FinArena: Human-Agent Financial Forecasting

• FinArena is a human-agent collaboration framework that fuses multimodal data with investor risk profiles to generate personalized buy, sell, or hold recommendations.
• It employs a mixture-of-experts approach with specialized agents for analyzing stock prices, news, and financial statements to enhance prediction accuracy.
• The framework features adaptive retrieval-augmented generation and iterative human feedback to balance precise market analysis with tailored investment decisions.

Searching arXiv for the specified paper to ground the article in the published record. FinArena is a Human-Agent collaboration framework for financial market analysis and forecasting that combines multimodal financial data analysis with user interaction, with the stated aim of improving stock trend predictions and supporting personalized investment decisions. It is explicitly inspired by the mixture of experts (MoE) approach and is organized around a Human Module, a LLM-based (LLM-based) multi-agent Machine Module, and a Universal Expert Agent that fuses multimodal outputs with investor-specific risk preferences to produce buy, sell, or hold recommendations (Xu et al., 4 Mar 2025).

1. Conceptual position and overall structure

FinArena is defined by a division between human preference elicitation and machine-side multimodal inference. The Human Module captures individual risk preferences through an interactive interface. The Machine Module integrates three financial modalities: stock prices, news articles, and financial statements. A Universal Expert Agent then combines the extracted features and preliminary forecasts with the investor’s stated risk preference to generate the final action recommendation (Xu et al., 4 Mar 2025).

The framework is presented as a tightly coupled human-agent system rather than a purely autonomous trading model. In the textual block-diagram given in the source description, the investor provides a risk preference $R$ to the Human Module; the Machine Module processes time-series data, news via an adaptive RAG pipeline, and financial statements via iterative reasoning; and the Universal Expert or Report Agent outputs a final action $A = \text{buy/sell/hold}$. Figure 1, described as “The Framework of FinArena,” reportedly depicts arrows from each specialized agent into the final ReportAgent together with a loop back when the investor provides feedback.

This architecture suggests that FinArena is not only a forecasting system but also a personalization layer over multimodal financial reasoning. A plausible implication is that its distinctive contribution lies in treating risk preference as a first-class input to the aggregation process, rather than as a post hoc portfolio constraint.

2. Core computational formulation

FinArena is described using a set of mathematical formulations that organize its expert aggregation, retrieval logic, time-series prediction, statement reasoning, and final decision fusion (Xu et al., 4 Mar 2025).

The MoE-style gating mechanism is written as

$$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$

where $x$ is the multimodal input feature vector and $w_{g,i}, b_{g,i}$ are the gating parameters for expert $i$. The weights $\alpha_i$ blend the outputs of the specialized experts.

For adaptive Retrieval-Augmented Generation, the relevance function is expressed as

$$\text{score}(q,d) = \text{sim}(\text{encode}(q), \text{encode}(d)).$$

The description further states that, during the News Agent’s uncertainty-driven RAG process, a small LLM-based Judge Module $\mathcal{J}$ decides, via a martingale stopping rule, whether to fetch external passages $i_k$ from a search API; if $A = \text{buy/sell/hold}$0, the top-$A = \text{buy/sell/hold}$1 retrieved documents are appended to the prompt.

The Time Series Agent is formulated as

$A = \text{buy/sell/hold}$2

where

$A = \text{buy/sell/hold}$3

is the time-cumulative feature space of prices $A = \text{buy/sell/hold}$4 and covariates $A = \text{buy/sell/hold}$5, and $A = \text{buy/sell/hold}$6. An equivalent conceptualization is given by the ODE-style expression

$A = \text{buy/sell/hold}$7

A function $A = \text{buy/sell/hold}$8 then binarizes price changes to yield the trend forecast $A = \text{buy/sell/hold}$9.

For financial statements, the three-step iterative reasoning pipeline is specified as

$$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$0

$$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$1

$$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$2

The output $$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$3 is described as the season-by-season trend plus a tokenized confidence level.

Final action selection is written as

$$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$4

where $$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$5 and $$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$6 is a state-transition correction factor updated by human feedback:

$$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$7

The source also includes a canonical mean-variance objective, while explicitly noting that it is “not spelled out in the paper”:

$$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$8

subject to

$$\alpha_i = \frac{\exp(w_{g,i}^{\top} x + b_{g,i})}{\sum_j \exp(w_{g,j}^{\top} x + b_{g,j})},$$9

Because this formulation is identified as canonical rather than native to the paper, it is best treated as an interpretive lens for understanding how risk preference $x$0 may constrain portfolio selection.

3. Human module and personalization logic

Personalization in FinArena begins at launch, when the Human Module prompts the investor to provide a risk profile $x$1, either through simple labels such as “conservative” and “moderately aggressive” or in quantitative terms such as a target Sharpe ratio (Xu et al., 4 Mar 2025). ReportAgent₁ encodes $x$2 into a latent vector, and this representation influences three distinct components of the framework.

First, it affects the gating coefficients $x$3 in the MoE aggregator, shifting weight toward more risk-averse or more risk-seeking expert outputs. Second, it modifies the constraint set in the final portfolio-construction sub-module, for example through a variance cap. Third, it alters the entry and exit thresholds in the buy or sell decision rule.

The source gives an illustrative contrast: a “conservative” risk profile yields a lower maximum drawdown tolerance and a gating pattern that places more weight on the Statement Agent’s slower-moving fundamental signal, whereas an “aggressive” profile tilts toward the higher-volatility Time Series Agent. Throughout backtesting, the feedback loop $x$4 gradually refines the mappings from $x$5 to $x$6 and from $x$7 to strategy parameters.

This design distinguishes prediction from preference expression. The framework uses investor preference to condition action selection, but it separately reports that explicit investor market sentiment at prediction time reduced prediction accuracy. That separation suggests a deliberate attempt to prevent subjective bias from contaminating the forecasting stage while still allowing downstream personalization in execution and allocation.

4. Modal data processing and feature extraction

FinArena’s Machine Module is organized by modality, with each data source assigned to a dedicated agent and preprocessing path (Xu et al., 4 Mar 2025).

For stock price time series, the raw inputs are open, close, and volume for each trading day. The preprocessing pipeline includes normalization and feature-engineering, with moving averages and RSI given as examples, after which the data are fed directly into TimeGPT through a simple JSON input. The description states that the LLM’s time-series tokens capture the entire integral representation $x$8.

For financial news, the system scrapes headlines and full text from National Business Daily for A-share data or Business Today for U.S. data. Text cleaning includes regex-based removal of boilerplate followed by an LLM self-reflection pass that highlights residual “bias tokens” for excision, with Figure 2 cited for this stage. When multiple articles appear on the same day, k-means clustering is used to select the most representative article. Search augmentation is not unconditional: the RAG pipeline triggers retrieval only when the Judge $x$9 signals low confidence through the martingale test, thereby linking external search cost to model uncertainty.

For financial statements, the pipeline pulls key ratios such as revenue, net income, total assets, and free cash flow via an API, packages Q1–FY tables into a single prompt, and applies the three-step LLM reasoning process described earlier: seasonality extraction, holistic analysis, and final stock-trend prediction with confidence.

Feature fusion is delegated to the Universal Expert Agent, which combines the binary trend forecasts $w_{g,i}, b_{g,i}$0, $w_{g,i}, b_{g,i}$1, and $w_{g,i}, b_{g,i}$2 together with the corresponding LLM-reported confidence values. The final weighted score is written as

$w_{g,i}, b_{g,i}$3

with $w_{g,i}, b_{g,i}$4 dependent on $w_{g,i}, b_{g,i}$5.

A plausible implication is that FinArena treats multimodal fusion as a confidence-weighted decision synthesis problem rather than a single end-to-end joint encoder. That makes the interpretive role of each modality more explicit, especially under heterogeneous risk profiles.

5. Experimental protocol and empirical findings

The empirical evaluation uses ten large-cap stocks, comprising five U.S. and five A-share names, each with approximately 300 trading days, approximately 150–300 news articles after k-means reduction, and one year of quarterly statement data. The test period is Jan 1, 2024 – Mar 27, 2024 (Xu et al., 4 Mar 2025).

The benchmark suite is divided into prediction and trading baselines. For stock movement prediction, the baselines are ARIMA(1,2,1), a 4-layer LSTM, and TimeGPT as a single-modal LLM. For trading simulation, the baselines are Random, BRSF (Buy-on-Rising-Streak / Sell-on-Falling-Streak), ARIMA-based and LSTM-based strategies, plus a pure news-only LLM ablation. Evaluation metrics are Accuracy and F1-Score for prediction, and Annualized Return (AR), Sharpe Ratio (SR), and Maximum Drawdown (MD) for trading.

Quantitative summary

Evaluation setting	Reported result	Context
U.S. stock prediction	FinArena mean Accuracy = 0.5833 vs. TimeGPT = 0.5745	FinArena higher on Accuracy
U.S. stock prediction	FinArena F1 = 0.5540 vs. TimeGPT = 0.5723	Slightly below TimeGPT on F1, with lower variance
A-share prediction	FinArena Accuracy = 0.5072 vs. TimeGPT = 0.4897	FinArena higher on Accuracy
A-share prediction	FinArena F1 = 0.5000 vs. TimeGPT = 0.4897	FinArena higher on F1
Adaptive RAG ablation	Consistent $w_{g,i}, b_{g,i}$6–$w_{g,i}, b_{g,i}$7 lift in both Acc and F1 over no-RAG	Reported in Fig. 4
Trading simulation	Moderately Aggressive FinArena: AR = 0.6271, SR = 2.4817	Compared with ARIMA-based AR = 0.2313, SR = 0.9954
Conservative profile	AR = 0.3153, MD = 0.0653	Lower return, lowest drawdown

The reported findings separate prediction quality from investment outcomes. In U.S. stocks, FinArena exceeds TimeGPT in mean Accuracy but is slightly lower in F1, though with lower variance. In A-share stocks, it exceeds TimeGPT on both Accuracy and F1. The adaptive RAG mechanism yields a consistent approximately $w_{g,i}, b_{g,i}$8–$w_{g,i}, b_{g,i}$9 improvement in both Accuracy and F1 relative to the no-RAG setting. Replacing the DeepSeek-v2 expert with language-specialized models such as LLaMA-3 or Kimi degrades performance, which the source interprets as evidence for a “mixture-of-economical-experts.”

The framework also addresses the effect of sentiment injection: adding explicit investor market sentiment labels such as sensitive, insensitive, or optimistic at prediction time reduced prediction accuracy. The study presents this as an answer to a research question, namely that the forecasting stage should not be biased by investor sentiment.

In trading simulation, FinArena-driven strategies reportedly outperform all baselines in average AR and SR across both markets. The Moderately Aggressive profile achieves AR = 0.6271 and SR = 2.4817, compared with the ARIMA-based strategy’s AR = 0.2313 and SR = 0.9954. Conservative investors obtain lower AR at 0.3153 but also the lowest MD at 0.0653, illustrating the stated risk-return trade-off. The paper also notes that some extreme cases, such as Nvidia’s massive up-trend, were better served by simple BRSF, even though multimodal fusion was superior on average.

6. Interpretation, constraints, and projected extensions

The main interpretation advanced for FinArena is that an MoE-inspired, human-in-the-loop, multi-agent LLM framework can exceed both classical baselines such as ARIMA and LSTM and a single-modal LLM baseline such as TimeGPT in the U.S. equity market (Xu et al., 4 Mar 2025). The adaptive RAG mechanism is described as reducing hallucinations on fresh news while economizing API calls for low-cost investors, and the human-agent interface is presented as crucial because it avoids “one size fits all” advice.

The reported limitations are equally central to understanding the framework’s scope. A-share trading results are negative across all FinArena profiles, and this is explicitly used to highlight the caveat that data quality and information asymmetry can overwhelm LLM analytics in less efficient markets. The source further notes that news clustering or RAG may still overfetch or underfetch Chinese-language context. Trading performance remains contingent on the quality and recency of the underlying data corpus, and the dataset is described as small-scale by design, insufficient to fully capture high-frequency or alternative data sources such as social media and order-book feeds.

The ODE-style time-series model is also characterized as a conceptual abstraction rather than a definitive mechanistic model. Future work is said to include possible fine-tuning of foundation time-series LLMs such as Time-LLM, as well as hybridization with domain-specific volatility models. Additional extensions proposed in the source are incorporation of social media sentiment or SEC filings such as 8-K and 10-K as new agents, more sophisticated portfolio optimizers that solve mean-variance or CVaR programs explicitly within the LLM chain-of-thought, and a live A/B test with real-money paper trading under transaction costs and market impact.

A plausible implication of these limitations is that FinArena should be read less as a universal claim about LLM superiority in finance than as a blueprint for structured human-agent collaboration under multimodal data heterogeneity. Its empirical profile is strongest where data quality, market transparency, and retrieval fidelity jointly support the multi-agent decomposition.