A system that routes news, fundamentals, price dynamics, and macro analysis through four separate LLM agents, then synthesises their outputs into a single stock recommendation. That is the pitch behind MarketSenseAI, and the two papers behind it (Fatouros et al., arXiv:2502.00415 and arXiv:2604.17327) claim cumulative returns of 125.9% versus 73.5% for the S&P 100 index over 2023-2024, and a 99.7th-percentile Monte Carlo ranking on S&P 500 strong-buy picks over 19 months. The numbers look impressive. The question that matters for anyone considering multi-agent LLM stock scoring in their own workflow is whether this validation actually holds up under the tests a skeptical trader would run.
I have spent months building and breaking agentic AI nowcasting systems for real-time factor scores, and one thing I have learned is that the architecture diagram is never the hard part. The hard part is proving the system adds value after you subtract transaction costs, adjust for regime, and compare against a baseline simple enough that it cannot overfit. That is the lens for this article: not whether multi-agent LLM systems are exciting, but whether they survive the validation gauntlet.
How MarketSenseAI’s Multi-Agent Architecture Works
MarketSenseAI, developed by Alpha Tensor Technologies, runs four specialist LLM agents in parallel. The News agent handles ticker-specific event analysis. The Fundamentals agent processes SEC filings, earnings transcripts, and quantitative financials via retrieval-augmented generation. The Dynamics agent reads price-action and technical signals. The Macro agent covers sector-level and macroeconomic context from institutional reports.
Each agent produces an independent text analysis for a given stock on a given date. A fifth synthesis agent reads all four outputs and generates two things: a free-text equity thesis and a five-point ordinal recommendation from strong-sell to strong-buy. The system runs on a first-Friday-of-the-month cadence and covers either S&P 100 or S&P 500 stocks depending on the cohort.
This is a different architecture from single-model scoring, where one LLM ingests all information types and produces a score directly. The multi-agent design forces specialisation. Each agent sees only its domain. The synthesis agent sees only the agent outputs, not the raw data. Whether that separation helps or hurts is an empirical question, not an architectural one.
What the Validation Actually Tested
The April 2026 validation paper (arXiv:2604.17327) does something most LLM-finance papers skip: it tests live signals against a proper null. All recommendations were generated at each observation date in real time. No retroactive signal generation. No look-ahead bias from training data leakage, since the S&P 500 cohort runs from September 2024 onward, past the LLM training cutoff.
The core test is a Monte Carlo portfolio comparison. On each observation date, the system’s strong-buy picks are formed into an equal-weight portfolio. Then 10,000 random portfolios of the same size are drawn from the same universe on the same date with equal weighting. The only variable is which stocks are selected. This controls for universe composition, market timing, position count, and sector exposure in expectation.
On the S&P 500 cohort (19 months, 467 stocks), the strong-buy portfolio earned +2.18% per month against a passive equal-weight benchmark of +1.15%, landing at the 99.7th percentile of the Monte Carlo distribution (p=0.003). Compound excess over the equal-weight benchmark was +25.2 percentage points. The strong-buy portfolio beat the benchmark in 11 of 19 months.
The S&P 100 cohort (35 months, 94 stocks) showed +30.5 percentage points of compound excess but did not reach statistical significance (p=0.17), constrained by an average selection of only 10 stocks per month creating wide variance in the null distribution.
Where the Multi-Agent Edge Appears
The most interesting finding is not the raw return but where it comes from. The researchers used non-negative least-squares (NNLS) decomposition to project thesis embeddings onto agent embeddings, recovering per-stock, per-date contribution weights for each agent. If one agent dominated every month, you would not need a multi-agent system. A single fine-tuned model would suffice.
That is not what they found. Agent contributions rotate with market regime. On the S&P 500 cohort, Fundamentals led on 5 of 19 dates, Macro on 6, Dynamics on 5, and News on 3. The Macro agent led during the Fed easing cycle and the US election period. Fundamentals dominated after the “Liberation Day” regime shift. Dynamics appeared episodically when momentum was informative.
I find this rotation pattern more convincing than the headline return. A system that always relies on the same signal source is a dressed-up single-factor model. A system where the dominant signal source changes with market conditions, and that change aligns with identifiable macro events, is doing something closer to what experienced discretionary traders do: shifting emphasis based on what matters right now.
The ordinal recommendation’s cross-sectional Information Coefficient reached ICIR = +0.489 on S&P 500 (p=0.024), confirming the recommendation label carries genuine rank information within the buy and strong-buy subset. The continuous NNLS weights carried even more predictive content than the discrete label, with Fundamentals pooled IC at +0.052 versus the ordinal score’s pooled IC of +0.006.
What the Validation Did Not Test
Here is where a skeptical trader should slow down. The validation has real gaps, and the authors acknowledge several of them.
Transaction costs are absent from the results. The paper notes that at monthly rebalancing across roughly 35 large-cap equal-weight positions, “typical implementation drag (bid-ask spread, market impact) would be well below 30 bps/month.” That is probably right for large-cap US equities, but it is an assertion, not a measurement. A proper net-of-costs backtest with realistic fill assumptions would be stronger evidence.
Regime coverage is thin. The S&P 500 cohort runs September 2024 to March 2026, a period characterised by generally positive equity returns. The system has not been tested through a sustained bear market, a liquidity crisis, or a correlation spike where diversification fails. The below-unity portfolio beta (0.865) and stronger down-market alpha preservation (+1.31% excess in down months versus +0.82% in up months) are encouraging, but 8 down months is not enough data to make a claim about crisis resilience.
Sell-side signals failed. Sell and strong-sell stocks earned positive average one-month returns (+1.65% and +2.98% respectively). The authors suggest short-squeeze dynamics in a risk-on environment. That is plausible. But it means the system’s bearish calls were actively wrong over this period, and short-selling based on its signals would have lost money.
There is no comparison against simple factor baselines. The Monte Carlo test compares against random selection, which is the right statistical null. But traders want to know: does this beat a momentum screen? A quality-value composite? A simple RSI filter combined with earnings revisions? The paper does not run those comparisons. A multi-agent LLM system that cannot beat a three-factor screen sorted by composite factor scores is an expensive noise generator.
The Baseline Problem: What Would Convince a Skeptical Trader
When I evaluate any new scoring system, I start with the simplest baseline that could plausibly explain the results. For MarketSenseAI’s strong-buy picks on S&P 500 large caps during a risk-on period, the obvious baselines are:
First, a momentum screen. Sort S&P 500 stocks by 12-1 month momentum, buy the top decile equal-weight, rebalance monthly. This is a well-documented factor that outperforms in trending markets. If MarketSenseAI’s picks correlate 0.7 or higher with a momentum screen, the multi-agent architecture may be adding marginal value on top of a factor you can run in a spreadsheet.
Second, an earnings-revision screen. Sort by analyst earnings estimate changes over the past 30 days, buy the top quintile. Earnings revisions are one of the strongest short-term return predictors in equity markets. The Fundamentals agent processes earnings transcripts, so there is a real question about whether the “Fundamentals” signal is mostly a proxy for revision momentum.
Third, a combined quality-momentum composite using regime-aware factor screening. If the multi-agent system’s key advantage is shifting factor emphasis with regime, then a simple regime-conditional factor rotation should capture most of that edge at a fraction of the cost.
None of these baselines appear in the MarketSenseAI papers. Until someone runs them, the question of whether the multi-agent architecture adds value over simpler alternatives remains open.
Feature Leakage and the Training Cutoff Question
The validation paper’s strongest design choice is live signal generation. All agent outputs were produced at each observation date, not retroactively. For dates beyond the LLM training cutoff, news, earnings, and price data are absent from the model’s pre-training corpus. This eliminates the most common form of look-ahead bias in LLM-finance studies.
But there are subtler leakage channels. The S&P 100 cohort starts in May 2023 and runs 35 months. The early months of that cohort likely fall within GPT-4’s training window, which raises the question of whether the model “remembers” stock-specific patterns from its pre-training data. The paper acknowledges this: it notes that the S&P 500 cohort (entirely post-cutoff) is the primary result, and the S&P 100 cohort is presented as robustness. That is honest framing.
A harder leakage question involves the retrieval-augmented generation pipeline. The Fundamentals agent uses RAG to process SEC filings and earnings calls. If the retrieval index contains documents with forward-looking analyst consensus embedded in earnings call transcripts, you can get implicit look-ahead through the retrieval layer even when the LLM itself has no future knowledge. The paper does not describe retrieval filtering by date, which would be the standard safeguard. I have seen this exact issue in my own RAG pipelines for financial data: a retrieval hit from a “timely” document that actually contains forward references buried in the Q&A section of an earnings call.
Overfitting in Multi-Agent Systems
Multi-agent architectures introduce overfitting surfaces that single-model systems do not have. Each agent has its own prompt, its own retrieval configuration, and its own output format. The synthesis agent has a prompt that determines how it weighs agent inputs. That is at least five sets of design choices, each of which can be tuned to historical data.
The MarketSenseAI papers do not describe a walk-forward analysis or any form of rolling out-of-sample testing. The system was presumably calibrated before the live period began, and then run forward. If prompt engineering or retrieval configuration was adjusted during the live period based on intermediate results, the live signals are no longer fully out-of-sample.
This is common in deployed systems and not necessarily disqualifying, but it needs to be stated. The paper does not say whether any system parameters were changed during the 19-month S&P 500 evaluation window. A frozen system with no mid-course corrections would be stronger evidence than one that was iteratively improved.
I track my own equity curve behavior for exactly this reason. If a system’s live performance degrades after an initial strong period, that is often the signature of an overfit setup encountering regime change. MarketSenseAI’s month-by-month returns would tell this story, but the paper presents them only in aggregate.
Explainability and the Embedding Attribution Question
The NNLS attribution is one of the more rigorous attempts at explainability in LLM finance research. By decomposing thesis embeddings onto agent embeddings, the researchers can say “this month, the Macro agent contributed 35% of the thesis content” in a way that is mathematically precise. The reconstruction quality is high: mean cosine similarity between thesis and reconstruction is 0.944 on S&P 500.
But embedding attribution is not the same as causal explanation. The NNLS weights tell you which agent’s text is most semantically similar to the final thesis. They do not tell you which agent’s information caused the recommendation to be correct. A Macro agent summary that discusses interest rates might be semantically close to a thesis about a bank stock, but the actual return driver might be an earnings beat that the Fundamentals agent flagged.
For practical use, what traders need is not attribution of text similarity but attribution of return prediction. Which agent’s signal, when it is elevated, predicts positive excess returns? The paper partially answers this through pooled IC analysis (Fundamentals carries the highest positive IC on S&P 500, Macro on S&P 100). But the date-level IC for individual agents is insignificant for all four agents in both cohorts. The prediction comes from the combination, not from any individual agent. That is either a feature (genuine diversification) or a problem (no single agent can be independently validated).
How This Differs from Single-Agent Scoring
The agentic AI nowcasting approach uses a single LLM agent to build real-time factor scores from multiple data sources. The key difference is architectural: one model sees everything versus specialised models seeing slices.
MarketSenseAI’s multi-agent design has one clear advantage in the validation data: the agent rotation pattern. Because each agent operates independently, the synthesis layer can shift emphasis without retraining. A single-model system would need to learn regime-conditional attention internally, which requires either explicit regime labels in the prompt or enough training data to learn the pattern implicitly.
The disadvantage is cost and complexity. Four agents plus a synthesis layer means five LLM calls per stock per observation date. On the S&P 500 at monthly frequency, that is roughly 2,335 LLM calls per month before counting embedding generation and NNLS computation. For a retail trader screening 50 stocks, a single well-prompted model with structured output is probably sufficient and costs 90% less.
The right question is not “multi-agent versus single-agent” in the abstract. It is: at what universe size and rebalancing frequency does the multi-agent overhead pay for itself through improved selection? The papers do not answer this directly, but the significant S&P 500 result versus the insignificant S&P 100 result suggests that broader universes benefit more from the multi-agent structure, possibly because there is more cross-sectional dispersion for the agents to exploit.
What a Proper Independent Validation Would Look Like
If I were designing an independent validation of a multi-agent LLM scoring system, here is the checklist:
Run at least three simple baselines head-to-head: momentum, quality, and a combined factor composite. All baselines must use the same universe, same dates, same rebalancing frequency, same equal weighting. If the multi-agent system cannot beat the best simple baseline by a statistically significant margin after costs, it is not worth the complexity.
Split the evaluation into regime buckets. Test separately in risk-on months, risk-off months, high-volatility months, and low-volatility months. A system that works only in trending markets is a momentum proxy with extra steps.
Apply realistic transaction costs. For S&P 500 large caps at monthly rebalancing, 10-15 bps per side is reasonable. For smaller-cap or higher-frequency applications, costs matter more. Report net-of-cost Sharpe and Sortino ratios.
Freeze the system. No prompt changes, no retrieval updates, no parameter adjustments during the evaluation window. If the system requires ongoing tuning, report the tuning frequency and the performance decay between tuning events.
Run a Monte Carlo simulation on the equity curve itself, not just on monthly returns. A system with a 99.7th-percentile average monthly return that achieves it through two outlier months and seventeen flat ones has a very different risk profile than one with consistent monthly excess.
Test the individual agents in isolation. If the Fundamentals agent alone, used as a single-model scorer, captures 80% of the multi-agent system’s alpha, then the multi-agent architecture is adding complexity without proportional value.
What Traders Should Take from This Research
MarketSenseAI’s validation is above average for the LLM-finance literature. The live signal generation, Monte Carlo null, and NNLS attribution are methodologically sound. The S&P 500 result (p=0.003) is statistically strong for a 19-month sample. The agent rotation finding adds genuine insight into how multi-agent systems can adapt to changing markets.
But the validation has the gaps that matter most to practitioners: no factor-baseline comparison, no transaction-cost adjustment, thin regime coverage, and no walk-forward design. The strong-buy signal works as a universe filter. The sell-side signal does not work at all. The system has not been tested through a market environment where its architecture would be most stressed.
For traders evaluating whether to build or adopt a multi-agent LLM scoring system, the research supports one narrow claim: four specialised agents with a synthesis layer can produce stock selections that outperform random selection from the same universe. It does not yet support the stronger claim that this architecture outperforms simpler, cheaper alternatives. That is the test that still needs to be run.
Educational content only. Not investment advice. Trading involves risk. You are responsible for your decisions.