F IN C ON: A Synthesized LLM Multi-Agent System

with Conceptual Verbal Reinforcement for Enhanced

Financial Decision Making

Yangyang Yu1,⋆ , Zhiyuan Yao1,⋆ , Haohang Li1,⋆ , Zhiyang Deng1,⋆ , Yuechen Jiang1,⋆ , Yupeng Cao1,⋆
Zhi Chen1,⋆ , Jordan W. Suchow1 , Zhenyu Cui1 , Rong Liu1 , Zhaozhuo Xu1 , Denghui Zhang1
arXiv:2407.06567v3 [[Link]] 7 Nov 2024

Koduvayur Subbalakshmi1 , Guojun Xiong2 , Yueru He3 , Jimin Huang 3 , Dong Li3 , Qianqian Xie3,†
1
Stevens Institute of Technology 2 Harvard University 3 The Fin AI
⋆
These authors contributed equally † Corresponding author: [Link]@[Link]

Abstract
Large language models (LLMs) have shown potential in complex financial tasks,
but sequential financial decision-making remains challenging due to the volatile
environment and the need for intelligent risk management. While LLM-based
agent systems have achieved impressive returns, optimizing multi-source infor-
mation synthesis and decision-making through timely experience refinement is
underexplored. We introduce F IN C ON, an LLM-based multi-agent framework
with CONceptual verbal reinforcement for diverse FINancial tasks. Inspired by
real-world investment firm structures, F IN C ON employs a manager-analyst hier-
archy, enabling synchronized cross-functional agent collaboration towards unified
goals via natural language interactions. Its dual-level risk-control component
enhances decision-making by monitoring daily market risk and updating systematic
investment beliefs through self-critique. These conceptualized beliefs provide
verbal reinforcement for future decisions, selectively propagated to relevant agents,
improving performance while reducing unnecessary peer-to-peer communication
costs. F IN C ON generalizes well across tasks, including single stock trading and
portfolio management. 1

1 Introduction
The intricacies and fluctuations inherent in financial markets pose significant challenges for making
high-quality, sequential investment decisions. In tasks such as single stock trading and portfolio
management, each intelligent decision is driven by multiple market interactions and the integration
of diverse information streams, characterized by varying levels of timeliness and modalities [1, 2].
The primary objective of these tasks is to maximize profit while managing present market risks in an
open-ended environment.
In practice, trading firms often depend on synthesized teamwork, structured hierarchically with
functional roles such as data analysts, risk analysts, and portfolio managers communicating across
levels [3, 4]. These roles are responsible for the careful integration of diverse resources. However,
the cognitive limitations of human team members can hinder their capacity to rapidly process market
signals and achieve optimal investment outcomes [5].
To enhance investment returns and address the limitations of human decision-making, various
studies have explored methods such as deep reinforcement learning (DRL) to develop agent systems
that simulate market environments and automate investment strategies [6, 7, 8]. Concurrently,
1
We will release the code and demo in the following repo [Link]

38th Conference on Neural Information Processing Systems (NeurIPS 2024).

advancements in large language models (LLMs) have shown great potential in performing complex
tasks, including reasoning [9, 10], tool-using [11], planning [12], decision-making [13, 14], and even
in various financial applications [15, 16, 17, 18, 19], suggesting they may surpass existing agent
architectures. Language agents, in particular, are distinguished by their human-like communication
and flexible, prompt-based structures, making them well-suited to diverse decision-making settings
[20, 21, 22, 23].
To achieve optimal decision-making performance, two critical factors must be considered: (1)
Organizing agents to facilitate effective teamwork and efficient communication, and (2) Enabling
agents to continuously learn and refine their actions. Studies have shown that mimicking human
organizational structures can successfully coordinate language agents for specific tasks [24, 25, 26].
Additionally, recent advances in textual gradient-based prompt optimization [27, 28] and verbal
reinforcement [29, 30] have proven effective in iteratively improving the reasoning and decision-
making capabilities of language agents.
Language agent systems designed for financial decision-making, such as F IN GPT [31], F IN M EM [32],
and F INAGENT [33], have shown strong performance. However, they face several limitations. First,
their reliance on agents’ risk preferences based on short-term market fluctuations fails to control long-
term risk exposure, potentially overlooking fundamental factors driving investment returns. A more
effective approach is to quantify investment risks using established measures of risk from quantitative
finance [34, 35]. Second, these systems are often limited to single-asset trading tasks, making them
less adaptable to multi-asset financial applications like portfolio management. Third, they place
significant pressure on a single agent to understand and process information within a constrained
context window, which can degrade decision quality. Although approaches like S TOCK AGENT [36]
use multi-agent systems for stock trading, their reliance on extensive discussions between numerous
LLM agents leads to high communication costs and slow decision-making. Moreover, the absence of
a clear optimization objective can compromise outcome effectiveness. Additional related work in the
literature is discussed in the Appendix A.1.
To address these issues, we propose F IN C ON, an LLM-based multi-agent framework for critical
financial tasks, such as single-stock trading and portfolio management, as shown in Figure 1. Our
main contributions are: 1) Inspired by real-world investment roles, we introduce a novel Synthesized
Manager-Analyst hierarchical communication structure with a risk-control component. This
structure allocates financial data from different sources to corresponding functional analyst agents,
allowing them to focus on specific insights, while the manager consolidates these inputs to make
informed trading decisions. The streamlined communication reduces redundant peer-to-peer interac-
tion, lowering costs and improving efficiency. 2) Our framework generalizes beyond stock trading to
handle portfolio management, an area not previously addressed by other financial language agent
systems. 3) We developed a dual-level risk control component to update risk assessments both
within and across episodes. Within episodes, risk is supervised using the Conditional Value at Risk
(CVaR), a quantile-based risk measure [37]. Across episodes, we introduced a verbal reinforcement
mechanism, where investment beliefs are updated based on reasoning trajectories and profit-and-loss
(PnL) trends, distilled into conceptual perspectives. These insights are selectively back-propagated
from the manager to relevant analyst agents. Our ablation studies demonstrate the effectiveness of
this risk control design in managing market risk and enhancing trading performance.

Figure 1: The general framework of F IN C ON.

2
2 Preliminaries

Here, we outline the mathematical notations for the two major financial decision-making tasks
that will be explicitly discussed in our work. We also formally present the generalized modeling
formulation using a Partially Observable Markov Decision Process (POMDP) [38] for financial
decision-making tasks.

2.1 Financial Decision-making Tasks Formulation

i I
Single Stock Trading Tasks. F IN C ON uses analyst agents group {Mpr }i=1 to process multi-modal
market information sources. The processed information is then used by a manager agent Ma to make
trading decisions (buy, sell, hold), and to provide relevant reasoning texts. Note that the “sell” signal
means the system makes a “short-selling” decision, that is, a negative trading position is allowed.
Additionally, F IN C ON evaluates the daily investment risk, followed by prompt-optimization for the
manager agent from risk-control component Mr .
Portfolio Trading Tasks. In addition to processing multi-modal market information, the analyst
agents also construct a stock pool for portfolio management by considering the statistical correlations
between stock returns. The manager agent then makes trading decisions for each stock in the
pool. Finally, the manager agent determines the portfolio weights for all stocks using an external
optimization solver that applies the mean-variance optimization described below [39]:


∈ [0, 1], “buy”
max⟨w, µ⟩ − ⟨w, Σw⟩ s.t. wn = ∈ [−1, 0], “sell” , ∀n ∈ {1, · · · , N } (1)
w 
= 0, “hold”

where w = (w1 · · · , wN ) ∈ RN is portfolio weights vector, µ and Σ are the shrinkage estimators
of N -dimensional sample expected return and N × N sample covariance matrix of chosen stocks’
daily return sequences respectively [35]. We note that portfolio weights are rebalanced on daily
basis. In our implementation, we begin by calculating the portfolio weights through solving the
aforementioned optimization problem. Next, the target positions are determined by linearly scaling
these portfolio weights from the previous step.

2.2 Modeling Quantitative Trading as POMDP

Formally, we model quantitative trading task as an infinite horizon POMDP [40, 41] with time index
T = {0, 1, 2, · · · } and discount factor α ∈ (0, 1]. The components of this model are as follows: (1) a
state space X × Y where X is the observable component and Y is unobservable component of the
QI
financial market; (2) the action space of analyst agents group is A = i=1 Ai , where Ai represents
the collection of processed market information in textual format done by agent i (total I analyst
agents), and for manager agent, its action space is A, which is modeled as {“buy", “sell", “hold"}
for single stock trading task and as ({“buy", “sell", “hold"} × [−1, 1])⊗N for portfolio management
task among N -stocks; (3) the reward function R(o, b, a) : X × Y × A → R uses daily profit &
loss (PnL) as the output; (4) the observation process {Ot }t∈T ⊆ X is an I-dimensional process,
with the ith entry {Oti }t∈T representing one type of uni-modal information flow solely processed
by the analyst agent i; (5) the reflection process {Bt }t∈T ⊆ Y represents the manager agent’s
self-reflection, which is updated from Bt to Bt+1 on daily basis [42]); (7) the processed information
flow Ôt = (Ôt1 , · · · , ÔtI ) ∈ A, ∀ t ∈ T, which represents the information processing outputs from
analyst agents group.
Then, our multi-agent system is supposed to learn the policies of all agents: the policies of analyst
agents πθi i : X → Ai , i ∈ {1, · · · , I} (the ways to process information, i.e. Ôti ∼ πθi i (·|Oti )),
and the policy of manager agent πθa : A × Y → A (the ways to make trading decisions, i.e.
At ∼ πθa (·|Ôt , Bt )) such that the system maximizes cumulative trading reward while controlling
risk [43]. All policies Πθ = ({πθi i }Ii=1 , πθa ) are parameterized by textual prompts θ = ({θi }Ii=1 , θa ).
By updating prompts via the risk-control component Mr , the whole system optimizes policies Πθ in
a verbal reinforcement manner. By denoting daily profit & loss (PnL) by RtΠθ = R(Ot , Bt , At ), the

3
optimization objective for the whole system can be written as:
hX θ
i
max E αt RtΠ (2)
θ
t∈T

is a risk-sensitive optimization problem that leverages textual gradient descent, fundamentally dif-
fering from DRL algorithms designed for POMDPs. Further details on the textual gradient descent
approach are provided in the Appendix A.2.

3 Architecture of F IN C ON
In this section, we present the architecture of F IN C ON using a two-level hierarchy. First, we describe
the hierarchical framework for coordinating the agents’ synchronous work and communication. Then,
we elaborate on the functionalities of each module that constitutes each agent in F IN C ON. Finally,
we aim to elaborate on how F IN C ON solves the objective function expressed as Equation (2) through
a verbal reinforcement approach.

3.1 Synthesized Multi-agent Hierarchical Structure Design

The agent system of F IN C ON consists of two main components: the Manager-Analyst Agent Group
component and the Risk-Control component.

Figure 2: The detailed architecture of F IN C ON contains two key components: Manager-Analyst

agent group and Risk Control. It also presents the between-component interaction of F IN C ON and
decision-making flow.

3.1.1 Manager-Analyst Agent Group

Analogous to human investment firm, F IN C ON establishes a unique hierarchical structure to organize
its multi-agent system, synthesizing their efforts to achieve superior decision-making outcomes.

4
3.1 Synthesized Multi-agent Hierarchical Structure Design

The primary goal is to enhance information presentation and comprehension while minimizing
unnecessary communication costs. The working mechanism of each agent is illustrated in Figure 2.
Analyst Agents. In F IN C ON, analyst agents distill concise investment insights from large volumes of
multi-source market data, each focused on a specific trading target. To ensure high-quality reasoning
by reducing task load and sharpening focus, each agent processes information from a single source in
a uni-modal fashion, providing pre-specified outputs based on prompts. This setup mimics an efficient
human team, where each analyst specializes in a specific function, filtering out market noise and
extracting key insights. These agents assist the manager agent by consolidating denoised investment
information from multiple perspectives. We implement seven distinct types of analyst agents using
LLMs, each producing unique investment insights, as shown in the upper section of Figure 2. Based
on input modalities, three textual data processing agents extract insights and sentiments from daily
news and financial reports. An audio agent uses the Whisper API to interpret investment signals from
earnings call recordings. Additionally, a data analysis agent and a stock selection agent compute
critical financial metrics, such as momentum and CVaR, using tabular time series data. The stock
selection agent also oversees portfolio selection by applying the classic risk diversification method in
quantitative finance [1].
Manager Agent. In F IN C ON, the manager agent acts as the sole decision-maker, responsible for
generating trading actions for sequential financial tasks. For portfolio management, it calculates
portfolio weights using convex optimization techniques constrained by directional trading decisions
(see optimization problem as presented in Formula (1)). Four key mechanisms support each decision:
1) Consolidating distilled insights from multiple analyst agents. 2) Receiving timely risk alerts and
conceptual investment updates from the risk control component. 3) Refining its investment beliefs
about the influence of different information sources on trading decisions for specific targets. 4)
Conducting self-reflection by reviewing reasoning outcomes from previous trading actions.

3.1.2 Risk-Control Component

We have innovatively designed a dual-level risk-control mechanism consisting of within-episode

and over-episode risk management. The within-episode mechanism detects market risk within a
single training episode, allowing the manager agent to promptly adjust trading actions to mitigate
potential losses by accounting for short-term trading performance and market fluctuations. This
mechanism also operates during the testing phase. In contrast, the over-episode mechanism functions
exclusively during the training stage, providing prompt optimization guidance by comparing the
trading performance of the current episode with the previous one. This reflection enables the manager
agent to update its investment beliefs based on performance differences. By drawing on prior
observations of market risk and profitability patterns, these two mechanisms help avoid repeated
investment errors, thereby enhancing future returns.
Within-Episode Risk Control: The within-episode risk alert is triggered by a sudden drop in the
CVaR value. Conditional Value at Risk (CVaR) represents the average of the worst-performing 1%
of daily trading Profits and Losses (PnLs). A decrease in CVaR typically indicates that recent trading
decisions have led to PnLs within this bottom percentile, signaling a potentially high-risk market
condition. When this occurs, the manager agent adopts a risk-averse stance for that day’s trading
actions, regardless of the prior risk status.
Over-Episode Risk Control: The over-episode investment belief updates facilitate adjustments in the
emphasis placed on analysts’ information distillation and the manager’s action generation. Through
the Actor-Critic mechanism, F IN C ON episodically optimizes its investment strategy for a given
trading target, as defined by objective (Equation (2)), by reflecting on a series of winning and losing
actions. This episodic reflection is powered by a unique Conceptual Verbal Reinforcement (CVRF).
CVRF assesses the performance of consecutive training episodes by analyzing the information per-
spectives provided by analysts and reflected in the manager’s decision-making. It then conceptualizes
and attributes the evaluation outcomes to these specific aspects. By comparing the conceptualized
insights from more profitable versus less profitable episodes, the system informs both the manager
and analyst agents about necessary belief adjustments, helping prioritize the most relevant market
information for increased profitability, as detailed in Algorithm 1. CVRF leverages text-based gradi-
ent descent to offer optimal conceptual investment guidance for the manager agent, refining prompts
with the latest investment beliefs. The guidance is organized according to perspectives provided by

5
3.2 Modular Design of F IN C ON Agents

the respective analyst agents, key financial indicators (such as historical momentum), or other crucial
viewpoints.

Factor Gradient-based model optimizer LLM-based prompt optimizer

Upgrade direction Model value gradient momentum Prompt reflection trajectory
Update method Learning rate descent Overlapping percentage of trading decisions

Table 1: Analogy between glossaries in model optimizer and prompt optimizer.

These belief updates are first received by the manager agent and then selectively propagated to
relevant agents, minimizing over-communication. Unlike the text-based gradient descent proposed
by Tang et al.[28], which uses prompt editing distance as a learning rate, we derive investment belief
updates by measuring the overlapping percentage of trading actions between two consecutive training
trajectories at each belief update, as presented in Table 1. This approach has proven effective in
improving the performance of a synthesized agent system, where each worker has a clearly defined
and specialized role. The above describes the workflow of F IN C ON during the training stage, while
the workflow during the testing stage is detailed in the Appendix A.3.

Algorithm 1 Training Stage Algorithm of F IN C ON: Conceptual Verbal Reinforcement using Textual-based
Gradient Descent
i
Initialize manager-analysts component {Mpr }Ii=1 &Ma , and risk-control component Mr .
Initialize trading start date s, stock pool of portfolio and portfolio weights w0 = 0.
Initialize Prompts θ, policy Πθ .
while episode k < M ax do
for 0 ≤ t ≤ T do
Run policy Πθ (collecting daily PnL rt , portfolio weights wt and daily CVaR value ρt ).
if ρt < ρt−1 or rt < 0 then
Trigger Ma self-reflection and generate self-reflection text Bt .
end if
Get the investment trajectory Hk and calculate the objective function value (Function (2)).
end for
Compare the objective function values of episodes k − 1 & k, and decide which episode has higher
performance;
Pass sustained profitable and losing trades from two episodes Hk−1 & Hk into risk-control component
Mr ;
Guide Mr to summarize conceptualized investment insights {c1k−1 , · · · , cn 1 m
k−1 } & {ck , · · · , ck };
Compare two sets of conceptualized insights and give the reasoning for higher performance (providing
textual optimization direction, i.e. meta prompt);
Calculate the overlapping percentage between trading decision sequences from two episodes (providing
the learning rate τ );
Update the prompts by textual gradient-descent:
θ ←− Mr (θ, τ, meta prompt).
end while

3.2 Modular Design of F IN C ON Agents

Here, we explain the modular design of F IN C ON agents. Inspired by the recent works of Park et al.
[44] and Sumers et al. [45] on developing the cognitive structure of language agents for human-like
behavior, agents in F IN C ON integrate four modules to support their necessary functionalities, along
with a shared general configuration, as detailed in Appendix A.4:
General Configuration and Profiling Module. This module defines task types (e.g., stock trading,
portfolio management) and specifies trading targets, including sector and performance details. The
profiling module outlines each agent’s roles and responsibilities. The concatenated textual content
from these parts is used to query investment-related events from the agents’ memory databases.
Perception Module. This module defines how each agent interacts with the market, specifying the
information they perceive, receive, and communicate, with interactions tailored to each agent’s role.
In detail, it converts raw market data, feedback from other agents, and information retrieved from
the memory module into formats compatible with large language models, enabling them to process
these inputs effectively. Memory Module. The memory module comprises three key components:

6
working memory, procedural memory, and episodic memory. Much like how humans process events
in their working memory [46], F IN C ON agents leverage their working memory to perform a range of
tasks, including observation, distillation, and refinement of available memory events, all tailored to
the specific roles of the agents. Procedural memory and episodic memory are critical for recording
historical actions, outcomes, and reflections during sequential decision-making. Procedural memory
is generated after each decision step within an episode, storing data as memory events. For trading
inquiries, top events are retrieved from procedural memory and ranked based on recency, relevance,
and importance, following a simplified version of the method proposed by Yu et al. [32], with further
details provided in Appendix A.13. Each functional analyst agent has distinct procedural memory
decay rates, reflecting the timeliness of various financial data sources, which is crucial for aligning
multi-type data influencing specific time points and supporting informed decision-making. The
manager agent enhances the procedural memory of analyst agents by providing feedback through
an access counter. Both analyst and manager agents maintain procedural memory, but they keep
different records, as illustrated in Appendix A.4. Episodic memory, exclusive to the manager agent,
stores actions, PnL series from previous episodes, and updated conceptual investment beliefs from
the risk control component.

4 Experiments
Our experiment answers the key research questions (RQs): RQ1: Does F IN C ON demonstrate robust-
ness across multiple financial decision-making tasks, especially single-asset trading and portfolio
management? RQ2: Is the within-episode risk control mechanism in F IN C ON effective in maintain-
ing superior decision-making performance? RQ3: Is the over-episode risk control mechanism in
F IN C ON effective in timely updating the manager agent’s beliefs to enhance trading performance?

4.1 Experimental Setup

(i) Multi-Modal Datasets. We construct a market environment representation using real-world

financial data, including stock prices, daily news, company filings (Form 10-Q, Form 10-K, etc.), and
ECC audio from January 3, 2022, to June 10, 2023, as detailed in Appendix. A.8. Each data source
is assigned to specific analyst agents based on its timeliness. (ii) Evaluation Metrics. We evaluate
F IN C ON and other state-of-the-art (SOTA) agents using metrics such as Cumulative Return (CR%),
Sharpe Ratio (SR), and Max Drawdown (MDD%). CR and SR are prioritized because they provide
comprehensive insights into overall performance and risk-adjusted returns, essential for informed
investment decisions. In contrast, MDD focuses on evaluating the potential for significant losses,
making it a secondary consideration in this context. Details are provided in Appendix A.10. (iii)
Comparative Methods. For single-stock trading, we compare F IN C ON with DRL agents (A2C, PPO,
DQN) and LLM-based agents (G ENERATIVE AGENT (GA), F IN GPT, F IN M EM, F INAGENT) as
well as the Buy-and-Hold (B & H) strategy. For portfolio management, we compare F IN C ON with
Markowitz MV, FinRL-A2C, and Equal-Weighted ETF strategy, with further details provided in
Appendix A.12. The detailed experiment parameter configurations of the above agent systems are
articulated in Appendix. A.14. (iv) Implementation Details. All LLM-based agents use GPT-4-Turbo,
with temperature set at 0.3. F IN C ON is trained from January 3, 2022, to October 4, 2022, and tested
from October 5, 2022, to June 10, 2023. DRL agents are trained over the period from January 1,
2018, to October 4, 2022, to ensure that there is sufficient data available for model convergence.
Performance is based on the median CR and SR from five repeated epochs. For a more detailed
explanation of the experimental setup, please refer to the Appendix A.5.

4.2 Main Results

In response to RQ1, we analyze F IN C ON’s performance on two types of financial decision-making

tasks: single-asset trading and portfolio management. The system’s ability to manage these sequen-
tially complex decisions is thoroughly evaluated in the following sections.

4.2.1 Single Asset Trading Task

In this task, we evaluate F IN C ON’s performance against other leading algorithmic trading models by
trading eight different stocks. As presented in the tables above, F IN C ON significantly outperforms
both LLM-based and DRL-based approaches in terms of CRs and SRs. Additionally, F IN C ON

7
4.2 Main Results

achieves one of the lowest MDD values across most trading assets, demonstrating effective risk man-
agement while still delivering the highest investment returns. For detailed performance comparisons
across all models and metrics, refer to Table 1.
Overall, even with extended training periods, DRL-based models tend to underperform, with the A2C
algorithm lagging significantly behind other agents in general. Notably, the training periods for Nio
Inc. (NIO) and Coinbase Global Inc. (COIN) require clarification. NIO, which completed its IPO
in September 2018, has a slightly shorter training period than other tickers, yet the DRL algorithms
for NIO still achieved convergence. In contrast, Coinbase Global Inc. (COIN), which completed its
IPO in April 2021, presented a more significant challenge due to the limited available trading data,
causing DRL algorithms to struggle with convergence. This limitation underscores a major drawback
for DRL agents when trading recently listed IPOs. Consequently, our analysis of COIN focuses on
comparisons between F IN C ON, LLM-based agents, and the buy-and-hold (B & H) strategy. In this
context, F IN C ON demonstrates a clear advantage, achieving a cumulative return of over 57% and a
Sharpe ratio of 0.825. Furthermore, LLM-based agents, which can leverage diverse data types and
require minimal training, effectively mitigate the challenges faced by DRL algorithms.
TSLA AMZN NIO MSFT
Categories Models
CR% ↑ SR↑ MDD% ↓ CR % ↑ SR↑ MDD% ↓ CR% ↑ SR↑ MDD% ↓ CR% ↑ SR↑ MDD% ↓
Market B&H 6.425 0.145 58.150 2.030 0.072 34.241 -77.210 -1.449 63.975 27.856 1.230 15.010
Our Model F IN C ON 82.871 1.972 29.727 24.848 0.904 25.889 17.461 0.335 40.647 31.625 1.538 15.010
GA 16.535 0.391 54.131 -5.631 -0.199 37.213 -3.176 -1.574 3.155 -31.821 -1.414 39.808
F IN GPT 1.549 0.044 42.400 -29.811 -1.810 29.671 -4.959 -0.121 37.344 21.535 1.315 16.503
LLM-based
F IN M EM 34.624 1.552 15.674 -18.011 -0.773 36.825 -48.437 -1.180 64.144 -22.036 -1.247 29.435
F INAGENT 11.960 0.271 55.734 -24.588 -1.493 33.074 0.933 0.051 19.181 -27.534 -1.247 39.544
A2C -35.644 -0.805 61.502 -12.560 -0.444 37.106 -91.910 -1.728 68.911 21.397 0.962 21.458
DRL-based PPO 1.409 0.032 49.740 3.863 0.138 28.085 -72.119 -1.352 62.093 -4.761 -0.214 30.950
DQN -1.296 -0.029 58.150 11.171 0.398 31.174 -35.419 -0.662 56.905 27.021 1.216 21.458

AAPL GOOG NFLX COIN

Categories Models
CR% ↑ SR↑ MDD% ↓ CR % ↑ SR↑ MDD% ↓ CR% ↑ SR↑ MDD% ↓ CR% ↑ SR↑ MDD% ↓
Market B&H 22.315 1.107 20.659 22.420 0.891 21.191 57.338 1.794 20.926 -21.756 -0.311 60.187
Our Model F IN C ON 27.352 1.597 15.266 25.077 1.052 17.530 69.239 2.370 20.792 57.045 0.825 42.679
GA 5.694 0.372 14.161 -1.515 -0.192 8.210 41.770 1.485 20.926 19.271 0.277 67.532
F IN GPT 20.321 1.161 16.759 0.242 0.011 26.984 11.925 0.472 20.201 -99.553 -1.807 74.967
LLM-based
F IN M EM 12.397 0.994 11.268 0.311 0.018 21.503 -10.306 -0.478 27.692 0.811 0.017 50.390
F INAGENT 20.757 1.041 19.896 -7.440 -1.024 10.360 61.303 1.960 20.926 -5.971 -0.106 56.882
A2C 13.781 0.683 14.226 8.562 0.340 21.191 -8.176 -0.258 49.579 - - -
DRL-based PPO 14.041 0.704 22.785 2.434 0.097 25.202 -33.144 -1.049 33.377 - - -
DQN 21.125 1.048 16.131 20.690 0.822 21.191 21.753 0.687 39.733 - - -

Table 2: Comparison of key performance metrics during the testing period for the single-asset trading tasks
involving eight stocks, between F IN C ON and other algorithmic agents. Note that the highest and second highest
CRs and SRs have been tested and found statistically significant using the Wilcoxon signed-rank test. The highest
CRs and SRs are highlighted in red, while the second highest are marked in blue.
In alignment with market trends, F IN C ON consistently exhibits superior decision-making quality
compared to other LLM-based agents, regardless of market conditions—whether bullish (e.g., GOOG,
MSFT), bearish (e.g., NIO), or mixed (e.g., TSLA). We attribute this performance to its high-quality
distillation of information through a synthesized multi-agent collaboration mechanism, combined with
its dual-level risk control design, positioning F IN C ON as a leader in the space. By contrast, F IN GPT
primarily relies on sentiment analysis of financial information, failing to fully exploit the potential of
LLMs to integrate nuanced textual insights with numerical financial indicators. Similarly, GA and
F IN M EM use single-agent frameworks without sophisticated information distillation processes or a
diverse toolset, placing heavy cognitive demand on the agent to process multi-source information,
especially when dealing with large and varied data modalities. Moreover, their static or minimal
investment belief systems result in weak filtering of market noise. As illustrated in Figure 7 (a) &
(b) of Appendix A.7.2, this limitation leads these models to consistently hold lower positions and
hesitate between ‘buy’ or ‘sell’ decisions, ultimately resulting in suboptimal performance.
F IN C ON overcomes these challenges through its innovative multi-agent synthesis, enabling it to
deliver superior outcomes. Although F INAGENT performs well when integrating images and tabular
data, it struggles to remain competitive when incorporating audio data, such as ECC recordings,
which are critical in real-world trading. Additionally, F INAGENT relies on similarity-based memory
retrieval, which can lead to decisions based on outdated information, often resulting in errors. In

8
4.3 Ablation Studies

contrast, F IN C ON’s memory structure accounts for the varying timeliness of multi-source financial
data, significantly enhancing decision quality and overall performance.

4.2.2 Portfolio Management Task

In this task, we compare F IN C ON’s performance with the Markowitz Mean-Variance (MV) portfolio
[47] and F IN RL [48] in managing two small portfolios: Portfolio 1 (TSLA, MSFT, and PFE) and
Portfolio 2 (AMZN, GM, and LLY). These assets were selected by the stock selection agent from a
pool of 42 stocks, each with sufficient news data (over 800 news articles during the combined training
and testing periods), as illustrated in Figure 9 in Appendix A.9. The training and testing periods, the
backbone model and the parameter settings are consistent with those used in the single-asset trading
task. For the Markowitz MV portfolio, we estimate the covariance matrix and expected returns using
the same training data. In the case of F IN RL, we use five years of training data prior to the test
period. As detailed in Table 3 and Figure 3, our results show that F IN C ON outperforms both the
Markowitz MV portfolio and F IN RL as well as the market baseline – Equal-Weighted ETF, achieving
significantly higher CRs and SRs, as well as MDDs.
However, managing multi-asset portfolios introduces more complexity, leading to a higher likeli-
hood of hallucination compared to single-asset trading. This is due to the increased input length
and complexity involved in multi-asset decision-making. While F IN C ON mitigates this issue by
distributing tasks across specialized agents that focus on critical investment insights, it occasionally
generates incorrect information, such as non-existent indices of memory events. Handling multi-asset
decision-making requires sophisticated logic and substantial market information, which poses a
significant challenge for LLMs when processing extended contexts. This complexity has left port-
folio management relatively unexplored in previous language agent studies. Nonetheless, F IN C ON
demonstrates considerable potential by constructing agent systems that can tackle complex financial
tasks through effective resource optimization, even when managing relatively compact portfolios.

Models CR % ↑ SR↑ MDD %↓ Models CR % ↑ SR↑ MDD %↓

F IN C ON 113.836 3.269 16.163 F IN C ON 32.922 1.371 21.502
Markowitz MV 12.636 0.614 17.842 Markowitz MV 10.289 0.540 25.099
F IN RL-A2C 19.461 0.831 26.917 F IN RL-A2C 11.589 0.649 15.787
Equal-Weighted ETF 9.344 0.492 21.223 Equal-Weighted ETF 15.061 0.867 14.662

Table 3: Key performance metrics comparison among all portfolio management strategies of Portfolio 1 & 2.
F IN C ON leads all performance metrics.

Figure 3: Portfolio values of Portfolio 1 & 2 changes over time for all the strategies. The computation of
portfolio value refers to Equation 7 in Appendix A.10.

4.3 Ablation Studies

In response to RQ2 and RQ3, we conduct a comprehensive evaluation of our unique risk control
component through two ablation studies. Both studies maintain consistency with the training and
testing periods used in the main experiments. The first study examines the effectiveness of the within-
episode risk control mechanism, which leverages Conditional Value at Risk (CVaR) to manage risk in
real-time, as detailed in Table 4. Comparisons on primary metrics illustrate that the success of utilizing
CVaR for within-episode risk control is evident in both bullish and bearish market environments in the

9
single asset trading case. Moreover, in portfolio trading with mixed price trends, our within-episode
risk control mechanism performs robustly by monitoring the entire portfolio’s value fluctuations.
The second study focuses on the over-episode risk control mechanism, demonstrating its critical
role in updating the trading manager agent’s beliefs to provide a more comprehensive understanding
of current trading conditions, as articulated in Table 5. The markedly improved CRs and SRs in
both decision-making scenarios underscore the effectiveness of using CVRF to update investment
beliefs episodically, guiding the agent towards more profitable investment strategies. Additionally,
F IN C ON demonstrates significant learning gains, achieving these results after only four training
episodes—substantially fewer than what is typically required by traditional RL algorithmic trading
agents. More visualizations and analysis are provided in the Appendix A.7.

Task Assets Market Trend Models CR %↑ SR↑ MDD %↓

w/ CVaR 25.077 1.052 17.530
GOOG General Bullish ↗
w/o CVaR -1.461 -0.006 27.079
Single Stock
w/ CVaR 17.461 0.335 40.647
NIO General Bearish ↘
w/o CVaR -52.887 -1.002 70.243
Portfolio (TSLA, w/ CVaR 113.836 3.269 16.163
Mixed
Management MSFT, PFE) w/o CVaR 14.699 1.142 17.511

Table 4: Key metrics F IN C ON with vs. without implementing CVaR for within-episode risk control. The
performance of F IN C ON with the implementation of CVaR won a leading performance in both single-asset
trading and portfolio management tasks.

Task Assets Market Trend Models CR %↑ SR↑ MDD %↓

w/ belief 25.077 1.052 17.530
GOOG General Bullish ↗
w/o belief -11.944 -0.496 29.309
Single Stock
w/ belief 17.461 0.335 40.647
NIO General Bearish ↘
w/o belief 8.197 0.156 55.688
Portfolio (TSLA, w/ belief 113.836 3.269 16.163
Mixed
Management MSFT, PFE) w/o belief 28.432 1.181 27.535

Table 5: Key metrics F IN C ON with vs. without implementing belief updates for over-episode risk control.
The performance of F IN C ON with the implementation of CVRF won a leading performance in both single-asset
trading and portfolio management tasks.

5 Conclusion
In this paper, we present F IN C ON, a novel LLM-based multi-agent framework for financial decision-
making tasks, including single stock trading and portfolio management. Central to F IN C ON is the
Synthesized Manager-Analyst hierarchical communication structure and a dual-level risk control
component. This communication method channels financial data from multiple sources to specialized
analyst agents, who distill it into key investment insights. The manager agent then synthesizes these
insights for decision-making. Our experimental evaluations demonstrate the efficacy of our risk
control mechanism in mitigating investment risks and enhancing trading performance. Additionally,
the streamlined communication structure reduces overhead. The dual-level risk control component
introduces a novel approach to defining agent personas, enabling dynamic updates of risk and
market beliefs within agent communication. A valuable future research direction would be to scale
F IN C ON’s framework to manage large-sized portfolios comprising tens of assets, while maintaining
the impressive decision-making quality demonstrated with smaller portfolios. Given the LLM’s
input length constraint, a critical challenge lies in striking an optimal balance between information
conciseness through agent distillation and potential performance deterioration when extending the
current context window. Addressing this will be essential for ensuring quality-assured outcomes.

10
REFERENCES

References
[1] Harry Markowitz. Portfolio selection. The Journal of Finance, 7(1):77–91, 1952.
[2] Xuan-Hong Dang, Syed Yousaf Shah, and Petros Zerfos. " the squawk bot": Joint learning
of time series and text data modalities for automated financial information filtering. arXiv
preprint arXiv:1912.10858, 2019.
[3] John L Maginn, Donald L Tuttle, Dennis W McLeavey, and Jerald E Pinto. Managing
investment portfolios: a dynamic process, volume 3. John Wiley & Sons, 2007.
[4] Roy Radner. The organization of decentralized information processing. Econometrica: Journal
of the Econometric Society, pages 1109–1146, 1993.
[5] George A Miller. The magical number seven, plus or minus two: Some limits on our capacity
for processing information. Psychological review, 63(2):81, 1956.
[6] Rundong Wang, Hongxin Wei, Bo An, Zhouyan Feng, and Jun Yao. Commission fee is not
enough: A hierarchical reinforced framework for portfolio management. In Proceedings of the
AAAI Conference on Artificial Intelligence, volume 35, pages 626–633, 2021.
[7] Weiguang Han, Boyi Zhang, Qianqian Xie, Min Peng, Yanzhao Lai, and Jimin Huang. Select
and trade: Towards unified pair trading with hierarchical reinforcement learning. In Proceed-
ings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages
4123–4134, 2023.
[8] Molei Qin, Shuo Sun, Wentao Zhang, Haochong Xia, Xinrun Wang, and Bo An. Earnhft:
Efficient hierarchical reinforcement learning for high frequency trading. In Proceedings of the
AAAI Conference on Artificial Intelligence, volume 38, pages 14669–14676, 2024.
[9] Jie Huang and Kevin Chen-Chuan Chang. Towards reasoning in large language models: A
survey. arXiv preprint arXiv:2212.10403, 2022.
[10] Mingyu Jin, Qinkai Yu, Haiyan Zhao, Wenyue Hua, Yanda Meng, Yongfeng Zhang, Mengnan
Du, et al. The impact of reasoning step length on large language models. arXiv preprint
arXiv:2401.04925, 2024.
[11] Tianle Cai, Xuezhi Wang, Tengyu Ma, Xinyun Chen, and Denny Zhou. Large language models
as tool makers. arXiv preprint arXiv:2305.17126, 2023.
[12] Jingqing Ruan, Yihong Chen, Bin Zhang, Zhiwei Xu, Tianpeng Bao, Hangyu Mao, Ziyue
Li, Xingyu Zeng, Rui Zhao, et al. Tptu: Task planning and tool usage of large language
model-based ai agents. In NeurIPS 2023 Foundation Models for Decision Making Workshop,
2023.
[13] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su.
Llm-planner: Few-shot grounded planning for embodied agents with large language models. In
Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998–3009,
2023.
[14] Bhargavi Paranjape, Scott Lundberg, Sameer Singh, Hannaneh Hajishirzi, Luke Zettlemoyer,
and Marco Tulio Ribeiro. Art: Automatic multi-step reasoning and tool-use for large language
models. arXiv preprint arXiv:2303.09014, 2023.
[15] Qianqian Xie, Weiguang Han, Xiao Zhang, Yanzhao Lai, Min Peng, Alejandro Lopez-Lira,
and Jimin Huang. Pixiu: A comprehensive benchmark, instruction dataset and large language
model for finance. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine,
editors, Advances in Neural Information Processing Systems, volume 36, pages 33469–33484.
Curran Associates, Inc., 2023.
[16] Xiao Zhang, Ruoyu Xiang, Chenhan Yuan, Duanyu Feng, Weiguang Han, Alejandro Lopez-
Lira, Xiao-Yang Liu, Meikang Qiu, Sophia Ananiadou, Min Peng, Jimin Huang, and Qianqian
Xie. Dólares or dollars? unraveling the bilingual prowess of financial llms between spanish
and english. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery
and Data Mining, KDD ’24, page 6236–6246, New York, NY, USA, 2024. Association for
Computing Machinery.
[17] Qianqian Xie, Weiguang Han, Zhengyu Chen, Ruoyu Xiang, Xiao Zhang, Yueru He, Mengxi
Xiao, Dong Li, Yongfu Dai, Duanyu Feng, et al. The finben: An holistic financial benchmark
for large language models. arXiv preprint arXiv:2402.12659, 2024.

11
REFERENCES

[18] Gang Hu, Ke Qin, Chenhan Yuan, Min Peng, Alejandro Lopez-Lira, Benyou Wang, Sophia
Ananiadou, Wanlong Yu, Jimin Huang, and Qianqian Xie. No language is an island: Unifying
chinese and english in financial large language models, instruction data, and benchmarks.
arXiv preprint arXiv:2403.06249, 2024.
[19] Yuzhe Yang, Yifei Zhang, Yan Hu, Yilin Guo, Ruoli Gan, Yueru He, Mingcong Lei, Xiao
Zhang, Haining Wang, Qianqian Xie, et al. Ucfe: A user-centric financial expertise benchmark
for large language models. arXiv preprint arXiv:2410.14059, 2024.
[20] Joon Sung Park, Joseph C O’Brien, Carrie J Cai, Meredith Ringel Morris, Percy Liang, and
Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. arXiv
preprint arXiv:2304.03442, 2023.
[21] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li,
Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via
multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023.
[22] Shuofei Qiao, Ningyu Zhang, Runnan Fang, Yujie Luo, Wangchunshu Zhou, Yuchen Eleanor
Jiang, Chengfei Lv, and Huajun Chen. Autoact: Automatic agent learning from scratch via
self-planning. arXiv preprint arXiv:2401.05268, 2024.
[23] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili
Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for
multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023.
[24] Xudong Guo, Kaixuan Huang, Jiale Liu, Wenhui Fan, Natalia Vélez, Qingyun Wu, Huazheng
Wang, Thomas L Griffiths, and Mengdi Wang. Embodied llm agents learn to cooperate in
organized teams. arXiv preprint arXiv:2403.12482, 2024.
[25] Junkai Li, Siyu Wang, Meng Zhang, Weitao Li, Yunghwei Lai, Xinhui Kang, Weizhi Ma, and
Yang Liu. Agent hospital: A simulacrum of hospital with evolvable medical agents. arXiv
preprint arXiv:2405.02957, 2024.
[26] Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu,
Yaxi Lu, Yi-Hsin Hung, Chen Qian, et al. Agentverse: Facilitating multi-agent collaboration
and exploring emergent behaviors. In The Twelfth International Conference on Learning
Representations, 2023.
[27] Reid Pryzant, Dan Iter, Jerry Li, Yin Tat Lee, Chenguang Zhu, and Michael Zeng. Au-
tomatic prompt optimization with" gradient descent" and beam search. arXiv preprint
arXiv:2305.03495, 2023.
[28] Xinyu Tang, Xiaolei Wang, Wayne Xin Zhao, Siyuan Lu, Yaliang Li, and Ji-Rong Wen.
Unleashing the potential of large language models as prompt optimizers: An analogical
analysis with gradient-based model optimizers. arXiv preprint arXiv:2402.17564, 2024.
[29] Weiran Yao, Shelby Heinecke, Juan Carlos Niebles, Zhiwei Liu, Yihao Feng, Le Xue, Rithesh
Murthy, Zeyuan Chen, Jianguo Zhang, Devansh Arpit, et al. Retroformer: Retrospective large
language agents with policy gradient optimization. arXiv preprint arXiv:2308.02151, 2023.
[30] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Re-
flexion: Language agents with verbal reinforcement learning. Advances in Neural Information
Processing Systems, 36, 2024.
[31] Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. Fingpt: Open-source financial
large language models. arXiv preprint arXiv:2306.06031, 2023.
[32] Yangyang Yu, Haohang Li, Zhi Chen, Yuechen Jiang, Yang Li, Denghui Zhang, Rong Liu,
Jordan W Suchow, and Khaldoun Khashanah. Finmem: A performance-enhanced llm trading
agent with layered memory and character design. arXiv preprint arXiv:2311.13743, 2023.
[33] Wentao Zhang, Lingxuan Zhao, Haochong Xia, Shuo Sun, Jiaze Sun, Molei Qin, Xinyi Li,
Yuqing Zhao, Yilei Zhao, Xinyu Cai, et al. Finagent: A multimodal foundation agent for finan-
cial trading: Tool-augmented, diversified, and generalist. arXiv preprint arXiv:2402.18485,
2024.
[34] Freddy Delbaen and Sara Biagini. Coherent risk measures. Springer, 2000.
[35] Frank J Fabozzi, Sergio M Focardi, and Petter N Kolm. Quantitative equity investing: Tech-
niques and strategies. John Wiley & Sons, 2010.

12
REFERENCES

[36] Chong Zhang, Xinyi Liu, Mingyu Jin, Zhongmou Zhang, Lingyao Li, Zhengting Wang,
Wenyue Hua, Dong Shu, Suiyuan Zhu, Xiaobo Jin, et al. When ai meets finance (stockagent):
Large language model-based stock trading in simulated real-world environments. arXiv
preprint arXiv:2407.18957, 2024.
[37] Keith Kuester, Stefan Mittnik, and Marc S Paolella. Value-at-risk prediction: A comparison of
alternative strategies. Journal of Financial Econometrics, 4(1):53–89, 2006.
[38] Matthijs TJ Spaan. Partially observable markov decision processes. In Reinforcement learning:
State-of-the-art, pages 387–414. Springer, 2012.
[39] Frank J Fabozzi, Harry M Markowitz, and Francis Gupta. Portfolio selection. Handbook of
finance, 2, 2008.
[40] Yang Liu, Qi Liu, Hongke Zhao, Zhen Pan, and Chuanren Liu. Adaptive quantitative trading:
An imitative deep reinforcement learning approach. In Proceedings of the AAAI conference on
artificial intelligence, volume 34, pages 2128–2135, 2020.
[41] Taylan Kabbani and Ekrem Duman. Deep reinforcement learning approach for trading
automation in the stock market. IEEE Access, 10:93564–93574, 2022.
[42] Thomas L Griffiths, Jian-Qiao Zhu, Erin Grant, and R Thomas McCoy. Bayes in the age of
intelligent machines. arXiv preprint arXiv:2311.10206, 2023.
[43] Ashwin Rao and Tikhon Jelvis. Foundations of reinforcement learning with applications in
finance. Chapman and Hall/CRC, 2022.
[44] Theodore Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths. Cognitive
architectures for language agents, 2023.
[45] Jintian Zhang, Xin Xu, and Shumin Deng. Exploring collaboration mechanisms for llm agents:
A social psychology view. arXiv preprint arXiv:2310.02124, 2023.
[46] Anthony D Wagner. Working memory contributions to human learning and remembering.
Neuron, 22(1):19–22, 1999.
[47] Harry M Markowitz and G Peter Todd. Mean-variance analysis in portfolio choice and capital
markets, volume 66. John Wiley & Sons, 2000.
[48] Xiao-Yang Liu, Hongyang Yang, Qian Chen, Runjia Zhang, Liuqing Yang, Bowen Xiao, and
Christina Dan Wang. Finrl: A deep reinforcement learning library for automated stock trading
in quantitative finance. arXiv preprint arXiv:2011.09607, 2020.
[49] Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflexion: an autonomous agent with
dynamic memory and self-reflection. arXiv preprint arXiv:2303.11366, 2023.
[50] Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. Expel:
Llm agents are experiential learners. In Proceedings of the AAAI Conference on Artificial
Intelligence, volume 38, pages 19632–19642, 2024.
[51] Yujia Li, David Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, Rémi Leblond,
Tom Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, et al. Competition-level code
generation with alphacode. Science, 378(6624):1092–1097, 2022.
[52] Bei Chen, Fengji Zhang, Anh Nguyen, Daoguang Zan, Zeqi Lin, Jian-Guang Lou, and Weizhu
Chen. Codet: Code generation with generated tests. arXiv preprint arXiv:2207.10397, 2022.
[53] Chen Qian, Xin Cong, Cheng Yang, Weize Chen, Yusheng Su, Juyuan Xu, Zhiyuan Liu,
and Maosong Sun. Communicative agents for software development. arXiv preprint
arXiv:2307.07924, 2023.
[54] Zelai Xu, Chao Yu, Fei Fang, Yu Wang, and Yi Wu. Language agents with reinforcement
learning for strategic play in the werewolf game. arXiv preprint arXiv:2310.18940, 2023.
[55] Weiyu Ma, Qirui Mi, Xue Yan, Yuqiao Wu, Runji Lin, Haifeng Zhang, and Jun Wang. Large
language models play starcraft ii: Benchmarks and a chain of summarization approach. arXiv
preprint arXiv:2312.11865, 2023.
[56] J de Curtò, I de Zarzà, Gemma Roig, Juan Carlos Cano, Pietro Manzoni, and Carlos T Calafate.
Llm-informed multi-armed bandit strategies for non-stationary environments. Electronics,
12(13):2814, 2023.

13
REFERENCES

[57] Huaqin Zhao, Zhengliang Liu, Zihao Wu, Yiwei Li, Tianze Yang, Peng Shu, Shaochen Xu,
Haixing Dai, Lin Zhao, Gengchen Mai, et al. Revolutionizing finance with llms: An overview
of applications and insights. arXiv preprint arXiv:2401.11641, 2024.
[58] Zhiwei Liu, Xin Zhang, Kailai Yang, Qianqian Xie, Jimin Huang, and Sophia Ananiadou.
Fmdllama: Financial misinformation detection based on large language models. arXiv preprint
arXiv:2409.16452, 2024.
[59] Yupeng Cao, Zhi Chen, Qingyun Pei, Fabrizio Dimino, Lorenzo Ausiello, Prashant Kumar,
KP Subbalakshmi, and Papa Momar Ndiaye. Risklabs: Predicting financial risk using large
language model based on multi-sources data. arXiv preprint arXiv:2404.07452, 2024.
[60] Qianqian Xie, Dong Li, Mengxi Xiao, Zihao Jiang, Ruoyu Xiang, Xiao Zhang, Zhengyu Chen,
Yueru He, Weiguang Han, Yuzhe Yang, et al. Open-finllms: Open multimodal large language
models for financial applications. arXiv preprint arXiv:2408.11878, 2024.
[61] Lorenzo Canese, Gian Carlo Cardarilli, Luca Di Nunzio, Rocco Fazzolari, Daniele Giardino,
Marco Re, and Sergio Spanò. Multi-agent reinforcement learning: A review of challenges and
applications. Applied Sciences, 11(11):4948, 2021.
[62] Kaiqing Zhang, Zhuoran Yang, and Tamer Başar. Multi-agent reinforcement learning: A
selective overview of theories and algorithms. Handbook of reinforcement learning and control,
pages 321–384, 2021.
[63] Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson. Learning
to communicate with deep multi-agent reinforcement learning. Advances in neural information
processing systems, 29, 2016.
[64] Toru Lin, Jacob Huh, Christopher Stauffer, Ser Nam Lim, and Phillip Isola. Learning to ground
multi-agent communication with autoencoders. Advances in Neural Information Processing
Systems, 34:15230–15242, 2021.
[65] Woojun Kim, Jongeui Park, and Youngchul Sung. Communication in multi-agent reinforce-
ment learning: Intention sharing. In International Conference on Learning Representations,
2020.
[66] Changxi Zhu, Mehdi Dastani, and Shihan Wang. A survey of multi-agent reinforcement
learning with communication. arXiv preprint arXiv:2203.08975, 2022.
[67] Zhiyuan Yao, Zheng Li, Matthew Thomas, and Ionut Florescu. Reinforcement learning in
agent-based market simulation: Unveiling realistic stylized facts and behavior. arXiv preprint
arXiv:2403.19781, 2024.
[68] HaoHang Li and Steve Y Yang. Impact of false information from spoofing strategies: An
abm model of market dynamics. In 2022 IEEE Symposium on Computational Intelligence for
Financial Engineering and Economics (CIFEr), pages 1–10. IEEE, 2022.
[69] Kuan Wang, Yadong Lu, Michael Santacroce, Yeyun Gong, Chao Zhang, and Yelong Shen.
Adapting llm agents through communication. arXiv preprint arXiv:2310.01444, 2023.
[70] Zhao Mandi, Shreeya Jain, and Shuran Song. Roco: Dialectic multi-robot collaboration with
large language models. arXiv preprint arXiv:2307.04738, 2023.
[71] Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum,
Tianmin Shu, and Chuang Gan. Building cooperative embodied agents modularly with large
language models. arXiv preprint arXiv:2307.02485, 2023.
[72] Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving
factuality and reasoning in language models through multiagent debate. arXiv preprint
arXiv:2305.14325, 2023.
[73] Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu,
and Zhiyuan Liu. Chateval: Towards better llm-based evaluators through multi-agent debate.
arXiv preprint arXiv:2308.07201, 2023.
[74] Frank Xing. Designing heterogeneous llm agents for financial sentiment analysis. arXiv
preprint arXiv:2401.05799, 2024.
[75] Irene de Zarzà i Cubero, Joaquim de Curtò i Díaz, Gemma Roig, and Carlos T Calafate.
Optimized financial planning: Integrating individual and cooperative budgeting models with
llm recommendations. AI, 5(1):91–114, 2024.

14
REFERENCES

[76] Xiangpeng Wan, Haicheng Deng, Kai Zou, and Shiqi Xu. Enhancing the efficiency and
accuracy of underlying asset reviews in structured finance: The application of multi-agent
framework. arXiv preprint arXiv:2405.04294, 2024.
[77] Chong Zhang, Xinyi Liu, Mingyu Jin, Zhongmou Zhang, Lingyao Li, Zhengting Wang,
Wenyue Hua, Dong Shu, Suiyuan Zhu, Xiaobo Jin, et al. When ai meets finance (stockagent):
Large language model-based stock trading in simulated real-world environments. arXiv
preprint arXiv:2407.18957, 2024.
[78] Patrick Bolton and Mathias Dewatripont. The firm as a communication network. The Quarterly
Journal of Economics, 109(4):809–839, 1994.
[79] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and
Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint
arXiv:2210.03629, 2022.
[80] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le,
Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.
Advances in neural information processing systems, 35:24824–24837, 2022.
[81] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik
Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.
Advances in Neural Information Processing Systems, 36, 2024.
[82] Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and
Zhiting Hu. Reasoning with language model is planning with world model. arXiv preprint
arXiv:2305.14992, 2023.
[83] Xizhou Zhu, Yuntao Chen, Hao Tian, Chenxin Tao, Weijie Su, Chenyu Yang, Gao Huang, Bin
Li, Lewei Lu, Xiaogang Wang, et al. Ghost in the minecraft: Generally capable agents for
open-world enviroments via large language models with text-based knowledge and memory.
arXiv preprint arXiv:2305.17144, 2023.
[84] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen,
Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous
agents. Frontiers of Computer Science, 18(6):1–26, 2024.
[85] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A
survey. Journal of artificial intelligence research, 4:237–285, 1996.
[86] Guojun Xiong, Shufan Wang, Daniel Jiang, and Jian Li. Personalized federated reinforce-
ment learning with shared representations. In Deployable RL: From Research to Practice@
Reinforcement Learning Conference 2024, 2024.
[87] Guojun Xiong, Ujwal Dinesha, Debajoy Mukherjee, Jian Li, and Srinivas Shakkottai. Dopl:
Direct online preference learning for restless bandits with preference feedback. arXiv preprint
arXiv:2410.05527, 2024.
[88] Zhiyuan Yao, Ionut Florescu, and Chihoon Lee. Control in stochastic environment with
delays: A model-based reinforcement learning approach. In Proceedings of the International
Conference on Automated Planning and Scheduling, volume 34, pages 663–670, 2024.
[89] Bin Zhang, Hangyu Mao, Jingqing Ruan, Ying Wen, Yang Li, Shao Zhang, Zhiwei Xu, Dapeng
Li, Ziyue Li, Rui Zhao, et al. Controlling large language model-based agents for large-scale
decision-making: An actor-critic approach. arXiv preprint arXiv:2311.13884, 2023.
[90] John Hull. Risk Management and Financial Institutions. John Wiley & Sons, 2007.
[91] William F. Sharpe. The sharpe ratio. The Journal of Portfolio Management, 21(1):49–58,
1994.
[92] Andrew Ang and Joseph Chen. Downside risk. Journal of Portfolio Management, 29(4):103–
112, 2003.
[93] Xiao-Yang Liu, Guoxuan Wang, and Daochen Zha. Fingpt: Democratizing internet-scale data
for financial large language models. arXiv preprint arXiv:2307.10485, 2023.
[94] Yu Qin and Yi Yang. What you say and how you say it matters: Predicting stock volatility
using verbal and vocal cues. In Proceedings of the 57th Annual Meeting of the Association for
Computational Linguistics, pages 390–401, 2019.

15
REFERENCES

[95] Linyi Yang, Tin Lok James Ng, Barry Smyth, and Riuhai Dong. Html: Hierarchical
transformer-based multi-task learning for volatility prediction. In Proceedings of The Web
Conference 2020, pages 441–451, 2020.
[96] Yupeng Cao, Zhi Chen, Qingyun Pei, Prashant Kumar, KP Subbalakshmi, and Papa Momar
Ndiaye. Ecc analyzer: Extract trading signal from earnings conference calls using large
language model for stock performance prediction. arXiv preprint arXiv:2404.18470, 2024.
[97] John C Hull. Options, Futures, and Other Derivatives. Pearson Education, 2017.
[98] Xiao-Yang Liu, Hongyang Yang, Jiechao Gao, and Christina Dan Wang. FinRL: Deep rein-
forcement learning framework to automate trading in quantitative finance. ACM International
Conference on AI in Finance (ICAIF), 2021.
[99] Xiao-Yang Liu, Ziyi Xia, Jingyang Rui, Jiechao Gao, Hongyang Yang, Ming Zhu, Christina
Wang, Zhaoran Wang, and Jian Guo. Finrl-meta: Market environments and benchmarks for
data-driven financial reinforcement learning. Advances in Neural Information Processing
Systems, 35:1835–1849, 2022.
[100] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap,
Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforce-
ment learning. In International conference on machine learning, pages 1928–1937. PMLR,
2016.
[101] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal
policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.
[102] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan
Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint
arXiv:1312.5602, 2013.
[103] Yuliya Plyakha, Raman Uppal, and Grigory Vilkov. Why does an equal-weighted portfolio
outperform value-and price-weighted portfolios? Available at SSRN 2724535, 2012.
[104] Jaap MJ Murre and Joeri Dros. Replication and analysis of ebbinghaus’ forgetting curve. PloS
one, 10(7):e0120644, 2015.
[105] Guardrails ai. [Link] Open source library for interacting with
Large Language Models.

16
A Appendix
A.1 Related Work

LLM Agents for Financial Decision Making. There are considerable efforts towards developing
general-purpose LLM agent for sequential decision-making [49, 50], and such type of tasks often
involve episodic interactions with environment and verbal reflections for action refinement, such as
coding competition [51, 52], software development [53, 23], game-playing [54, 55]. Furthermore,
researchers have started to exploit how LLM agents can perform better in harder decision-making
tasks from finance [56, 57, 58, 59, 60], in which there are more volatile environments, leading to
that the numerous unpredictable elements can obscure an agent’s ability to reflect accurately on the
reasons for poor decision outcomes. FinMem [32] enhances single stock trading performance by
embedding memory modules with LLM agent for reflection-refinement, and FinAgent [33] improved
trading profits via using external quantitative tool to fight against volatile environment.
Multi-Agent System and Communication Structures. In traditional multi-agent systems [61, 62],
the way for agents’ communication is pre-determined, like sharing data or state observations [63, 64,
65, 66, 67, 68]. The emergence of large language model brings flexibility for human-understandable
communications [69, 20, 23, 70], so some work tries to elevate decision-making ability of LLM-
based multi-agent system by letting agents engage in discussions [71, 21] or debates [72, 73]. The
similar peer-communication strategy was as well utilized by the multi-agent system for financial tasks
[74, 75, 76]. However, such approach are not optimal for unified-goal financial tasks that prioritize
profits [77], because they suffer from potentially ambiguous optimization objectives and are unable
to control the unnecessary communication costs [78].
Prompt Optimization and Verbal Reinforcement. To enhance the reasoning or decision-making
of LLM agents, many prompt optimization techniques have been proposed, like ReAct [79], Chain
of Thought (CoT) [80], Tree of Thoughts (ToT) [81], ART [14], intended for that LLM agents can
automatically generate intermediate reasoning steps as an iterative program. In addition, to make
LLM agents make decisions like humans and generate more understandable reasoning texts, some
researchers recommend incorporating cognitive structures [82, 83, 44, 84]. Inspired by these previous
work and DRL algorithms [85, 86, 87, 67, 88], verbal reinforcement [29, 30, 89, 24] was developed
for LLM agents such that they can update actions based on iterative self-reflection while integrating
additional LLM as a prompt optimizer [27, 28].

A.2 Textual Gradient-Descent

In an LLM-based prompt optimizer, a meta-prompt [27, 28] is used to refine the task prompt for
better performance. For example, for a mathematical reasoning task, the task prompt might be "Let’s
solve the problem," while the meta-prompt could be "Improve the prompt to help a model better
perform mathematical reasoning."
Although prompt optimization lacks explicit gradients to control the update direction, we can simulate
“textual gradient” by using LLMs’ reflection capabilities. By generating feedback from past successes
and failures on trading decisions, LLMs can produce "semantic" gradient signals that guide the
optimization process.
Adjusting the optimization process’s direction is crucial, similar to tuning the learning rate in
traditional parameter optimization. An inappropriate learning rate can cause the process to oscillate
or converge too slowly. Similarly, without proper control, the LLM-based optimizer might overshoot
or oscillate during prompt optimization.
To mimic learning rate effects, we measure the overlapping percentage between trading decision
sequences from consecutive iterations. We then directly edit the previous task prompt to enhance
performance. The meta-prompt instructs the LLM to modify the current prompt based on feedback,
ensuring a stable and incremental improvement process. This method allows for effective exploitation
of existing prompts, leading to gradual performance enhancement.

17
A.3 F IN C ON Testing Stage Workflow

A.3 F IN C ON Testing Stage Workflow

During the testing stage, F IN C ON will utilize the investment beliefs learned from the training stage,
and the over-episode risk control mechanism will no longer operate. However, the within-episode
risk control mechanism will still function, allowing the manager agent to adjust trading actions in
real-time based on short-term trading performance and market fluctuations. This ensures that even
during testing, F IN C ON can promptly respond to market risks and potentially prevent losses while
leveraging the knowledge gained during training.

Algorithm 2 Testing Stage Algorithm of F IN C ON

Initialize trading start date s, stock pool of portfolio and portfolio weights w0 = 0.
i
Inherit manager-analysts component {Mpr }Ii=1 &Ma .
Inherit the reflections B, learned prompts θ, the trained policy Πθ .
for T + 1 ≤ t ≤ S do
Run policy Πθ (collecting daily PnL rt , portfolio weights wt and daily CVaR value ρt ).
if ρt < ρt−1 or rt < 0 then
Trigger Ma self-reflection.
end if
Get one investment trajectory H.
end for
Output performance metrics calculation results based on H.

18
A.4 Figure of Modular Design of Agents in F IN C ON

A.4 Figure of Modular Design of Agents in F IN C ON

General Configuration
1. Investment task introduction 2. Trading target background
3. Trading sectors 4. Historical financial performance overview

Profiling Module
Manager Agent:
1. Role assignment: You are an experienced trading manager in the investment firm ...
2. Role description: Your responsibilities are to consolidate investment insights from analysts
and make trading actions on {asset symbols} ...
Analyst Agents:
1. Role assignment: You are the investment analysts for news/ market data/ Form 10-K (Q)/
ECC audio recording ...
2. Role duty description: Your responsibilities are to distill investment insights and other
indicators like financial sentiment for {asset symbols} ...

Perception Module
Manager Agent:
1. Perceive: Investment insights from analyst agents; daily risk alert and episode-level
investment belief updates from the risk control component.
2. Send: Feedback to analyst agents about their contribution to significant investment
earnings & losses.
Analyst Agents:
1. Perceive: Market information from certain information sources.
2. Send: Relevant market insights to manager agent.
3. Receive: Feedback from the manager agent.

Memory Module
Manager Agent:
1. Working: - Consolidation -Refinement
2. Procedural: - Trading action records - Reflection records
3. Episodic: - Trajectory history
Analyst Agents:
1. Working: - Observation - Retrieval - Distillation
2. Procedural: - Distilled Investment-related insights - Financial sentiment
- Investment report recommended actions

Action Module
Manager Agent:
1. Conduct: Trading actions.
2. Reflect: Trading reasons and analyst agents’ contribution assessment.

Figure 4: The detailed modular design of the manager and analyst agents. The general configuration
and profiling modules generate text-based queries to retrieve investment-related information from the
agents’ memory databases. The perceptual and memory modules interact with LLMs via prompts to
extract key investment insights. The action module of the manager agent consolidates these insights
to facilitate informed trading decisions.

19
A.5 Experimental Setup

A.5 Experimental Setup

Multi-Modal Datasets. We collect a comprehensive multi-modal dataset to simulate a realistic market

environment. This dataset includes stock price data, daily financial news, company filing reports
(10K and 10Q), and ECC (Earnings Call Conference) audio recordings spanning from January 3,
2022, to June 10, 2023. Each data source is assigned to specific analyst agents based on the timeliness
of the information. For example, annual filings (10K) exhibit longer-term persistence, quarterly
filings (10Q) and ECC data have medium-term relevance, and daily financial news provides the most
immediate information.
Evaluation Metrics. We evaluate F IN C ON and benchmark it against other state-of-the-art LLM-based
and DRL-based agent systems using three key financial performance metrics: Cumulative Return
(CR%) [90], Sharpe Ratio (SR) [91], and Max Drawdown (MDD%) [92]. These metrics help quantify
each model’s profitability, risk-adjusted returns, and risk management performance, respectively.
Comparative Methods. In the single-stock trading task, we compare F IN C ON against seven algo-
rithmic agents and the widely accepted Buy-and-Hold (B & H) baseline. The three DRL-based
agents—A2C, PPO, and DQN—are from the FinRL framework [48], while the four state-of-the-art
LLM-based agents include G ENERATIVE AGENT [20], F IN GPT [93], F IN M EM [32], and F INA-
GENT [33]. For portfolio management, we benchmark F IN C ON against the classical Markowitz MV
portfolio selection strategy [1], the RL-based FinRL-A2C agent [48], and the B & H strategy, which
holds an equal-weighted position across all assets (equal-weighted ETF). Our focus on classical,
RL-based, and B & H methods is due to the current lack of mature LLM-based agents for portfolio
management tasks.
Implementation Details. In our experiments, all LLM-based agent systems, including F IN C ON, use
GPT-4-Turbo as the backbone model, with the temperature parameter set at 0.3 to balance response
consistency with creative reasoning. F IN C ON is trained on financial data from January 3, 2022, to
October 4, 2022, and tested on data from October 5, 2022, to June 10, 2023. Since deep reinforcement
learning (DRL) agents require extensive data for convergence, their training period is extended to
nearly five years (January 1, 2018, to October 4, 2022) to ensure fair comparison. The testing period
remains the same across all models. The final performance metrics are based on the test trajectory
with the median CR and SR values from five repeated epochs. If the median CR and SR occur in
different epochs, performance is assessed based on the trajectory with the median CR value.

20
A.6 Single Stock Trading Result Graphs

A.6 Single Stock Trading Result Graphs

Figure 5: CRs over time for single-asset trading tasks. F IN C ON outperformed other comparative strategies,
achieving the highest CRs across all six stocks by the end of the testing period, regardless of market conditions.

A.7 Detailed Ablation Study

21
A.7 Detailed Ablation Study

A.7.1 The Effectiveness of Within-Episode Risk Control mechanism via CVaR

To answer the RQ2, we conduct the first ablation study. We assess the efficacy of F IN C ON’s within-
episode risk control mechanisms by monitoring system risk changes through CVaR. To demonstrate
the robustness of F IN C ON, we compare the performance of F IN C ON with versus without CVaR
implementation across two task types: single-asset trading and portfolio management. Furthermore,
in single-asset trading tasks, we consider assets in both general bullish and bearish market conditions
in the testing phase for comprehensive consideration.
Our results demonstrate that implementing CVaR in F IN C ON is highly effective across all finan-
cial metrics for both task types, as shown in Table 4 and Fig 6. For single-asset trading tasks,
F IN C ON without within-episode risk control yields negative CRs and significantly higher MDDs,
underperforming compared to the Buy-and-Hold strategy (CR of GOOG: 22.42%, CR of NIO:
−77.210%), highlighting the severe consequences of ignoring environmental risks. In portfolio
management, the CR increases dramatically from 14.699% to 113.836% with within-episode risk
control, demonstrating its effectiveness in risk supervision even amid non-uniform market trends.

(a) [1] Single Stock: General Bullish (b) [2] Single Stock: General Bearlish

(c) Multi-Assets
Figure 6: CRs of F IN C ON with vs. without implementing CVaR for within-episode risk control show that
the CVaR mechanism significantly improves F IN C ON’s performance. This is evident from two metrics: (a)
cumulative returns over time for single stocks in both bullish and bearish market conditions, and (b) portfolio
value over time for a multi-asset portfolio. In both cases, F IN C ON with CVaR demonstrates substantially higher
gains.

Specifically, the success of utilizing CVaR for within-episode risk control is evident in both bullish
and bearish market environments, as shown in the single asset trading case. In bullish markets,
CVaR sharply captures immediate market shocks and timely informs F IN C ON to exercise caution,
even amidst general optimism. Conversely, in bearish markets, CVaR consistently alerts F IN C ON to
significant price drops, ensuring awareness of market risks. Moreover, in portfolio trading with mixed
price trends, our within-episode risk control mechanism performs robustly by monitoring the entire
portfolio’s value fluctuations, enabling the trading manager agent to adjust potentially aggressive
operations for each asset promptly.

22
A.7 Detailed Ablation Study

A.7.2 The Efficacy of Over-Episode Belief Updates Using CVRF

In the second ablation study, to answer RQ3, we use the same assets to examine the effectiveness
of F IN C ON’s over-episode risk control mechanisms. This is achieved by consistently improving
F IN C ON’s beliefs about market conditions for the targeted assets. To ensure consistent belief output
for each training episode, we set the temperature parameter to 0 specifically for belief generation.
We collect third-time belief updates over four training episodes using our innovative CVRF mecha-
nism. The overlap of trading actions between the last two adjacent episodes increases to over 80%,
and the updated investment beliefs are mostly aligned. To illustrate F IN C ON’s evolving investment
beliefs through iterative training, we use the GOOG investment belief update as an example, as
shown in Figure 8. Compared to the initial and final belief updates, each conceptual aspect, such as
historical momentum and news insights, is enriched with executable information through our CRVF
mechanism, leading to more profitable actions.
The results in Table 5 and Figure 7 indicate that the over-episode belief update mechanism is more
critical than within-episode risk control in enhancing F IN C ON’s decision-making. Without this
functionality, key metrics like CR, SR, and MDD are lower than without the within-episode risk
control in single asset trading. Although the CR of 28.432% outperforms the Equal-Weighted ETF
strategy’s 9.344% for portfolio management, the SR of 1.181 is higher than Equal-Weighted ETF
strategy’s 0.492, with the belief update feature, performance significantly further improves. It can
achieve a CR of 113.836% and an SR of 3.269. These results demonstrate that using CVRF to
update investment beliefs over episodes efficiently steers the agent’s investment beliefs towards more
profitable directions. F IN C ON achieves superior performance on multiple tasks, with learning gains
evident after just four training episodes, requiring far fewer episodes than traditional RL algorithmic
trading agents.

(a) [1] Single Stock: General Bullish (b) [2] Single Stock: General Bearlish

(c) Multi-Assets
Figure 7: CRs of F IN C ON with vs. without belief updates for over-episode risk control. (a) The CRs over
time for single stocks. The performance of F IN C ON with belief updates consistently leads in both bullish and
bearish market conditions. (b) The portfolio values over time for multi-asset portfolio. F IN C ON’s performance
with belief updates also won a substantially higher gains.

23
A.7 Detailed Ablation Study

Updated Investment Belief Contents

First Update:
{’historical momentum’: ’Enhance the use of momentum indicators by establishing systematic rules for
entry and exit based on momentum values to improve consistency and predictability in trading actions.’,
’news insights’: ’Integrate advanced real-time news sentiment analysis to better understand immediate
market reactions and adjust trading strategies accordingly...’,
’Form 10-Q’: ’Incorporating annual report insights (10-K) could provide a comprehensive view of the
company’s long-term performance and strategic direction, aiding in more informed decision-making.’...,
’other aspects’: [’sector trends’, ’technological advancements’], ...}
Last Update:
{’historical momentum’: ’The more profitable strategy effectively utilizes negative momentum values to
make timely sell decisions, avoiding potential losses during downward trends, and effectively utilizes
positive momentum values to make timely buy decisions, facilitating potential gains during upward
trends. It is suggested to refine the use of momentum indicators, possibly by adjusting thresholds or
incorporating additional short-term momentum metrics to enhance the timing and accuracy of buy/sell
decisions.’,
’news insights’: ’It is recommended to further develop a systematic approach to quantify the impact of
news, especially the sentiment scores and regulatory changes, to refine trading decisions.’,
’Form 10-Q’: ’It is suggested that even if there is a strong financial performance present 10-Q reports, it
is better to prioritize immediate market signals. For future strategies, it is suggested to balance these
aspects more evenly, especially in stable or bullish market conditions, to avoid premature exits from
potentially profitable positions.’,...,
’other aspects’: [’sector trends’, ’technological advancements’, ’macroeconomic factors’, ’regulatory
risks’], ...}

Tradng Action Overlapping Percentages

1. Second vs. First Training Episode: 46.939%
2. Third vs. Second Training Episode: 71.429%
3. Fourth vs. Third Training Episode: 81.633%

Figure 8: The first time and last time LLM generated investment belief updates by CVRF for GOOG.

24
A.8 Raw Data Sources

A.8 Raw Data Sources

We assessed the performance of F IN C ON using multi-modal financial data from January 3, 2022, to
June 10, 2022, sourced from reputable databases and APIs including Yahoo Finance (via yfinance),
Alpaca News API, and Capital IQ, detailed explained in Table. These data, initially stored in the Raw
Financial Data Warehouse as available observations of the financial market environment, are diverged
into the corresponding F IN C ON’s Analysts’ Procedural Memory Databases based on timeliness
through working memory’s summarization operation.

Data Sources
News data associated with ticker: News data is sourced from REFINITIV REAL-TIME NEWS mainly
contains news from Reuters.
Form 10-Q, Part 1 Item 2 (Management’s Discussion and Analysis of Financial Condition and Results of
Operations): Quarterly reports (Form 10-Q) are required by the U.S. Securities and Exchange Commission
(SEC).
Form 10-k, Section 7 (Management’s Discussion and Analysis of Financial Condition and Results of
Operations): Annual reports (Form 10-K) are required by the U.S. Securities and Exchange Commission
(SEC), sourced from EDGAR, and downloaded via SEC API.
Historical stock price: Daily open price, high price, close price, adjusted close price, and volume data from
Yahoo Finance.
Zacks Equity Research:
Zacks Rank: The Zacks Rank is a short-term rating system that is most effective over the one- to three-month
holding horizon. The underlying driver for the quantitatively determined Zacks Rank is the same as the
Zacks Recommendation and reflects trends in earnings estimate revisions.
Zacks Analyst: Reason to Sell, Reason to Buy, and potential risks.
Earning Conference Calls (ECC): ECC is a type of unstructured financial data (audio) that is crucial for
understanding market dynamics and investor sentiment. The company executive board delivers ECC about
recent financial outcomes, future projections, and strategic directions. Recent studies have underscored the
importance of not only the textual content of these calls but also the audio feature. Analyses have revealed
that the audio elements—such as tone, pace, and inflections—offer significant predictive value regarding
company performance and stock movements [94, 95, 96].
Table 6: Raw data and memory warehouses of F IN C ON

A.9 Distribution of Data

Figure 9: The distribution of news from REFINITIV REAL-TIME NEWS for the 42 stocks in the
experiments

25
A.10 Formulas of Classic Financial Metrics for Risk Estimator and Decision-making Task Performance
Evaluation

Figure 10: The distribution of 10k10q from Securities and Exchange Commission (SEC) for the 42
stocks in the experiments

Figure 11: The distribution of Analyst Report from Zacks Equity Research for the 42 stocks in the
experiments

A.10 Formulas of Classic Financial Metrics for Risk Estimator and Decision-making Task
Performance Evaluation

The risk estimator uses the following metrics:

Profit and Loss (PnL)[97]: PnL quantifies the net outcome of trading activities over a specified
period by accounting for the realized gains and losses from financial instruments like stocks and
derivatives.
Value at Risk (VaR) of PnL[97]: VaR is a statistical tool used to estimate the potential loss in a
portfolio, within a defined confidence interval. Mathematically, it is defined as Equation 3:
VaRα (P nL) = inf {l ∈ R : P(P nL ≤ l) ≥ α} (3)
where α is the confidence level.
Conditional Value at Risk (CVaR) of PnL[97]: CVaR is a statistical tool used to estimate the
expected potential loss worse than the VaR value in a portfolio, within a defined confidence interval.
Mathematically, it is defined as Equation 4:
n o
CVaRα (P nL) = E P nL|P nL ≤ VaRα (P nL) (4)

26
A.11 Baseline and Comparative Models on Single Stock Trading Task

where α is the confidence level.

The performance evaluation of algorithmic trading agents incorporates the following metrics:
Cumulative Return of PnL [90]: Cumulative Return is a key trading performance metric because
it provides a comprehensive insight into investment performance, especially for strategies that
emphasize long-term growth and reinvestment. The effectiveness of different investment strategies is
evaluated based on their Cumulative Returns, which reflect the total change in value over time. In
this study, we compute Cumulative Returns over the specified period by summing daily logarithmic
returns, as outlined in Equation 5. This method is widely accepted in the finance area due to its ability
to precisely capture minor price fluctuations and symmetrically address gains and losses. In essence,
a higher Cumulative Return typically indicates a more effective strategy.

n
X
Cumulative Return = ri
t=1
n    
X pt+1
= ln · actiont , (5)
t=1
pt

where ri represents the PnL for day t + 1, pt is the closing price on day t, pt+1 is the closing price
on day t + 1, and actiont denotes the trading decision made by the model for that day.
Portfolio Value: Portfolio value represents the total worth of all the investments held in a portfolio at
a given point in time. It is a metric used only in the portfolio management task.
t
Y
Cumulative Simple Returnt = (1 + Daily Simple Returnt ) − 1 (6)
k=1

Portfolio Valuet = Initial Investment Amount × (1 + Cumulative Simple Returnt ) (7)

, where the initial amount is set as $1, 000, 000.
Sharpe Ratio of PnL [91]: Sharpe Ratio is another core metric for evaluating investment performance
and adjusting returns for risk. It is calculated by dividing the portfolio’s average PnL (Rp ) over the
risk-free rate (Rf ) by its volatility (σp ), as shown in Equation 8. This metric adjusts returns for risk,
with a higher ratio indicating better risk-adjusted performance. Essential in comparing different
portfolios or strategies, it contextualizes performance against similar investments. Although a Sharpe
Ratio above 1 is typically considered favorable and above 2 as excellent, these benchmarks can vary
depending on the context of comparison.

Rp − Rf
Sharpe Ratio = (8)
σp

Max Drawdown of PnL [92]: Max Drawdown is a metric for assessing risk. It represents the most
significant decrease in a portfolio’s value, from its highest (Ppeak ) to its lowest point (Ptrough ) until a
new peak emerges, detailed in Equation 9. Indicative of investment strategy robustness, a smaller
Max Drawdown suggests reduced risk.

Ppeak − Ptrough
Max Drawdown = max( ) (9)
Ppeak

A.11 Baseline and Comparative Models on Single Stock Trading Task

Buy-and-Hold strategy (B&H):

A passive investment approach, where an investor purchases stocks and holds onto them for an
extended period regardless of market fluctuations, is commonly used as a baseline for comparison of
stock trading strategies.

27
A.12 Portfolio Management

LLM trading agents:

We evaluate F IN C ON against four LLM agents in the context of stock trading.
• G ENERAL - PURPOSE G ENERATIVE AGENTS – GA:The generative AI agent by Park
et al. [20], originally intended to simulate realistic human behavior and make everyday
decisions, has been adapted here for specific stock trading tasks. This agent’s architecture
includes a memory module that employs recency, relevance, and importance metrics to
extract pivotal memory events for informed decision-making. However, it does not provide
a layered memory module to effectively differentiate the time sensitivities unique to various
types of financial data. Additionally, although it features a profiling module to define agent
attributes like professional background, the model does not specify the agent’s persona. In
our experiments, we modified the original prompt template created by Park et al., which
was intended for general daily tasks, to suit financial investment tasks.
• F IN GPT: A novel open-source LLM framework specialized for converting incoming textual
and numeric information into informed financial decision-making, introduced by Yang et
al[31]. It claims superiority over the traditional buy-and-hold strategy.
• F IN M EM:F IN M EM employs a specialized profiling module and self-adaptive risk settings
for enhanced market robustness. Its memory module integrates working memory and layered
long-term memory, enabling effective data processing. This allows F IN M EM to leverage
market insights and improve trading decisions [32].
• F IN AGENT:F INAGENT developed upon F IN M EM, which leverages the use of tool-using
capabilities of LLMs to incorporate multi-modal financial data [33]. It claims an further
improved trading performance on single asset trading (stocks and cryptocurrencies).
DRL trading agents:
As the F IN M EM is practiced and examined on the basis of single stock trading and discrete trading
actions, we choose three advanced DRL algorithms fitting into the same scenarios according to the
previous and shown expressive performance in the work of Liu et al [98, 99]. The DRL training
agents only take numeric features as inputs.
• Advantage Actor-Critic (A2C): A2C ([100]) is applied to optimize trading actions in the
financial environment. It operates by simultaneously updating both the policy (actor) and
the value (critic) functions, providing a balance between exploration and exploitation.
• Proximal Policy Optimization (PPO): PPO ([101]) is employed in stock trading due to
its stability and efficiency. One salient advantage of PPO is that it maintains a balance
between exploration and exploitation by bounding the policy update, preventing drastic
policy changes.
• Deep Q-Network (DQN): DQN ([102]) is an adaptation of Q-learning, that can be used
to optimize investment strategies. Unlike traditional Q-learning that relies on a tabular
approach for storing Q-values, DQN generalizes Q-value estimation across states using deep
learning, making it more scalable for complex trading environments.

A.12 Portfolio Management

Markowitz Portfolio Selection [1]: introduced by Harry Markowitz in 1952, is a framework for
constructing portfolios that optimize expected return for a given level of risk or minimize risk for
a given level of expected return. This method uses expected returns, variances, and covariances of
asset returns to determine the optimal asset allocation, thereby balancing risk and return through
diversification.
FinRL-A2C [48]: is an RL algorithm proposed to address single stock trading and portfolio opti-
mization problems in Liu et al.. The RL models make trading decisions (i.e., portfolio weights) based
on the observation of previous market conditions and the brokerage information of the RL agents.
The implementation of this algorithm 2 is provided and is used as baselines in our study.
Equal-Weighted ETF [103]: is a portfolio giving equal allocation to all stocks, similar to a buy-and-
hold strategy in single-stock trading, can provide a benchmark on market trends.
2
[Link]

28
A.13 Ranking Metrics for Procedural Memory in F IN C ON

A.13 Ranking Metrics for Procedural Memory in F IN C ON

Upon receiving an investment inquiry, each agent in F IN C ON retrieves the top-K pivotal memory
events from its procedural memory, where K is a hyperparameter. These events are selected based on
their information retrieval score. For any given memory event E, its information retrieval score γ E is
defined by
γ E = SRelevancy
E E
+ SImportance (10)
which is adpated from Park et al [20] but with modified relevancy and importance computations, and
is scaled to [0, 1] before summing up. Upon the arrival of a trade inquiry P in processing memory
E
event E via LLM prompts, the agent computes the relevancy score SRelevancy that measures the cosine
similarity between the embedding vectors of the memory event textual content mE and the LLM
prompt query mP , which is defined as follows:
E mE · mP
SRelevancy = (11)
∥mE ∥2 × ∥mP ∥2
Note that the LLM prompt query inputs trading inquiry and trader characteristics. On the other hand,
E
the importance score SImportance is inversely correlates with the time gap between the inquiry and
the event’s memory timestamp δt = tP − tE , mirroring Ebbinghaus’s forgetting curve [104]. More
precisely, if we denote the initial score value of memory event v E and degrading ratio θ ∈ (0, 1),
then the importance score is computed via
E
SImportance = v E × θδt (12)
Note that the ratio θ measures the diminishing importance of an event over time, which is inspired by
design of [20]. But in our design, the factors of recency and importance are handled by one equation.
Different agents in F IN C ON admit different choices of {v E , θ} for memory event E.
Additionally, an access counter function facilitates memory event augmentation, so that critical events
impacting trading decisions can be augmented by F IN C ON, while trivial events are gradually faded.
This is achieved by using the LLM validation tool Guardrails AI [105] to track critical memory ID. A
E
memory ID deemed critical to investment gains receives +5 to its importance score SImportance . This
access counter implementation enables F IN C ON to capture and prioritize crucial events based on type
and retrieval frequency.

A.14 Detailed Configurations in Experiments

The training period was chosen to account for the seasonal nature of corporate financial reporting and
the duration of data retention in F IN C ON’s memory module. The selected training duration ensures
the inclusion of at least one publication cycle of either Form 10-Q, ECC, or Form 10-K. This strategy
ensures that the learned conceptualized investment guidance considers a more comprehensive scope
of factors. Additionally, the training duration allowed F IN C ON sufficient time to establish inferential
links between financial news, market indicators, and stock market trends, thereby accumulating
substantial experience. Furthermore, we set the number of top memory events retrieved for each agent
at 5. We ran F IN C ON. The reported performance outcomes are based on the setting that achieved the
highest cumulative return during the testing phase.
To maintain consistency in the comparison, the training and testing phases for the other three LLM-
based agents were aligned with those of F IN M EM. For parameters of other LLM-based agents that
are not encompassed by F IN M EM’s configuration, they were kept in accordance with their original
settings as specified in their respective source codes.
F IN C ON’s performance was benchmarked against that of the most effective comparative model, using
Cumulative Return and Sharpe Ratio as the primary evaluation metrics. The statistical significance of
F IN C ON’s superior performance was ascertained through the non-parametric Wilcoxon signed-rank
test, which is particularly apt for the non-Gaussian distributed data.

A.15 F IN C ON performance on extreme market conditions

To further illustrate the robustness of F IN C ON’s performance, we assess its effectiveness in two
distinct scenarios: (1) a single-asset trading task using TSLA and (2) a portfolio management task
involving a combination of TSLA, MSFT, and PFE. Our evaluation focuses on key financial metrics,

29
A.15 F IN C ON performance on extreme market conditions

including Cumulative Returns (CRs), Sharpe Ratios (SRs), and Maximum Drawdown (MDD). The
training period spanned from January 17, 2022, to March 31, 2022, while the testing phase covered
April 1, 2022, to October 15, 2022. This specific timeframe was chosen due to the elevated levels
of the CBOE Volatility Index (VIX), which averaged above 20, signaling greater market volatility
during these months.
As demonstrated in Table 7 and Figure 12, FINCON is the sole agent system to achieve positive
Cumulative Returns (CRs) and Sharpe Ratios (SRs) in single stock trading tasks. Regarding portfolio
management tasks, the results of all baselines (four benchmarks) are detailed in Table 8 and Figure 13.
In these comparisons, FINCON consistently attained the highest scores in the primary performance
metrics.

Models CR % ↑ SR↑ MDD %↓

B&H -56.738 -1.625 52.077
F IN C ON 22.460 0.695 45.215
GA -51.251 -1.547 48.763
F IN GPT -20.035 -0.805 32.199
F IN M EM -47.809 -1.549 49.560
F INAGENT -31.119 -1.933 33.224
A2C -73.251 -2.142 56.998
PPO -78.007 -2.284 59.003
DQN -8.452 -1.328 8.463

Table 7: Key performance comparison for single

asset trading under the high volatility condition Figure 12: CR changes over time across all the strategies
using TSLA as an example. F IN C ON leads all under the high volatility condition using TSLA as an ex-
performance metrics. ample of the single asset trading task.

Models CR % ↑ SR↑ MDD %↓

F IN C ON -8.429 -0.294 26.176
Markowitz MV -28.996 -1.805 31.831
F IN RL-A2C -15.932 -1.195 21.569
Equal-Weighted ETF -28.008 -1.731 30.070

Table 8: Key performance comparison among

all portfolio management strategies of Portfolio1
under the high volatility condition. F IN C ON leads Figure 13: Portfolio1 value changes over time for all the
all performance metrics. strategies under the high volatility condition.

30