords: | artificial intelligence; financial advice; assest market experiment; mispricing; Human-AI interaction |
| JEL: | C90 C91 G12 |
| Date: | 2026–09–15 |
| URL: | https://d.repec.org/n?u=RePEc:wai:econwp:26/05 |
| By: | Moore, Alexander K.; Li, Bingqing; Wang, Shane; Thomson, Matt |
| Abstract: | Large language models (LLMs) are increasingly adopted in the social sciences as surrogates for human coders. Standard practice validates them by their correspondence with human measurements of the same stimuli using statistics like correlation. However, overall correspondence can conceal biases that distort specific claims. We demonstrate this with open-weight LLM ratings of 568 food and 1, 811 scene photographs. Ratings inflate, attenuate, or reverse effects, and improving overall correspondence with prompts can improve one contrast while worsening another. Rankings also elevate some categories and demote others beyond what random errors of the same size typically produce. Across six prompts, some of these errors are shared, so sampling across prompts may continue to mask biases. We propose validating LLM coders against specific claims they will test rather than the general coding task. |
| Date: | 2026–09–21 |
| URL: | https://d.repec.org/n?u=RePEc:osf:socarx:dz3qe_v1 |
| By: | Melissa Dell; Ashesh Rambachan |
| Abstract: | Artificial intelligence (AI) is transforming measurement in economics. AI models convert unstructured data, such as text and images, into structured variables at low cost, making previously prohibitive measurement feasible at scale. This shifts the bottleneck from finding any scalable measure of a phenomenon to choosing among many plausible ones, which may support different empirical conclusions. This review provides guidance for navigating that shift. We describe three stages at which AI enters the measurement pipeline—discovery, construct definition, and observation—and what each demands of researchers. We argue that credible inference with AI-generated variables requires appropriately designed validation: anchoring measurement to explicit criteria, rather than informal claims that a proxy is reasonable. We then examine how validation samples support valid inference even when AI predictions are arbitrarily biased, and what can be done when a random validation sample is unavailable. |
| JEL: | C1 C45 |
| Date: | 2026–09 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:35744 |
| By: | Doruk Cetemen; Emre Ozdenoren |
| Abstract: | Early-career work both signals a worker's ability and builds long-run skills. We introduce an AI industry into a career signaling model. The industry sells tokens, which workers convert into output indistinguishable from their own. We compare three regimes: no token trade, purchases only, and open trade with side sales. With no token trade, market learns every type, signalling incentives create a rat race and workers work above the full information level. With token purchases only, lower types buy output but the market still learns every type. Buying by lower types raises the signal higher types must post, so the high types can work harder than before AI. Buyers work less with AI, even below the full-information level. With side sales and abundant AI, the signal no longer depends on effort and every type under-invests in skill. Workers gain from token trade only when welfare improves, never the reverse, and the difference is the AI industry's profit. In education, where output has no value, every worker loses in the separating equilibrium, but with valuable output and cheap tokens everyone gains. Token taxes, wage taxes, and watermarking have ambiguous effects on workers' payoffs and welfare, and their effects depend on whether workers can sell output on the side. |
| Date: | 2026–09 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2609.17752 |
| By: | Iñaki Aldasoro; Sebastian Doerr; Daniel Rees |
| Abstract: | Artificial intelligence (AI) and robots have the potential to offset the macroeconomic consequences of an ageing workforce, but the ability to automate jobs depends on which industries are ageing. AI and robots substitute most readily for jobs in industries with younger workforces (eg finance), while older, high-employment industries (eg agriculture, health) have less scope for automation. A new ageing–automation scorecard for over 130 economies shows that some of the fastest-ageing economies may face more substantial challenges in easing demographic pressures through automation. |
| Date: | 2026–08–26 |
| URL: | https://d.repec.org/n?u=RePEc:bis:bisblt:135 |
| By: | Drydakis, Nick |
| Abstract: | This study examines whether signalling AI-related digital skills improves employment outcomes for women from underrepresented groups in England, defined by race, age, sexual orientation, and autism-spectrum disclosure. Using correspondence evidence, the study finds that underrepresented women receive fewer interview invitations and are considered for lower-paid vacancies than majority-group women. In pooled analyses, signalling AI-related digital skills increases interview invitations for underrepresented applicants, but does not eliminate the disadvantage. These findings are consistent with AI Capital and productivity-signalling frameworks, as employers appear to value AI-related capabilities while the returns to such credentials remain constrained by persistent demographic inequalities. The study therefore shows that positive returns to AI-related skills and labour-market disadvantage can coexist. Its broader implication is that digital upskilling can strengthen the recruitment prospects of underrepresented women, but cannot by itself deliver parity in employment outcomes. A dual policy response is therefore required, combining wider and more equitable access to AI education and training with stronger anti-discrimination enforcement, greater transparency in shortlisting, improved oversight of recruitment processes, and systematic evaluation of how employers recognise applicant credentials. The study contributes to the economics of AI by linking correspondence evidence to the AI Capital framework and highlighting barriers to converting AI-related resources into meaningful employment opportunities. |
| Keywords: | AI Capital, Artificial Intelligence, Skills, Discrimination, Hiring, Wages, Sexual Orientation, Race, Age, Autism Spectrum |
| JEL: | J71 J24 J31 J64 C93 J15 J16 J14 O33 M51 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:glodps:1823 |
| By: | Oleš, Tomáš; Roncone, Francesco |
| Abstract: | Does artificial intelligence automate work or augment workers' productive capabilities? Our results show that it can do both, through distinct channels that conventional exposure measures obscure. We distinguish these channels by linking U.S. patents to occupations' tasks and tools, with patent-task matches capturing automation and patent-tool matches capturing augmentation. Using the origins and diffusion of historical AI breakthroughs to predict contemporary local exposure, we find large but opposing labor-market effects. A one-standarddeviation increase in AI automation reduces occupational employment by 21 percent, with little effect on wages; AI augmentation instead raises wages by 8 percent, alongside some employment growth. We also find that a larger AI share within each channel strengthens labor displacement through automation while shifting augmentation's effects from employment growth toward wage gains. The labor-market consequences of AI therefore depend critically on the balance between automation and augmentation rather than overall exposure alone. |
| Keywords: | Artificial Intelligence, Automation, Augmentation, Local Labor Markets, Occupational Tasks, Occupational Tools |
| JEL: | J23 J24 J30 O33 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:glodps:1821 |
| By: | Robert Fairlie; Jane Wu; Robert W. Fairlie |
| Abstract: | The impact of AI on the employment prospects of recent college graduates is hotly debated with no consensus on the magnitude of impacts nor even the timing of those potential impacts. Using CPS microdata, we provide the first estimates of the effects of AI on the unemployment of recent college graduates in June, July and August 2026. We provide evidence suggesting that unemployment rates are especially high for summer months and that 2026 might be the first year of widespread enough AI use in the workplace to detect impacts of AI on recent college graduates, the group argued to be most vulnerable to AI replacement. Taking an agnostic approach to defining treatment timing, we find that unemployment rates did not spike in summer 2026 relative to summer months in previous years and did not rise in a significant way relative to older college graduates or young workers without a college degree. We also provide the first analysis of an expanded definition of unemployment that includes those who report “wanting a job” which adds nearly two percentage points to the unemployment rate of recent college graduates but we find no evidence of a statistically significant increase in summer 2026 even after adding these “sidelined unemployed.” We estimate difference-in-differences and event-study interaction models using both older college graduates and young non-college graduates as comparison groups and do not find evidence of an increase in relative unemployment rates. Finally, we estimate models in which we interact 2026 unemployment with AI exposure and remote work availability by occupation and find some evidence of a positive relationship with remote work availability. |
| Keywords: | AI, technological change, unemployment, job loss, college graduates |
| JEL: | J21 J24 J63 J64 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_12994 |
| By: | Asik Erpek, Gunes Arkadas; Hari, Siddharth; Ozen, Efsan Nas; Raju, Dhushyanth |
| Abstract: | Artificial intelligence (AI) has attracted growing attention as a potential source of productivity growth and structural transformation in emerging economies. Yet the labor market implications of AI depend not only on technological capabilities but also on how those capabilities interact with firm behavior, workers’ skills, and institutional conditions. This paper provides a systematic diagnostic of how AI-related technological potential is distributed across Türkiye’s labor market and examines the conditions under which that potential may translate into observable labor market outcomes. The analysis applies occupation-level measures of AI exposure and task complementarity to nationally representative labor force survey data and employer-employee administrative records. It documents how AI-related technological potential is distributed across workers, jobs, firms, sectors, and regions. The results show that approximately 41 percent of employment in Türkiye is in occupations with relatively high exposure to AI, although the scope for productivity-enhancing complementarity varies substantially across occupations and worker groups. Given the early stage of AI diffusion and the rapid evolution of AI capabilities, these estimates should be interpreted as a descriptive mapping of technological exposure and potential complementarity rather than as forecasts of realized employment, wage, or product ivity effects. Despite this widespread exposure, employment and labor demand patterns across exposure categories have remained largely stable in recent years, suggesting that the translation of technological potential into observable labor market effects depends critically on adoption dynamics and complementary adjustments within firms and labor markets. Drawing on additional evidence on firm adoption, hiring patterns, and labor market structure, the paper argues that the economic consequences of AI are likely to unfold gradually through changes in task organization, skill demand, and production processes rather than through abrupt economywide disruption. By distinguishing technological exposure from complementarity and adoption capacity, the paper provides an analytical framework for interpreting emerging evidence on AI and work. It also highlights the policy margins most relevant to shaping labor market adjustment in middle-income economies. |
| Date: | 2026–07–24 |
| URL: | https://d.repec.org/n?u=RePEc:wbk:hdnspu:213012 |
| By: | Alemán-Pericón, Christian |
| Abstract: | I study how human validation requirements shape the response of wages and labor demand to AI adoption. I introduce fixed, recurring validation labor into a general equilibrium task model with endogenous AI adoption. It features a validation threshold determining whether the labor created by AI adoption offsets the productive labor it displaces after accounting for demand expansions. I estimate the model using U.S. occupation-task data on human versus AI involvement at work, and AI adoption rates. The estimated U.S. validation requirement lies below the threshold, implying that productive labor displacement dominates and wages fall. For the occupations in the estimation sample, a ten year projection of declining AI costs raises adoption from 18 to 62 percent and nearly triples output, while reducing employment by 13 percent and wages by 43 percent. Extending the exercise to the entire U.S. economy attenuates the employment and wage declines to 5 and 18 percent, respectively. |
| Keywords: | Artificial Intelligence, Technology Adoption, Validation Labor, Labor Demand, Task-Occupation Labor |
| JEL: | E17 E24 J22 J23 J31 O33 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:esprep:343972 |
| By: | Andrea Gazzani (Bank of Italy); Filippo Natoli (Bank of Italy) |
| Abstract: | Using detailed data on the artificial intelligence (AI) content of U.S. patents from 1980 to 2019, we construct a novel monthly measure of AI intensity in innovation and identify AI related technology shocks. These shocks generate delayed increases in total factor productivity, output, employment, hours, and wages, alongside persistent declines in consumer prices, consistent with a positive supply shock. Although AI-intensive patents largely fall within the broader ICT and automation domains, they stand out as higher-quality innovations, attracting more citations and exhibiting greater technological and market value. AI shocks consistently have substantially larger aggregate effects than broader ICT or automation shocks, suggesting that AI has had a particularly high-impact role in the ICT revolution. Unlike general technology shocks, however, AI shocks reduce the labor share and increase wealth inequality, indicating that their gains are not distributed evenly across the economy. To connect these historical findings to the most recent wave of innovation, we develop a new time series of Generative-AI patents within the U.S. patent universe. A shock based exclusively on Gen-AI patents produces the same qualitative supply-side effects as broader AI shocks, linking the macroeconomic consequences of earlier AI innovation to those of the emerging Generative-AI era. |
| Keywords: | artificial intelligence, technology shocks, local projections, business cycle, patents |
| JEL: | E32 C32 C36 O33 O34 |
| Date: | 2026–07 |
| URL: | https://d.repec.org/n?u=RePEc:bdi:wptemi:td_1542_26 |
| By: | Lucia Esposito (Bank of Italy); Elisa Guglielminetti (Bank of Italy); Elia Moracci (Bank of Italy); Andrea Papetti (Bank of Italy); Massimiliano Pisani (Bank of Italy) |
| Abstract: | This paper assesses the impact of artificial intelligence (AI) on monetary policy transmission and on central banks' reaction functions. By speeding up routine and repetitive tasks through automation and accelerating the incorporation of incoming information into decision-making, AI may affect productivity, market structure and broader macroeconomic developments. AI may also increase macro-financial complexity and systemic vulnerabilities by amplifying interconnectedness, heightening operational and cyber risks, and reinforcing procyclical and herding dynamics in an environment characterized by faster and partly algorithmic decision-making. Central banks' reaction functions will need to adapt to a setting marked by a different transmission mechanism, new uncertainty surrounding the natural rate of interest and tail risks. At the same time, AI could enhance the effectiveness of monetary policy by improving assessments of the macroeconomic outlook and the distribution of risks around it, as well as by strengthening monetary policy communication and the management of expectations. |
| Keywords: | artificial intelligence, credit markets, financial markets, inflation, business cycle, monetary policy |
| JEL: | O33 E44 E32 E31 E52 |
| Date: | 2026–07 |
| URL: | https://d.repec.org/n?u=RePEc:bdi:opques:qef_1051_26 |
| By: | Chakraborty, Lekha (National Institute of Public Finance and Policy); Prasanth C. (Christ (Deemed to be University), Bengaluru) |
| Abstract: | This paper investigates whether AI adoption has induced structural changes in the determinants of Indian sovereign bond yields across the maturity spectrum. The rapid global adoption of artificial intelligence (AI) in times of macroeconomic turmoil, particularly generative AI technologies since late 2022, has prompted intense debate about its potential macroeconomic consequences, including the measurement issues. Using monthly data from 2000 to 2025 and autoregressive distributed lag (ARDL) models augmented with an AI dummy variable and slope interactions on expected inflation and broad money (M3) growth, we identify significant regime shifts. Results indicate that in the post-AI period, longer-maturity yields exhibit markedly reduced sensitivity to expected inflation and money supply growth. This dampening is statistically significant, with interaction terms largely offsetting baseline positive elasticities. By contrast, short-term yields (91-day Treasury bills) show heightened inflation sensitivity in the AI era, while intermediate yields display mixed patterns. These findings are consistent with theoretical predictions that AI-driven productivity gains could lower equilibrium real interest rates and weaken traditional monetary transmission channels at the long end of the yield curve. For an emerging market like India, where inflation expectations have historically influenced borrowing costs, such changes may enhance monetary policy independence but complicate fiscal-monetary coordination. The paper contributes to the sparse empirical literature on AI’s financial market implications in emerging economies. Policy implications include the need for the Reserve Bank of India (RBI) to recalibrate forward guidance and liquidity operations in light of evolving yield dynamics. |
| Keywords: | Sovereign bond yields ; artificial intelligence ; ARDL bounds testing ; structural change ; monetary transmission ; India |
| JEL: | E43 E44 O33 G12 C22 |
| Date: | 2026–06 |
| URL: | https://d.repec.org/n?u=RePEc:npf:wpaper:26/450 |
| By: | Marco Albori (Bank of Italy); Andrea Zaghini (Bank of Italy) |
| Abstract: | We study how the emergence of generative artificial intelligence (AI) affected corporate bond pricing. Using bond-level issuance data for U.S. non-financial corporations and a difference-in-differences design centered on the launch of ChatGPT, we examine whether debt investors differentiate firms according to their position in the AI value chain. We find that hyperscalers experienced a significant decline in borrowing costs relative to comparable issuers, while software firms faced less favorable financing conditions. The results suggest that credit markets rapidly incorporated expectations regarding the distribution of gains and losses from generative AI. More broadly, the evidence indicates that technological shocks affect not only firm valuations but also the cost of external finance, thereby influencing the allocation of capital across firms during periods of technological transition. |
| Keywords: | artificial intelligence, asset pricing, bonds, corporate finance |
| JEL: | G12 G32 C23 |
| Date: | 2026–09 |
| URL: | https://d.repec.org/n?u=RePEc:bdi:opques:qef_1057_26 |
| By: | Kemal Kirtac |
| Abstract: | Financial language models can transform unstructured firm-specific news into structured decision signals, but financial AI research lacks an integrated deployment framework for evaluating whether those signals remain useful in financial decision systems. Computer science research has developed strong methods for time-series forecasting, text classification, multimodal stock prediction, graph-based market modeling, and machine-learning operations, yet these streams do not provide a domain-specific protocol that jointly tests financial language-model outputs under event-time observability, probability calibration, execution timing, transaction costs, liquidity constraints, capacity limits, operational diagnostics, and statistical inference. We introduce MFAST, a Market-Friction-Aware Sentiment-to-Trading framework that converts timestamped financial text into auditable, reproducible, and market-feasible trading decisions. The application is news-based trading, where firm-specific text must be linked to securities before portfolio decisions can be evaluated. The framework links Refinitiv News Analytics to Center for Research in Security Prices (CRSP) equity data, restricts the primary out-of-sample evaluation to post-release news outside disclosed foundation-model data-freshness periods, and adds a public replication arm using open financial text and public price data. Results show that decoder-only language models outperform encoder baselines and dictionary sentiment in classification, calibration, return prediction, and net portfolio performance, while operational diagnostics reveal trade-offs among accuracy, latency, memory, throughput, and inference cost. The paper shows that credible evaluation of financial language models requires an end-to-end engineering approach combining language understanding, temporal discipline, market-friction-aware deployment, and reproducible validation. |
| Date: | 2026–09 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2609.23703 |