The AI That Learned to Read Wall Street Without Breaking the Bank
Every trading day, the world’s equity markets produce an avalanche of headlines, earnings summaries, and macroeconomic commentary, and buried inside that torrent are the signals that quietly move prices. For several years, the pressing question in quantitative finance has been whether artificial intelligence can reliably extract those signals from raw text. That question has largely been answered in the affirmative. The harder question—whether anyone can afford to run such systems at scale—has remained stubbornly open. A study published on 29 August 2026 in the journal Complex & Intelligent Systems confronts it directly. Yukai Su and Lailong Luo of the College of System Engineering at the National University of Defense Technology in Changsha, China, together with Hui Chen of the School of Computing at Macquarie University in Sydney, Australia, introduce MENAP—Multimodal Efficiency-constrained News-Augmented Pricing—a framework that trains large language model agents to interpret financial news with substantially fewer tokens and refinement steps, while maintaining or even improving the economic quality of the trading decisions that flow from those interpretations.
The intellectual foundation of MENAP is the recognition that financial signals are inherently multimodal. Textual news captures the narrative dimension of markets—what companies said, what policymakers signaled, which shocks unfolded overnight. Market and macroeconomic factors encode the quantitative state of the system: where interest rates sit, how volatility behaves, which sectors are running hot. Auxiliary metadata adds texture that neither stream provides on its own. Together, these inputs offer complementary views of investor attention and risk, the two quantities that sit at the heart of asset pricing, and a model that fuses them sees a richer world than one restricted to any single source. In MENAP, these heterogeneous streams are jointly interpreted, so a headline about a policy shift is read in the context of prevailing macro conditions. The challenge is that exploiting them at scale increasingly relies on large language model agents, whose strengths—multi-step reasoning over messy inputs and natural-language synthesis of disparate evidence—are precisely the capabilities that carry the steepest computational price tag.
MENAP plugs its optimized agent into a well-established architecture known as the news-to-state-to-pricing-to-portfolio pipeline. In this canonical design, textual news and its companion signals are first translated into a compact representation of the market’s state; that state description is then passed to a pricing model, which estimates expected returns and mispricings; and those estimates ultimately drive the construction of portfolios. The pipeline is attractive because each stage is modular and auditable. But in its modern, LLM-powered incarnation, the first stage has become the dominant cost center. An agent asked to interpret a day’s multimodal inputs does not simply read and summarize. It plans, cross-references the news against market and macro conditions, drafts an interpretation, critiques its own output, and revises—sometimes repeatedly. The authors formalize this behavior by treating each daily interpretation as a trajectory: a chain of reasoning steps and generated text whose total token count and number of refinement steps translate directly into real computational expenditure.
Left to its own devices, such an agent tends toward verbosity, and the incentive structure of language modeling explains why. Longer, more hedged, more elaborated answers statistically improve benchmark scores even when the marginal informational value of the extra words approaches zero. In a research setting, that verbosity is a nuisance; in live trading, it becomes a structural disadvantage. A pipeline that reinterprets the news every day multiplies the agent’s per-run cost by the trading calendar, and the bill compounds further across strategies, asset universes, and backtests. Every redundant sentence the agent writes is paid for at inference prices, twice over if the critique-and-revise loop forces a second pass. It is this economic reality—not any deficiency of model intelligence—that MENAP was designed to address, by making efficiency an explicit, trainable objective rather than an afterthought bolted on at deployment time.
The framework’s central mechanism is offline preference optimization with efficiency regularization. Rather than letting the agent explore and learn from live feedback—an expensive proposition when every exploratory step consumes tokens—MENAP performs its optimization against a fixed set of trajectories. Each daily multimodal interpretation is treated as a trajectory, and preferred interpretations are held up as exemplars while dispreferred ones are pushed down in likelihood. The innovation lies in what happens to the preferred set: they are further regularized by efficiency rewards, explicit penalty terms on total token usage and on the number of refinement steps a trajectory consumes. Learning therefore pulls in two directions at once—toward interpretations that are accurate and informative, and toward interpretations that reach their conclusions economically. The intended outcome, and the one the experiments bear out, is an agent that produces concise yet informative multimodal summaries: short enough to be cheap, and rich enough to keep the downstream pricing model fully informed.
The mechanics of the reward structure deserve attention because they encode a delicate balance. In standard preference optimization, the learning signal is comparative: among candidate interpretations of the same day’s inputs, one is designated preferred and another dispreferred, and the model’s parameters are adjusted to raise the probability of the former relative to the latter. MENAP extends this recipe by weighting the preferred side with efficiency rewards, so that a trajectory delivering the same informational conclusion in fewer tokens, or with fewer self-correction loops, is favored over an equally accurate but bloated alternative. Calibration matters enormously: penalize tokens too aggressively and the agent collapses into terse summaries that discard the very signals the pipeline exists to capture; penalize too lightly and the verbosity problem survives intact. The reward design explicitly encourages summaries that remain informative about the full multimodal input—news text, market and macro factors, and metadata—even as their length and step count shrink. The offline character of the procedure also means the heavy lifting—generating and comparing candidate trajectories—happens once during training rather than continuously in production.
Perhaps the most consequential design decision in MENAP is what its authors chose not to change. The downstream mixed pricing network—the component that converts the agent’s summaries into pricing signals—and the entire evaluation protocol remain completely untouched. This conservatism carries two payoffs. Scientifically, it enables fair comparisons: any difference in performance between MENAP and a costlier baseline can be attributed to the quality and efficiency of the news interpretation itself, rather than to compensating adjustments elsewhere in the system. Practically, it enables straightforward deployment, because an organization adopting MENAP can slot the efficiency-optimized agent into an existing pipeline without re-engineering its pricing models, its portfolio construction logic, or its risk controls. It also keeps the science honest: with the evaluation protocol frozen, improvements cannot be manufactured by quietly relaxing how success is measured. In an industry where a single misconfigured day of live trading carries real financial consequences, that plug-and-play property is more than a convenience; it is the difference between a laboratory curiosity and a technology a trading desk could actually contemplate adopting.
The empirical evaluation follows a deliberately standard construction, a choice that makes the results legible against prior literature. The team assembled two years of Wall Street Journal news, spanning 29 September 2021 to 29 September 2023, and aligned it with daily stock returns drawn from CRSP, the Center for Research in Security Prices. Market and risk-free rates came from the Ken French library, and the macroeconomic factors followed the specification of Jensen and colleagues, giving the pricing side of the experiment a modern multi-factor foundation. The data were divided with a fixed split: nine months for training, three months for validation, and one full year for testing. The period itself was no gentle test bed: it spanned rapid interest-rate hikes and a punishing equity bear market, conditions that stress-test any model’s stability. That final untouched year matters enormously, because it subjects the optimized agent to an extended stretch of market conditions it never saw during training, approximating the out-of-sample discipline of live deployment while allowing an unusually thorough assessment of both performance and cost.
The evaluation is two-pronged, reflecting the twin demands of quantitative practice. On the portfolio side, the researchers measured the Sharpe ratio and maximum drawdown for three portfolio types—TP, EW, and VW—capturing risk-adjusted return and worst-case peak-to-trough loss across different weighting schemes. On the pricing side, they examined average absolute alpha, the magnitude of excess return left unexplained by the model, together with its t-statistics and the GRS test computed across 78 anomaly portfolios, a broad panel widely used to determine whether a pricing model leaves systematic return patterns unexplained. The GRS statistic, a classical test of Gibbons, Ross, and Shanken, asks whether a model’s pricing errors are jointly indistinguishable from zero across the panel, making it a stringent arbiter of pricing skill. In parallel, the study logged LLM efficiency metrics—tokens consumed and refinement steps taken—turning the computational bill into a measured experimental outcome rather than an anecdote. Reporting both dimensions with equal rigor is uncommon in this literature, and the pairing is exactly what allows the effectiveness-efficiency trade-off, the study’s true subject, to be quantified rather than merely asserted.
The results, the authors report, show that MENAP delivers a superior trade-off between effectiveness and efficiency: it reduces inference cost while maintaining or improving economic outcomes. The efficiency-constrained agent did not purchase its token savings at the expense of signal quality. Its concise summaries remained informative enough that the unchanged pricing network produced portfolios performing as well as or better than those driven by costlier, more verbose agents, and the pricing accuracy diagnostics held up across the anomaly panel. The savings registered directly in the reported token and step counts, the metrics that map most closely to real-world inference expenditure. The study, published open access in Complex & Intelligent Systems, arrives as organizations everywhere grapple with the operational cost of LLM-driven workflows, and its template—keep the downstream machinery fixed, reshape the agent through preference optimization with explicit efficiency rewards, and validate on standard data so results remain comparable—extends well beyond finance. The future of AI in markets may belong not to the biggest models, but to the most disciplined ones.
Subject of Research: Cost-aware preference optimization of large language model agents for multimodal, news-augmented asset pricing, balancing signal quality against LLM inference efficiency (token usage and refinement steps) across portfolio performance and pricing accuracy evaluation.
Subject of Research: Technology and Engineering
Article Title: MENAP: multimodal efficiency-constrained news-augmented asset pricing via cost-aware preference optimization
Article References: Su, Y., Chen, H., & Luo, L. (2026). MENAP: multimodal efficiency-constrained news-augmented asset pricing via cost-aware preference optimization. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02433-x
Image Credits: AI Generated
DOI: 10.1007/s40747-026-02433-x
Keywords: News-augmented asset pricing, Large language models, Preference optimization, Multimodal financial NLP, Cost-aware preference optimization, Inference efficiency, Efficiency rewards, Token efficiency, Sharpe ratio, Maximum drawdown, GRS test, Anomaly portfolios
Cite Scienmag News
APA MLA Chicago
Denise Maddox. (August 30, 2026). MENAP: efficient news-augmented asset pricing via cost-aware preference optimization. Scienmag. https://scienmag.com/menap-efficient-news-augmented-asset-pricing-via-cost-aware-preference-optimization/
Denise Maddox. “MENAP: efficient news-augmented asset pricing via cost-aware preference optimization.” Scienmag, 30 August 2026, https://scienmag.com/menap-efficient-news-augmented-asset-pricing-via-cost-aware-preference-optimization/. Accessed 30 August 2026.
Denise Maddox. “MENAP: efficient news-augmented asset pricing via cost-aware preference optimization.” Scienmag. August 30, 2026. https://scienmag.com/menap-efficient-news-augmented-asset-pricing-via-cost-aware-preference-optimization/
Copy citation Download RIS
Tags: AI in quantitative financeAI-driven stock market predictionasset pricing optimizationcost-aware AI frameworks for financial marketscost-aware preference optimizationcost-efficient large language modelseconomic quality of AI trading decisionsefficient AI for market analysisefficient AI-driven investment decision-makingfinancial news analysisfinancial news interpretation with minimal tokensFinancial news signal extractionlarge language models in financemultimodal data in asset pricingmultimodal data integration in asset pricingmultimodal financial data interpretationmultimodal financial signal processingnews-augmented trading algorithmsnews-augmented trading strategiesnews-based asset valuationpreference optimization in financescalable AI for stock market analysisscalable quantitative financetoken-efficient language models for trading
