TL;DR for operators

Finance AI is moving past the parlour trick stage. The interesting question is no longer whether a large language model can read a financial headline and produce a plausible answer. Of course it can. The useful question is whether that answer can be converted into a measurable, governed, risk-aware decision process without accidentally building a very expensive rumour amplifier.

Three recent papers point to a practical answer. One paper studies how to fine-tune and evaluate financial LLMs on finance-specific tasks.1 A second paper inserts LLM-generated news recommendation and risk scores into reinforcement-learning trading agents.2 A third paper builds a simulated stock market populated by LLM-based trading agents, allowing researchers to observe how heterogeneous agents react to shocks.3

The combined lesson is not “LLMs can trade”. That is the shiny but slightly unserious reading. The better reading is this:

Financial LLMs become useful when they are treated as one layer in a decision stack: benchmarked model capability, structured signal extraction, risk-aware action, and market-level simulation.

For business leaders, that means the near-term opportunity is not fully autonomous portfolio management. It is decision support: faster financial text processing, interpretable signal generation, scenario testing, stress simulation, and stronger risk governance around human and machine decisions.

The awkward part, naturally, is that this requires more engineering discipline than simply adding “AI-powered” to a deck and hoping procurement has had a long lunch.

The problem: finance does not need cleverer chat; it needs controlled judgement

Finance is full of language: earnings calls, central bank minutes, analyst notes, regulatory filings, policy statements, breaking news, and the ritual theatre of executive guidance. Large language models are good at language, so the attraction is obvious.

But finance is not merely a reading-comprehension contest. It is a decision environment where the cost of being confidently wrong compounds. A model that summarises a headline well may still produce a poor trading signal. A signal that works in one backtest may fail in another regime. A trading agent that performs well alone may behave oddly when copied across a market. And a market full of similar agents can create feedback loops that are not visible when each model is evaluated in isolation.

That is why these three papers are best read as a chain rather than as separate artefacts. They occupy different layers of the same emerging architecture.

Layer Question Paper role Operator takeaway
Foundation model Can the model perform financial tasks reliably enough to produce usable outputs? Fine-tune and evaluate financial LLM capability Do not assume general reasoning equals finance readiness
Decision agent Can financial text become a structured recommendation or risk signal inside a trading policy? Inject LLM signals into PPO and CVaR-PPO trading agents Treat the LLM as a feature generator, not a portfolio deity
Market system What happens when many LLM-like agents interact through market mechanisms? Simulate stock trading with heterogeneous LLM-based agents Use simulation for stress testing and mechanism discovery, not prophecy

This chain is useful because it keeps the argument honest. A FinLLM that improves on a benchmark is not automatically a trader. A trading agent that improves in a backtest is not automatically robust. A simulated market that resembles some known economic patterns is not automatically a real market. Each layer adds capability, but also adds assumptions. Lovely. Finance has discovered software architecture.

Layer one: first make the model financially literate, then worry about decisions

The Open FinLLM paper starts at the most basic layer: financial capability. It fine-tunes Qwen2.5 and DeepSeek-R1 models using supervised fine-tuning, direct preference optimisation, synthetic data, chain-of-thought annotation, and reinforcement-learning-style iterative data generation. The target environment is the Open FinLLM Leaderboard, which covers financial tasks such as information extraction, textual analysis, question answering, generation, risk management, forecasting, and decision-making.

What the paper shows is practical and slightly inconvenient. Supervised fine-tuning improves performance on the task being trained. That is unsurprising, but important. The more useful finding is that transfer is uneven. Fine-tuning on named entity recognition improves named entity recognition, but can hurt causal classification. Fine-tuning on one sentiment-style task can help another sentiment-style task because the structures are more similar. Sequential fine-tuning does not reliably beat direct fine-tuning.

That matters for operators because financial workflows are not one task. A credit analyst may need extraction, classification, summarisation, anomaly detection, reasoning over covenants, and decision support in one pipeline. A model improved for one of those jobs may become worse at another. Model evaluation therefore cannot stop at the leaderboard score. It needs workflow-level testing.

The paper also reports a very human failure mode: after fine-tuning, a model may over-generate or repeat the correct answer tokens rather than stopping cleanly. The authors use DPO to reduce this overlength behaviour. The business translation is straightforward: some “model intelligence” problems are actually output-control problems. In production, prompt design, decoding settings, regular expressions, schema validation, and post-processing may matter as much as another round of training.

That is not glamorous. It is also what makes systems work.

The foundation-layer lesson is this: before a financial LLM is trusted as a source of trading, credit, compliance, or treasury signals, it needs task-specific evaluation and operational containment. It should be judged not only on whether it knows financial vocabulary, but on whether its outputs are short enough, stable enough, parseable enough, and transferable enough for downstream systems.

Layer two: turn language into signals, not commandments

The FinRL-DeepSeek paper moves from model capability to trading-agent design. It extends PPO and CVaR-PPO by adding LLM-generated stock recommendation and risk assessment signals from financial news. The news source is FNSPID, and the paper uses models including DeepSeek V3, Qwen 2.5, and Llama 3.3 to generate stock-level scores.

The important design choice is that the LLM does not directly trade. It produces signals that perturb the trading algorithm.

For PPO, the LLM-generated recommendation score modifies the trading action:

$$ a^{mod}_t = S_f \cdot a_t $$

The paper keeps the perturbation close to 1, so the LLM nudges the action rather than hijacking the policy. Positive recommendations can amplify a compatible buy or sell action; negative recommendations can dampen it.

For CVaR-PPO, the LLM generates risk scores, which affect the trajectory return:

$$ D_{R_f}(\pi_\theta) = R_f \cdot D(\pi_\theta) $$

This is the right mental model for business use. The LLM is not the trader. It is a structured alternative-data processor feeding a risk-aware decision system.

The results are more useful because they are mixed. In one test, integrating LLM stock recommendations improves PPO cumulative returns but does not beat the Nasdaq-100 benchmark. In another test with a longer training and trading period, the use of LLM signals worsens performance. After longer training, PPO and CPPO-DeepSeek can outperform other methods in one stochastic run, but the pattern is regime-sensitive: PPO appears stronger in bull markets, while CPPO-DeepSeek appears more useful in bear conditions. When the paper varies LLM infusion strength, stronger LLM influence mostly degrades PPO performance, while stronger infusion improves CPPO-DeepSeek in the reported variants.

This is precisely the sort of evidence business readers should prefer. Not because it is triumphal, but because it is diagnostic. The result says: “The value of LLM signals depends on the surrounding risk architecture, the regime, the training window, the signal quality, and the strength of intervention.” That sentence will not sell as many conference tickets as “AI beats Wall Street”, but it has the advantage of being less silly.

The paper’s future-work section reinforces the operational point. Longer training consumes much more memory. Faster market reaction requires shorter decision timescales. Better performance depends on improving news-signal quality. These are engineering and data-quality constraints, not mystical limitations of machine intelligence.

For operators, the lesson is clear: use LLMs to produce typed, auditable financial signals. Then test those signals inside conservative decision policies. Measure the effect of signal strength. Compare against benchmarks. Split by market regime. Assume degradation is possible. In short, behave like adults. I appreciate this is a burden.

Layer three: once agents act, markets become systems again

The ASFM paper moves one layer higher. Instead of asking whether one model can produce better financial outputs, it asks what happens when many LLM-based agents participate in a simulated stock market.

The paper constructs an agent-based simulated financial market with listed companies, an order-matching mechanism, stock price updates, and LLM-based trading agents. Agents have profiles, observations, and tool-based actions: buy, sell, or hold. Their observations include recent stock prices, order-book history, and economic policy news. Their profiles encode different strategies such as value investing, institutional investing, contrarian investing, and aggressive investing.

This is not a trading-bot paper in the usual sense. It is closer to experimental economics without the expensive human-subject logistics. The authors use the system to test controlled scenarios: interest-rate cuts, inflation shocks, behavioural composition, and large-trader effects.

The paper reports that a rate-cut shock produces rising average stock returns in the simulation, consistent with standard expectations about expansionary policy. It also reports sectoral differences, such as real estate and finance responding differently. Under inflation shocks, both very high and very low inflation damage returns, while moderate inflation produces better simulated outcomes. In behavioural-composition experiments, a market of only value investors shows lower volatility, while a market of aggressive investors produces higher order quantity, turnover, and volatility. In the large-trader experiment, smaller traders show higher return growth over the simulated period, while large traders are less flexible because their orders are harder to match.

Again, the business interpretation must be controlled. This does not prove that the simulator predicts real markets. It shows that LLM-based agents can be placed inside a market mechanism and produce behaviours that resemble some known economic intuitions. That is weaker than prophecy, but still useful.

Why? Because many business decisions are not about predicting one exact price. They are about exploring plausible reactions under uncertainty.

A bank can ask how different customer segments might react to interest-rate changes. A trading platform can test how agentic strategies behave under liquidity stress. A regulator can explore how market composition affects volatility. A treasury team can use a synthetic environment to test policy narratives before committing to a real-world hedge. None of these uses requires the simulator to be a crystal ball. They require it to be a disciplined scenario engine.

That is the larger point. Once financial AI agents are allowed to act, even in simulation, the object of analysis changes. You are no longer evaluating a model answer. You are evaluating a system.

The chain: capability, signal, action, market

Read together, the papers suggest a practical architecture for FinLLM adoption.

Step What the papers show Business interpretation Main control question
1. Adapt the model Financial fine-tuning can improve target tasks, but transfer is uneven Build task-specific capability, not generic “finance reasoning” theatre Which tasks improved, and which degraded?
2. Structure the output LLMs can extract recommendation and risk scores from news Convert unstructured text into auditable features Are the scores stable, calibrated, and parseable?
3. Insert into decisions LLM signals can help or hurt trading agents depending on architecture and regime Use LLMs as controlled signal modules inside risk-aware systems How much influence should the signal have?
4. Simulate interaction LLM-based agents can interact through a market mechanism under shocks Use multi-agent environments for stress testing and scenario design What behaviours emerge when many agents react together?
5. Govern the stack Each layer introduces new assumptions Manage the pipeline, not just the model Where can failure propagate?

The central business takeaway is that financial AI should be built like a controlled pipeline, not like a chatbot with a brokerage account.

A sensible stack might look like this:

  1. Financial model layer Fine-tune or prompt models for specific financial tasks. Evaluate across extraction, sentiment, risk, forecasting, and decision support. Track transfer failures.

  2. Signal layer Convert text into structured fields: event type, sentiment, risk score, recommendation score, uncertainty level, evidence span, time sensitivity, and affected asset.

  3. Decision layer Feed those fields into existing models or reinforcement-learning agents with strict caps on influence. Monitor whether LLM-derived features improve risk-adjusted performance, not just raw returns.

  4. Simulation layer Test multiple agent types under policy shocks, news shocks, liquidity constraints, and behavioural mixes. Look for fragility, crowding, and regime dependence.

  5. Governance layer Log prompts, model versions, signal distributions, action changes, rejected outputs, and human overrides. Without this, “AI governance” becomes a PDF stored in a folder named Final_v7.

What the papers show versus what operators should infer

It is worth drawing a hard boundary between evidence and interpretation.

What the papers show

The papers show that financial LLMs can be improved through domain-specific tuning and data synthesis, but that task transfer is uneven. They show that news-derived LLM signals can be inserted into risk-aware trading agents, but that performance effects are mixed and architecture-dependent. They show that LLM-based agents can populate a simulated stock market and react to shocks in ways that align with some economic expectations.

What they do not show

They do not show that LLMs can reliably beat the market in live trading. They do not show that benchmark performance translates automatically into portfolio performance. They do not show that simulated market behaviour is causally valid for real markets. They do not show that more LLM influence is always better. In one of the papers, more influence is explicitly worse for one class of agent.

What business readers can reasonably infer

The reasonable inference is architectural. Financial AI is becoming modular. The model reads, extracts, scores, and explains. A decision engine acts under constraints. A simulator explores system-level consequences. Human operators define objectives, monitor failures, and decide where automation is permitted.

The value is not that the LLM “knows finance”. Many people know finance and still manage to lose money with admirable efficiency. The value is that a properly governed LLM can process financial language at scale, produce structured signals, and feed those signals into systems that can be tested.

The misconception to kill early: reasoning is not readiness

The likely reader misconception is that finance AI improves mainly by making the base model smarter. Better reasoning, better answers, better trades. A neat little staircase into alpha. How comforting.

The papers suggest a more annoying reality. Readiness depends on fit between task, data, output format, decision architecture, and market context.

A general model may be overqualified in the abstract and underprepared in the workflow. It can discuss monetary policy fluently but fail to produce a clean label. It can generate a sensible risk explanation but destabilise a trading policy if its signal is overweighted. It can simulate an investor profile but still require a carefully designed market mechanism before its behaviour means anything.

This is why FinLLMs matter more than reasoning in the narrow business sense. Not because reasoning is irrelevant, but because finance work is deeply procedural. It needs validated tasks, controlled outputs, audit trails, risk thresholds, and feedback loops. The cleverness of the model is only one component. Sometimes it is not even the bottleneck.

A practical operator framework

For organisations exploring FinLLM systems, the three-paper chain suggests five questions.

1. What task is the model actually doing?

“Financial analysis” is not a task. Named entity recognition is a task. Classifying central-bank language as hawkish or dovish is a task. Extracting covenant terms is a task. Generating a risk score from a news item is a task. Ranking the likely impact of a policy shock across sectors is a task.

The more precisely the task is defined, the easier it becomes to evaluate whether fine-tuning, prompting, retrieval, post-processing, or plain old rules are the right tool. Yes, rules still exist. They are unfashionable, not dead.

2. Is the output designed for a downstream system?

A beautiful paragraph is useless if the next system needs a number, a label, or a JSON field. Financial LLMs should often produce structured outputs with confidence, evidence, and failure states.

A production signal should answer:

  • What asset, entity, or sector does this relate to?
  • What is the event?
  • What is the expected direction?
  • What is the risk score?
  • What evidence supports the score?
  • How stale is the information?
  • Should the system abstain?

If the model cannot abstain, it is not cautious. It is merely verbose.

3. How much influence is the LLM allowed to have?

The FinRL-DeepSeek paper is valuable because it treats LLM influence as a tunable perturbation. That is the right instinct. In regulated or high-risk environments, an LLM signal should rarely have unlimited authority.

Influence can be capped through position limits, score thresholds, human review, confidence gates, fallback models, and regime-specific rules. The right question is not “Should the LLM decide?” It is “Under what conditions may its signal change the decision, and by how much?”

4. Has the system been tested under bad conditions?

Backtests that only celebrate good periods are marketing brochures with charts. Operators need stress periods, volatility spikes, liquidity limits, stale news, contradictory sources, adversarial headlines, and missing data.

The ASFM-style simulation layer is useful here. Even if the simulation is imperfect, it can reveal how strategies behave when agent composition, policy news, inflation assumptions, or liquidity conditions change. In business, a model that is only tested on polite data will eventually meet reality, and reality is not known for its customer service.

5. Where is accountability logged?

Every layer needs observability: prompt version, model version, input data, extracted signal, confidence, downstream action, rejected action, human override, realised outcome. Without this, organisations cannot tell whether success came from the model, the market regime, the risk engine, or luck wearing a nice suit.

This is especially important in finance because errors are rarely isolated. A bad extraction can become a bad score. A bad score can become a bad trade. A bad trade can become a risk event. A risk event can become a meeting with people who use the phrase “root cause analysis” with visible enthusiasm.

The strategic implication: build the stack before selling the magic

The most useful near-term FinLLM products will probably not be fully autonomous traders. They will be tools that sit inside existing decision processes:

  • research assistants that extract and classify financial events;
  • risk systems that convert news into structured warnings;
  • portfolio tools that test how signals affect decisions under constraints;
  • treasury dashboards that simulate macro shocks;
  • compliance systems that track language changes across filings and regulation;
  • market-intelligence platforms that model heterogeneous participant reactions.

The common feature is not autonomy. It is controlled augmentation.

This may feel less dramatic than replacing analysts. It is also more commercially plausible. Businesses do not need a black-box oracle that occasionally loses money with great confidence. They need tools that reduce processing time, improve consistency, surface risk earlier, and support decisions that can be explained after the fact.

The three papers therefore support a measured thesis: FinLLMs are not valuable because they “reason about markets” in some general sense. They are valuable when they become disciplined components in a financial decision system.

That system begins with domain adaptation. It converts text into structured signals. It tests those signals inside risk-aware agents. It explores interactions in simulated markets. And, if the organisation has any sense of self-preservation, it governs the whole chain.

The future of finance AI is not one giant model shouting “buy”. It is a stack of specialised components, each constrained by the layer above and tested by the layer below. Less cinematic, perhaps. But finance has quite enough drama already.

Sources

Cognaptus: Automate the Present, Incubate the Future.

  1. Varun Rao, Youran Sun, Mahendra Kumar, Tejas Mutneja, Agastya Mukherjee, and Haizhao Yang, “LLMs Meet Finance: Fine-Tuning Foundation Models for the Open FinLLM Leaderboard”, arXiv:2504.13125, 2025, https://arxiv.org/abs/2504.13125. ↩︎

  2. Mostapha Benhenda, “FinRL-DeepSeek: LLM-Infused Risk-Sensitive Reinforcement Learning for Trading Agents”, arXiv:2502.07393, 2025, https://arxiv.org/abs/2502.07393. ↩︎

  3. Shen Gao, Yuntao Wen, Minghang Zhu, Jianing Wei, Yuhan Cheng, Qunzi Zhang, and Shuo Shang, “Simulating Financial Market via Large Language Model based Agents”, arXiv:2406.19966, 2024, https://arxiv.org/abs/2406.19966. ↩︎