Automate Strategy Finding with LLM in Quant investment

Zhizhuo Kou Holam Yu

[email protected] [email protected]
HKUST(GZ) HKUST(GZ)
Guangzhou, China Guangzhou, China

Jingshu Peng Lei Chen

[email protected] [email protected]
HKUST HKUST(GZ)
arXiv:2409.06289v1 [q-fin.PM] 10 Sep 2024

Hong Kong, China Guangzhou, China

Abstract ACM Reference Format:

Despite significant progress in deep learning for financial trad- Zhizhuo Kou, Holam Yu, Jingshu Peng, and Lei Chen. 2018. Automate
Strategy Finding with LLM in Quant investment. In Proceedings of Make
ing, existing models often face instability and high uncertainty,
sure to enter the correct conference title from your rights confirmation emai
hindering their practical application. Leveraging advancements in (Conference acronym ’XX). ACM, New York, NY, USA, 13 pages. https://doi.
Large Language Models (LLMs) and multi-agent architectures, we org/XXXXXXX.XXXXXXX
propose a novel framework for quantitative stock investment in
portfolio management and alpha mining. Our framework addresses
these issues by integrating LLMs to generate diversified alphas and 1 Introduction
employing a multi-agent approach to dynamically evaluate market According to the survey by Grand View Research, in 2023 the global
conditions. This paper proposes a framework where large language alternative data market was valued at 7.2bn USD and is expected
models (LLMs) mine alpha factors from multimodal financial data, to expand at the compound annual growth rate (CAGR) of 52.1%
ensuring a comprehensive understanding of market dynamics. The from 2023 to 2030 representing a complex and dynamic ecosystem
first module extracts predictive signals by integrating numerical that attracts millions of investors seeking profitable investment
data, research papers, and visual charts. The second module uses opportunities. [1] Recent advancements in quantitative trading
ensemble learning to construct a diverse pool of trading agents have significantly enhanced the ability to analyze and process vast
with varying risk preferences, enhancing strategy performance amounts of financial data, leading to the development of sophisti-
through a broader market analysis. In the third module, a dynamic cated trading strategies. One of the key areas of focus in quantitative
weight-gating mechanism selects and assigns weights to the most trading is Alpha mining—the process of discovering and refining
relevant agents based on real-time market conditions, enabling the predictive signals, or alpha factors, that forecast asset returns. [2]
creation of an adaptive and context-aware composite alpha formula. Eugene Fama proposed the Efficient Market Hypothesis (EMH) to
Extensive experiments on the Chinese stock markets demonstrate present that stock price will be the reflection of all the market avail-
that this framework significantly outperforms state-of-the-art base- able information. [3] However with the market condition change
lines across multiple financial metrics. The results underscore the the same information may make different influence on the market,
efficacy of combining LLM-generated alphas with a multi-agent some researchers also found information asymmetry. [4] Here we
architecture to achieve superior trading performance and stabil- found three major challenges to the current Alpha mining process
ity. This work highlights the potential of AI-driven approaches (1) Rigidness of traditional methods: traditional methods for iden-
in enhancing quantitative investment strategies and sets a new tifying alpha factors in finance often rely on heuristic rules and
benchmark for integrating advanced machine learning techniques financial expertise. These rule-based approaches, typically perform
in financial trading can also be applied on diverse markets. well only under specific market conditions and lack the flexibil-
ity to adapt to varying market dynamics. (2) Data diversity and
integration, recent research has explored the application of ma-
Keywords
chine learning techniques to uncover alpha factors from diverse
Alpha mining, Large Language Models, Quantitative Trading data sources, such as economic news, social media trends, and his-
torical price movements. [4–6] (3) Adapting to market variability,
Market dynamics are fluid, and strategies that perform well in one
environment may fail in another. In recent years many researchers
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed proposed to use deep learning methods to predict the market [7]
for profit or commercial advantage and that copies bear this notice and the full citation and form their strategy. [8] Efficiently mining and utilizing alpha
on the first page. Copyrights for components of this work owned by others than the factors across different market conditions remains a significant
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission challenge.
and/or a fee. Request permissions from [email protected]. To address this challenge, we propose a novel framework that
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY leverages the exploratory power of Large Language Models (LLMs)
© 2018 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-XXXX-X/18/06 for Alpha mining and strategy optimization. (1) Flexible Alpha
https://doi.org/XXXXXXX.XXXXXXX mining with LLMs, the framework starts with the collection and
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al.

categorization of potential alpha factors from academic research,

financial literature, and proprietary datasets. These seed alphas
are filtered and categorized using the potential of the LLM, which
ensures a comprehensive and diverse set of candidate seed Alphas
for further research. All these seed alpha factors are well catego-
rized as momentum factors, fundamental factors, liquidity factors,
etc., and based on traditional finance research we believe that these
factors are independent of each other. [7] Once new research work
is published, our model will be able to incremental update the seed
alphas factory. This dynamic approach allows for the discovery
of adaptable alpha factors, overcoming the rigidity of traditional
methods. (2) Multi-Agent multimodal market evaluation, the seed
Alpha factors undergo rigorous back-testing against historical mar- Figure 1: Pipeline of Quant Trading Company
ket data to assess their performance. This includes back test under
different market conditions to determine which factors can respond
to the market for excess returns under different market conditions. robust and adaptive investment strategies without human
To handle the variability in market conditions, we employ a multi- interval.
Agent approach. This method allows specialized agents to evaluate The proposed framework is versatile and can be applied across
the relevance and effectiveness of each alpha factor under multi- various asset classes, enhancing its utility and effectiveness. To
modal market scenarios. The multi-Agent architecture captures the support future research and the reproducibility of the results, we
nuanced dynamics of various market environments from different make our source code publicly available at https://anonymous.
perspectives. Based on the current market condition the multi- 4open.science/r/Automate-Strategy-Finding-with-LLM-in-Quan
Agent part can filter the out performance seed alphas for us further t-investment-ADA2/.
strategy design. (3) Dynamic strategy optimization, following the
evaluation and back-testing phases, a weight gating layer assigns 2 Problem Formulation
optimal weights to each seed alpha based on current market condi- This section introduces a comprehensive framework for generating
tions. This dynamic weighting mechanism ensures that the strategy alpha factors and strategies in quantitative trading. It begins by
adapts to changing market environments, optimizing overall perfor- outlining how alpha factors are calculated for a set of stocks over
mance. The weighted alpha strategies are then aggregated to form multiple trading periods, with the best-performing seed alphas se-
a comprehensive investment strategy that maximizes returns while lected based on back-test results and current market conditions. The
managing risk. Our approach integrates state-of-the-art techniques framework leverages Large Language Models (LLMs) and a multi-
in machine learning and finance to develop a robust system capable agent system to dynamically mine and select seed alphas, ensuring
of identifying and optimizing alpha strategies across multiple asset adaptability to market variability. The seed alphas are expressed
classes. As Figure 7 shows here the work inspired from the idea as mathematical formulas, incorporating both cross-sectional and
in real-life quant company practice, for the real quant investment time-series operators. The framework also includes a process for
company they generally have four major departments. The first one refining and updating these alphas over time, ensuring they remain
is the research department’s response to find the valuable trading relevant to evolving market conditions.
signals and record them; the second part is the trading department
need to analyze the current market condition and select the ideal
2.1 Alpha Factor & Strategy
alphas for future portfolio construction; the third part is the risk
management part, they need to evaluate and make the strategy We suppose there are 𝑛 stocks in next trading period 𝑇 . For each
suitable for the risk preference; the last part will be the portfolio trading period 𝑡 ∈ {1, 2, . . . ,𝑇 }, for every stock 𝑖 can calculate out
manager, they mix up the traders idea construct the final trading several alphas from each category value 𝛼 11, 𝛼 12, . . . , 𝛼𝑘𝑘 . The value
strategy for profit.[9] of each stock may firstly have 𝑚 initial features like the volume &
Our main contributions are three-fold: price etc. now we can use the 𝛼 value to represent the potential
value of the stocks of different category. Then from each category
we choose the domain best performance seed alpha from back test
• A new framework for searching formulaic alpha factors performance linked with the current market condition as the raw
based on LLM, to utilize the strong exploratory power of material for our final strategy. The overall Alpha Strategy can be
Í
LLM to provide Alpha factory from multimodal information represent by 𝛼 = 𝑤𝑖 𝛼𝑖 .
with incremental updates features.
• We introduce the multi-Agent approach to the financial do- 2.2 Seed Alpha Mining & Selection
main to evaluate the relationship between the market con- In quantitative trading, the static nature of traditional heuristic
dition and Alpha factors, enabling specialized evaluation of and rule-based alpha mining methods poses significant limitations
alpha strategies under different senior. (e.g. Information coefficient), particularly in the context of rapidly
• By integrating advanced techniques from both fields, our changing market conditions. [10] These methods often fail to adapt
work represents a significant step forward in developing dynamically, leading to suboptimal performance and the selected
Automate Strategy Finding with LLM in Quant investment Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

3 Methodology
This section details the methodology of a comprehensive frame-
work for constructing and optimizing alpha strategies in quantita-
tive trading. It introduces three key components: the Seed Alphas
Factory, a multi-agent decision-making process, and a weight opti-
mization approach. The Seed Alphas Factory is built using Large
Language Models (LLMs) to filter and categorize multimodal data,
creating a robust set of seed alphas. The multi-agent system evalu-
ates and selects these alphas based on diverse risk preferences and
market conditions, while a confidence scoring mechanism ensures
reliability. Finally, the selected alphas’ weights are optimized using
Figure 2: (A) An example of the seed alpha. (B) Its equivalent a Deep Neural Network (DNN) to predict future yields, forming the
expression tree. (C) Step-by-step computation of this seed foundation for a dynamic and adaptable investment strategy.
alpha on an example time series.

3.1 Overview of the Framework

Our framework is structured into three primary components: the
Alpha may loss function. Moreover, the reliance on structured fi- Seed Alphas Factory, multi-agent decision-making, and the weight
nancial data, such as historical prices and fundamental indicators, optimization approach. (Figure 3) The initial phase utilizes a Large
overlooks the rich insights that can be derived from alternative data Language Model (LLM) to filter and categorize multimodal docu-
sources like economic news and social media trends. To address ments, constructing the Seed Alphas Factory. The LLM’s ability
these challenges, we propose a novel framework that leverages to process vast and diverse datasets ensures a comprehensive and
the exploratory power of Large Language Models (LLMs) and a robust set of seed alphas categorized into independent alpha cat-
multi-agent system for seed alpha mining and selection. Our ap- egories, as established by finance alpha mining research. In the
proach integrates state-of-the-art machine learning techniques with second phase, the framework employs a multimodal multi-agent
financial principles to create a robust system capable of dynami- decision-making process. This multi-agent approach allows for the
cally adjusting to market variability and optimizing alpha strategies incorporation of varied risk perspectives, enhancing the adaptabil-
across 𝑘 categories of seed alphas. This multimodal framework links ity and robustness of the strategy across different market conditions.
current market conditions with alpha back-test results, allowing The outcome of this phase is an alpha list suited to the current mar-
LLMs to identify the most suitable seed alphas while assigning ket status and risk preferences. The final phase involves a weight
confidence scores 𝜃 to ensure reliability. Specialized agents further optimization approach using deep learning methods to assess the
customize the evaluation based on risk preferences and expertise, performance of each seed alpha and construct the overall strategy.
ensuring a comprehensive and adaptive solution for maximizing The framework is designed to be dynamic, allowing for contin-
returns and managing risks in diverse market environments. uous improvement and adaptation. As new research emerges and
market conditions evolve, the Seed Alphas Factory can be incremen-
2.3 Seed Alpha Formulaic tally updated by providing additional information to the LLM. This
Our framework sets a rule for the output of the LLM-generated seed continuous update mechanism ensures that the Seed Alphas Factory
alphas, ensuring they are expressed as mathematical expressions remains relevant and robust over time. Similarly, the multi-agent
that combine various operators with raw financial features. These decision-making process can adapt to changing market conditions
operators include elementary functions such as addition ("+") and by selecting new sets of seed alphas, thereby maintaining the ef-
logarithms ("log"), which are used on one-day data and are referred ficacy of the strategy. This methodology is versatile and can be
to as cross-section operators. Additionally, there are time-series applied to any structured market globally. By integrating advanced
operators that require data spanning multiple days. An example of LLM capabilities and multi-agent systems, this proposed methodol-
a time-series operator is " (CLOSE - DELAY(SMA(CLOSE, 14),7)", ogy aims to replicate and enhance the sophisticated strategies used
which measures the momentum by comparing the current closing in professional investment research.
price to the 14-period Simple Moving Average (SMA) value from
7 periods ago, indicating whether the price is trending upward or
downward relative to that past average, as illustrated in Figure 2 3.2 LLM Filter & Category
(B). The first stage of our methodology involves the LLM Filter & Cat-
The comprehensive list of seed alpha formulas used in our exper- egory process for alpha-related research. This stage utilizes the
iments can be found in Appendix 2. It is important to note that with summarization and categorization capabilities of Large Language
ongoing research and new data, some alpha formulas may undergo Models (LLMs) to expedite the understanding of recent alpha min-
numerical adjustments, and others might be replaced with new ing research and to establish our Seed Alpha Factory for further
seed alpha formulas. This reflects our commitment to incremental development.
updates and the continuous refinement of our seed alpha factory For this study, we employed custom versions of ChatGPT, named
to adapt to the latest market conditions and research insights. “Alpha Grail,” to perform the LLM filtering and categorization tasks.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al.

Figure 3: Overview of the strategy generate process (CS stands for confidence score; RP stands for risk preference).

Table 1: Multimodal data types are derived from the LLM’s summarization and categorization pro-
cess. The detailed generated Seed alphas factory can be seen in the
Data Type Description Examples Appendix 2. This structured output forms the foundation for the
Textual Data Financial reports, aca- Trading forums’ sentiment analysis and
demic papers, news ar- stock predictions, company disclosures, subsequent multi-agent decision-making and weight optimization
ticles, and other textual financial statements, Sina Finance stages in our framework.
documents.
Numerical Data Historical stock market Returns, log returns, annualized returns,
data, financial metrics, volatility 3.3 Multimodal & multi-Agent Evaluate Alpha
and performance indi-
cators. set
Visual Data Charts, graphs, and Kline charts, trading charts The second stage of our methodology involves a comprehensive
other visual representa-
tions of financial data. evaluation and selection of alpha factors through a multimodal
Audio Data Financial news broad- Financial morning news radio, stock re- and multi-agent system. This approach ensures a diversified and
casts. view radio, market discussion radio thorough assessment of the seed alphas by integrating diverse data
Video Data Financial news chan- CCTV Securities Information Channel,
nels. CCTV News Broadcast (news affecting sources and leveraging multiple analytical perspectives.
China’s stock market) In this stage, we incorporate multimodal data encompassing
textual, numerical, visual, and multimedia inputs to enrich the eval-
uation process. By integrating these varied data types, we provide a
comprehensive view of market conditions and the performance of
The primary function of Alpha Grail is to assist quantitative re- alpha factors. Table 2 outlines the five types of data utilized in our
searchers in building the Seed Alpha Factory by processing alpha- evaluation framework. This holistic approach allows for a nuanced
related research documents. The specific instruction given to Alpha understanding of the market dynamics and enhances the robustness
Grail was: "Summarize the document information to help quanti- of the evaluation process.
tative researchers build the Seed Alpha Factory according to tra- The multi-agent system comprises various agents, each designed
ditional financial categories, ensuring that each category of seed with distinct risk preferences and investment strategies. These
alphas is independent." [11] Initially, we provided 11 documents agents analyze the multimodal data and evaluate the seed alphas
encompassing diverse backgrounds in both financial and computer based on their respective criteria. Each agent conducts a risk prefer-
science alpha mining research. Details of these documents are avail- ence analysis, assessing the seed alphas according to its predefined
able in Appendix 1. Through this process, Alpha Grail generated risk parameters. This ensures that the evaluation covers a wide spec-
nine categories comprising 100 seed alphas. trum of risk perspectives. Agents then select alpha factors that align
The LLM Filter & Category stage also integrates multimodal pro- with their risk preferences and market analyses, ensuring relevance
cessing capabilities. By incorporating multimodal data sources such and robustness under different market scenarios. To enhance the
as text, images, tables, and figures from the research documents, reliability of the selected alpha factors, a confidence scoring mech-
the LLM can provide a more comprehensive analysis and catego- anism is implemented. Each agent assigns a confidence score to the
rization of the seed alphas. This multimodal approach enhances the alpha factors based on their analysis and historical performance,
LLM’s ability to capture intricate details and relationships within reflecting the agent’s assessment of the alpha factor’s predictive
the research, leading to a more robust and diverse Seed Alpha power and robustness. Selected alpha factors undergo rigorous
Factory. backtesting using historical market data to evaluate their perfor-
The output of this stage is a structured set of seed alphas cate- mance under various market conditions. This step is crucial for
gorized into distinct financial domains such as Momentum, Mean identifying the most promising alpha factors for the final strategy.
Reversion, Volatility, and Fundamental analysis. Each category in- Key evaluation metrics include performance indicators such as the
cludes specific alpha names and corresponding alpha codes, which Information Coefficient (IC) and Sharpe Ratio etc.[12] Additionally,
Automate Strategy Finding with LLM in Quant investment Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

Algorithm 1 Category-Based Alpha Selection generalizability and prevent overfitting, a separate validation set
Input: categories C = {𝐶 1, ..., 𝐶𝑚 }, each 𝐶𝑖 containing a set of is utilized. This methodology establishes a robust framework for
Alphas, confidence threshold 𝑋 selected Alphas A = {} predicting future yields based on historical alpha values, thus form-
1: for each category 𝐶𝑖 ∈ C do ing an effective investment strategy. The DNN model processes the
2: A𝑖 = SelectBestAlphas(𝐶𝑖 ); input data through the hidden layer, transforming it with learned
3: for each 𝛼 ∈ A𝑖 do weights and biases. The final output is generated by applying an-
4: if ConfidenceScore(𝛼) > 𝑋 then other set of weights, biases, and an activation function in the output
5: A = A ∪ {𝛼 }; layer. Our approach demonstrates the efficacy of leveraging deep
6: end if learning techniques to optimize alpha weights, thereby enhancing
7: end for the predictive performance of investment strategies.
8: end for
9: return A 4 Experiment
Our research aims to develop a comprehensive LLM-driven alpha
mining framework that operates without human intervention. This
seed alpha factors from each category must meet specific confidence framework is uniquely capable of processing multimodal infor-
score thresholds to be considered for the final selection, ensuring mation and adapting to varying market conditions. A critical in-
that only the most reliable and robust alpha factors are chosen. novation in our system is the introduction of a confidence score,
To automate the selection process, we designed the Category- designed to mitigate potential hallucinations during the LLM’s seed
Based Alpha Selection algorithm. This algorithm identifies and alpha selection process. By integrating the results of seed alpha back
selects alphas from different categories based on their confidence tests with multimodal market conditions, our framework enables a
scores. The input consists of multiple categories, each containing more thorough understanding of market dynamics, and the senti-
a set of alphas and a confidence threshold. The algorithm begins ments associated with each category of seed alphas. The primary
with an empty set of selected alphas and iterates through each objectives of our research are as follows: Evaluate Framework Effec-
category, using the SelectBestAlphas function to determine the most tiveness: Assess whether our LLM-driven framework can identify
promising candidates. Each alpha’s confidence score is evaluated, alphas that are better suited to prevailing market conditions com-
and those exceeding the threshold are included in the final set. This pared to traditional methods. Multimodal knowledge extraction:
process repeats for all categories, ensuring a rigorous selection of Analyze the framework’s capability to process and integrate diverse
alphas that meet the confidence criteria from all categories. The data inputs, enhancing its understanding of market dynamics and
final output is a set of selected alphas that have successfully passed sentiment analysis. Confidence Scoring Mechanism: Investigate
the confidence evaluation. the effectiveness of the confidence score in reducing the likelihood
The multimodal and multi-agent evaluation stage leverages ad- of hallucinations during the alpha selection process. Overall Strat-
vanced data integration and diverse analytical perspectives to en- egy Performance: Determine whether our comprehensive strategy,
sure a comprehensive assessment of alpha factors. The incorpora- leveraging the LLM-driven framework, can outperform the market.
tion of confidence scoring adds an additional layer of reliability, To validate the effectiveness of our framework, we have con-
making the selection process more robust. This stage is crucial ducted a series of experiments addressing the following research
for identifying the most promising alpha factors and constructing questions (RQs):
a resilient investment strategy that can adapt to varying market • RQ1: Does the framework can achieve multimodal knowl-
conditions. The integration of multimodal data sources and the edge extraction and capture the new alpha with the market
utilization of a multi-agent system provide a nuanced and in-depth condition difference?
evaluation, laying a solid foundation for the subsequent stages of • RQ2: Does our LLM-driven alpha mining framework outper-
the framework. form current alpha factory in terms of trading effectiveness?
• RQ3: Can our overall strategy, incorporating the LLM-driven
3.4 Optimal weight alpha strategy framework and confidence scoring, consistently beat the
To optimize the weights of selected seed alphas, we employ a Deep market?
Neural Network (DNN) to fit alpha values to future yields. The
network architecture comprises three layers: an input layer, a hid- 4.1 Datasets
den layer, and an output layer. The input layer takes daily alpha Our study focuses on financial data from the Chinese A-shares
calculations derived from historical closing prices as input features. market, specifically targeting the SSE 50 Index. The dataset encom-
The hidden layer, equipped with ten nodes, utilizes ReLU activa- passes six primary features as original inputs for our Alpha factors:
tion functions to introduce non-linearity and enhance the model’s open, close, high, low, volume, and volume-weighted average price
learning capability. [13] The output layer, consisting of a single (VWAP). To ensure rigorous evaluation and robust model perfor-
node, predicts future yields. During the training phase, the network mance, the dataset is divided into distinct time periods: Training
employs backpropagation and gradient descent to minimize the loss Set: January 1, 2021, to December 31, 2022 and Test Set: January 1,
function, which quantifies the discrepancy between predicted and 2023, to December 31, 2023. Our experiments integrate financial
actual yields. This iterative process adjusts the network’s weights reports and factor performances of the 50 constituent companies
and biases to improve prediction accuracy. To ensure the model’s of the SSE 50 Index, providing a comprehensive view of the market.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al.

Table 2: Summary of the experiment dataset

Aspect Details
Primary Features Open, Close, High, Low, Volume, VWAP
Alpha Factors Custom factors based on price, volume,
financial ratios, moving averages, senti-
ment analysis
Financial Reports Quarterly and Annual reports from SSE
50 constituent companies
Time Periods - Training: Jan 1, 2021 - Dec 31, 2022
- Test: Jan 1, 2023 - Dec 31, 2023
Market Coverage SSE 50 Index, covering major sectors in
the Chinese A-shares market
Evaluation Criteria Causal relationships, Alpha factor per-
formance, model robustness

Figure 4: Multimodal LLM Analysis architecture

Experiments demonstrate the framework’s effectiveness in multi-

modal knowledge extraction and dynamic alpha selection under dif-
ferent market conditions. (Figure 5) In Case 1, using SSE50 company
announcements, financial statements, K-line charts, and trading
charts from December 31, 2021, to September 30, 2022, the selected
[14]The evaluation considers various datasets, including financial
alphas included momentum and volume-based indicators such as
reports from the specified periods and performance metrics of dif-
price momentum, RSI, MACD, moving averages, Bollinger Bands,
ferent Alpha factors. This extensive dataset allows us to identify
volume, market cap, and EPS adjustments. In Case 2, focusing on
causal relationships and patterns between financial features and
incremental updated SSE50 related news, stock commentary, and
Alpha factor performances. The diversified portfolio in our eval-
the China Macroeconomic Index, the selected alphas emphasized
uation comprises prominent companies within the SSE 50 Index,
volatility and economic factors, including ATR, Bollinger Bands
selected for their extensive market coverage and representation
ratios, volume-based indicators with delays, gross profit and operat-
across different sectors of the Chinese A-shares market. This selec-
ing income ratios, and high/low price comparisons. The experiment
tion underscores the versatility and consistency of our model across
underscores the framework’s adaptability, capturing new opportu-
various financial environments, showcasing its broad applicability
nities by integrating diverse data sources and employing advanced
and reliability.
analytical models. This dynamic approach ensures the selected
alphas are relevant and responsive to current market conditions,
enhancing the robustness and effectiveness of investment strategies.

4.2 Research Questions 4.2.2 RQ2: Does our LLM-driven alpha mining framework outper-
4.2.1 RQ1: Does the framework can achieve multimodal knowledge form current alpha factory in terms of trading effectiveness? When
extraction and capture the new alpha with the market condition dif- evaluating the performance of selected seed alpha signals, the pri-
ference? We present a prompt architecture (Figure 4) that incorpo- mary metric is the Information Coefficient (IC). [25] These metrics
rates multimodal market information into a Large Language Model provide insights into the predictive power of alpha values for future
(LLM) to achieve multimodal knowledge extraction and select opti- returns. The Information Coefficient (IC) measures the correlation
mal seed alphas under varying market conditions. By integrating between predicted alpha values and actual future returns. It is quan-
textual data from financial news sentiment, numerical data from tified using Pearson’s correlation coefficient, which assesses the
company financial statements, and visual data from trading charts, linear relationship between two variables. A higher IC indicates
our framework provides comprehensive stock analysis. This data a stronger predictive relationship, meaning the alpha values are
fusion ensures a holistic market view, enhancing decision-making. more effective at forecasting returns.
Contextual analysis adjusts parameters based on market trends and
sector performance, ensuring relevant and accurate alpha selection. 𝐼𝐶 = 𝜎 (𝑢, 𝑣)
Automate Strategy Finding with LLM in Quant investment Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

Sel
ect
edAl
phadependsondi
ffer
entcont
ext

Case1 Case2
the weight combination will drop to -0.055; once we remove the
Cont
ext
:Text+Tabl
e+l
mage. Cont
ext
:Text+Tabl
e+l
mage. alpha #11, the weight combination will drop to merely 0.0491. This
Text
:SSE50Li
stedCompanyAnnouncementi
n2021.
12.
31-2022.
09.
30 Text
:SSE50r
elat
edNewswi
thsent
imentDat
aset
in2021.
12.
31-2022.
09.
30
suggests that the seed alpha set selected by the LLM synergizes
Tabl
e:SSE50Li
stedCompanyFi
nicalSt
atement
in2021.
12.
31-2022.
09.
30 Tabl
e:St
ockcomment
aryi
nChi
nesef
orums2021.
12.
31-2022.
09.
30

l
mage:Kl
i
nechar
t& Tr
adi
ngchar
tof$SE50i
n2021.
12.
31-2022.
09.
30 l
mage:Chi
naMacr
oeconomi
cIndexi
n2021.
12.
31-2022.
09.
30
effectively, providing robust predictive power. [10]

Sel
ect
edAl
pha Sel
ect
edAl
pha Table 3: An example combination of 12 alphas.
(
CLOSE-
DELAY(
CLOSE.
14)
)
(
RSI.
DELAY(
RSI
,14)
) (
ATR-DELAY(
ATR,14)
)
(
CLOSE-
DELAY(
SMA(
CLOSE,
14)
,7)
)
(
UPPERBAND-LOWERBAND)
/SMA(
CLOSE,
20)
(
MA(
CLOSE.
20)
-CLOSE)
(
VOLUME-DELAY(
VOLUME,14)
)/DELAY(
VOLUME,14)
(
SMA(
(
CLOSE,
MAX(
HI
20)
-
GH,20)
-
CLOSE)
CLOSE) (
GROSSPROFI
T/REVENUE)
# Alpha Weight IC(SSE50)
(
100-
RSI
)
(
OPERATI
NG I
NCOME/
REVENUE)
(
BOLLUP.
BOLLDOWN)
/SMA(
CLOSE,20)
STD(
CLOSE,
10)
/STD(
CLOSE,50) (
MAX(
HIGH,
14)
-CLOSE)
/(MAX(
HIGH,
14)
-MI
N(LOW,
14)
))
*-100 1 (CLOSE - DELAY(CLOSE, 14)) -0.1459 0.0209
VOLUME/
MARKETCAP (
MAX(
HIGH,
20)
-CLOSE)

(
VOLUME*
CLOSE
EPS/DELAY(
EPS,
1)-
1)
2 (RSI - DELAY(RSI, 14)) -1.0265 -0.0225
3 (CLOSE - DELAY(SMA(CLOSE, 14), 7)) -0.1978 0.0193
4 (MA(CLOSE, 20) - CLOSE) 0.0556 -0.0186
5 (SMA(CLOSE, 20) - CLOSE) -0.945 -0.0186
Figure 5: Sample Experiment on Different Market Status 6 (MAX(HIGH, 20) - CLOSE) -0.4053 -0.0185
Input and Alpha Selection 7 (100-RSI) -0.3199 0.0194
8 (BOLL_UP - BOLL_DOWN) / SMA(CLOSE, 20) 3.6186 0.0278
9 STD(CLOSE, 10) / STD(CLOSE, 50) -0.183 0.0236
10 VOLUME / MARKET_CAP -3.2145 -0.0194
11 VOLUME * CLOSE -0.0058 0.0187
12 (EPS / DELAY(EPS, 1) - 1) -1.8351 -0.0215
Weighted Combination -0.0587

To address the question of whether our overall strategy, incor-

porating the LLM-driven framework and confidence scoring, can
consistently beat the market, we conducted a backtest using a
straightforward investment approach during the period from Janu-
ary 1, 2023, to December 31, 2023, on the SSE50 dataset. Our strategy
involves a daily portfolio reconstruction based on a top- 𝑘/ drop
−𝑛 selection process. Each trading day, we rank the stocks by their
alpha values and select the top 𝑘 stocks for investment. We aim to
distribute investments evenly among these selected stocks, but we
limit our transactions to a maximum of 𝑛 stocks per day to reduce
trading costs. For this experiment, we set 𝑘 to 13 and 𝑛 to 5 . The
net worth progression of the respective strategies over the testing
Figure 6: Average IC Comparation after LLM Selection
period is shown in Figure 7. Despite not explicitly optimizing for
absolute returns, our framework demonstrates impressive perfor-
where 𝜎 (𝑢, 𝑣) is Pearson’s correlation coefficient between pre- mance in the backtest, achieving the highest profit compared to
dicted alphas 𝑢 and actual future returns 𝑣. We evaluated four most other methods. Our cumulative return for 2023 backtest comes to
common alpha categories: Momentum, Mean Reversion, Volatility, 53.17% positive and meanwhile the index performance is -11.73%,
Fundamental and Growth. The figure 6 below shows the average EFund performance is -9.17% and the Boshi Fund performance is
IC values for each category. The results demonstrate that our LLM- -8.81%. We also do the compression with others alpha mining work
driven framework consistently achieves higher average IC values shows our framework are able to gain most profit in the market.
across all categories, particularly in Volatility and Fundamental, in- This approach shows the potential of our LLM-driven frame-
dicating superior trading effectiveness compared to the traditional work combined with confidence scoring to consistently outperform
alpha factory. the market by adapting dynamically to changing conditions and
minimizing costs through a disciplined investment strategy.
4.2.3 RQ3: Can our overall strategy, incorporating the LLM-driven
framework and confidence scoring, consistently beat the market? 5 Related Work
Table 3 presents an example combination of 12 alphas generated by
our framework, evaluated on the SSE50 constituent stock set. The 5.1 Formulaic alphas
table details the seed alphas selected by the LLM from each category, Formulaic alphas in quantitative investment refer to systematic,
along with their respective weights and Information Coefficient rule-based strategies that generate excess returns by exploiting
(IC) values. The weight combination IC value is quite high as - specific market patterns, anomalies, or inefficiencies. [15] These
0.0587. Although some of the seed alphas exhibit relatively low strategies use predefined mathematical models or algorithms and
IC values individually, their removal results in a significant drop are often implemented through automated trading systems. The
in the retrained combination weight, indicating their critical role common method for creating formulaic alphas is genetic program-
in the overall performance. For example, if we remove alpha #6 ming involves structure and numerical mutations to generate new
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al.

Figure 7: Backtest Result on SSE50 Comparison Between Index and Our performance

alphas. [16] The gplearn library is enhanced with time-series opera- 5.3 Multimodal & Multi-Agent LLM
tors specific forming an alpha-mining framework that incorporates The integration of multimodal large language models (LLMs) in
mutual information as the fitness measure. [17] AlphaEvolve evolv- the investment sector has been extensively researched, highlight-
ing new alphas from existing ones, using algorithmic graphs for ing their potential to transform financial technology. Li et al. [35]
more complex and accurate predictions, though this can reduce demonstrate how multimodal LLMs can process diverse data types
interpretability and increase the risk of overfitting. [5] Another such as text, images, and numerical data, offering comprehensive
approach is Machine learning-based alphas are predictive models insights and predictive capabilities. Ouyang et al. [36] emphasize
designed to generate trading signals or investment strategies using the robustness of these models in developing investment strategies
advanced machine learning techniques. Key algorithms include neu- that mitigate market volatility risks. By analyzing textual data from
ral networks, such as Long Short-Term Memory (LSTM) [18] and news articles and social media, these models can gauge investor
Transformer models [19], which excel at capturing complex tempo- emotions and predict market movements with higher accuracy. [37].
ral dependencies in time series data. Decision trees, exemplified by This is particularly beneficial in high-frequency trading, where an-
models like XGBoost [20] and LightGBM [21], are valued for their ticipating market shifts can provide a competitive edge [38]. In
interpretability and ability to handle high-dimensional data. Ex- portfolio management, multimodal LLMs offer personalized invest-
plainable AI (XAI) [22] methods and tools provide deeper insights ment advice tailored to individual risk profiles and financial goals.
into how these models make decisions, fostering greater under- [39]. These models analyze a client’s financial history and current
standing and trust among investors. Specialized network structures market conditions to recommend optimal asset allocations and
for stock forecasting, such as SFM [23], which employs a DFT-like strategies, enhancing client satisfaction and portfolio performance.
mechanism, have also been developed. More recent research focuses Additionally, they automate routine tasks such as data entry and
on integrating non-standard data with time series data. For example, report generation, allowing investment professionals to focus on
REST [24] combines multi-granular time series data with historical strategic activities [40].
event data to model the overall market. HIST [25] uses concept The application of multi-agent LLMs further enhances market
graphs alongside regular time series data to capture commonalities analysis and portfolio management. Zhang et al. [38] highlight how
in future trends across different stock groups. these models leverage vast datasets and advanced natural language
processing to interpret financial reports, news, and social media sen-
timent, providing nuanced insights. Multi-agent systems simulate
various market scenarios and stress tests, enabling fund managers
to devise robust strategies resilient to market volatility [41]. The
5.2 Financial Domain LLM collaborative operation of multiple agents ensures diverse strate-
gies can be tested and optimized in parallel, promoting a balanced
The development of general-domain large language models (LLMs) approach to risk and return [42]. However, the implementation
has spurred interest in Finance Large Language Models (Fin-LLMs). of these advanced models raises ethical and governance consid-
While general LLMs are well-studied, Fin-LLMs are still emerging erations. Ensuring transparency, accountability, and adherence to
[26–29]. Open-source LLMs, such as LLaMA[30], BLOOM [31], and ethical guidelines and regulatory standards is crucial to maintaining
Flan-T5[32], provide flexibility and privacy on self-hosted servers investor trust and market integrity. Managing potential biases in
but may underperform compared to proprietary models. Fine-tuned data and model training is also essential to prevent systemic risks
financial LLMs outperform base models (e.g., LLaMA) and other and ensure fair outcomes [43, 44].
open-source LLMs (e.g., BLOOM) in finance tasks, showing en-
hanced domain-specific comprehension [33]. However, in gener-
ative tasks, their performance is comparable or inferior to gen- 5.4 Discussion
eral LLMs like ChatGPT and GPT-4, indicating the need for better The term “formulaic alpha” is often associated with investing, the
domain-specific datasets [34]. concept of simple and interpretable formulaic predictors, which can
Automate Strategy Finding with LLM in Quant investment Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

be combined into more expressive models and many more applica- Computational Linguistics (ACL), pages 1970–1979, 2018.
tion scenarios. The concept of the “formulaic alpha” can solve many [8] Chi Chen, Li Zhao, Jiang Bian, Chunxiao Xing, and Tie-Yan Liu. Investment
behaviors can tell what inside: Exploring stock intrinsic properties for stock
forecasting problems in other subjects such as energy consumption trend prediction. In Proceedings of the 25th ACM SIGKDD International Conference
prediction, anomaly detection, and biomedical applications. [45– on Knowledge Discovery & Data Mining, pages 2376–2384, 2019.
[9] Frank J Fabozzi, Sergio M Focardi, and Petter N Kolm. Quantitative equity
47]The work also has potential future work, The current framework investing: Techniques and strategies. John Wiley & Sons, 2010.
utilizes a multi-Agent architecture to evaluate market conditions [10] Feng Zhang, Ruite Guo, and Honggao Cao. Information coefficient as a perfor-
and select alpha factors. Future work can explore transitioning mance measure of stock selection models, 2020.
[11] OpenAI. Gpt-4 technical report, 2023.
this architecture to a Mixture of Experts (MoE) structure, which [12] William F Sharpe. The sharpe ratio. Journal of portfolio management, 21(1):49–58,
can improve learning efficiency and model performance. MoE can 1994.
dynamically allocate computational resources to different "experts" [13] Yinpeng Chen and et al. Dynamic relu. In European Conference on Computer
Vision. Springer International Publishing, 2020.
based on the specific characteristics of the input data, thereby en- [14] Weiping Li and Feng Mei. Asset returns in deep learning methods: An empirical
hancing the model’s ability to learn and distinguish category seed analysis on sse 50 and csi 300. Research in International Business and Finance,
54:101291, 2020.
alphas features more effectively. [48]Another promising avenue for [15] Zura Kakushadze. 101 formulaic alphas. Wilmott, 2016(84):72–81, 2016.
future research is the incorporation of financial knowledge graph [16] L. W. Cong, K. Tang, J. Wang, and Y. Zhang. Alphaportfolio: Direct construction
techniques into the strategy development process. Financial mar- through deep reinforcement learning and interpretable ai, 2021.
[17] Xiaoming Lin, Ye Chen, Ziyu Li, and Kang He. Stock alpha mining based on
kets are inherently complex and interconnected, with numerous genetic algorithm. Technical report, Huatai Securities Research Center, 2019.
dependencies between stocks, sectors, and economic indicators. [18] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural
[49] By introducing the knowledge graph information may help in Computation, 9(8):1735–1780, 1997.
[19] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
our stock selection phase and achieve better performance in the Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need.
future. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob
Fergus, S. V. N. Vishwanathan, and Roman Garnett, editors, Advances in Neu-
ral Information Processing Systems 30: Annual Conference on Neural Information
6 Conclusion Processing Systems 2017, pages 5998–6008, Long Beach, CA, USA, 2017. Neural
Information Processing Systems Foundation.
In this paper, we proposed a novel framework for quantitative stock [20] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system.
investment and portfolio management by leveraging advancements In Balaji Krishnapuram, Mohak Shah, Alexander J. Smola, Charu C. Aggarwal,
Dou Shen, and Rajeev Rastogi, editors, Proceedings of the 22nd ACM SIGKDD
in Large Language Models (LLMs) and multi-Agent architectures. International Conference on Knowledge Discovery and Data Mining, pages 785–794,
Our framework addresses the challenges of instability and high San Francisco, CA, USA, 2016. ACM.
uncertainty in existing models by integrating LLMs to generate [21] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei
Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree.
diversified alphas and employing a multi-Agent approach to dy- In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob
namically evaluate market conditions. The LLMs mine and generate Fergus, S. V. N. Vishwanathan, and Roman Garnett, editors, Advances in Neural
Information Processing Systems 30: Annual Conference on Neural Information
potential alpha factors from a multimodal financial information Processing Systems 2017, pages 3146–3154, Long Beach, CA, USA, 2017. Neural
landscape, providing a comprehensive understanding of market Information Processing Systems Foundation.
dynamics and a rich set of predictive signals. Subsequently, we con- [22] Alejandro Barredo Arrieta and et al. Explainable artificial intelligence (xai):
Concepts, taxonomies, opportunities and challenges toward responsible ai. Infor-
struct a diverse pool of trading agents with various risk preferences mation Fusion, 58:82–115, 2020.
through an multi-Agent approach, enabling personalized market [23] Liheng Zhang, Charu C. Aggarwal, and Guo-Jun Qi. Stock price prediction via
insights and trading strategies. The final stage involves a sophis- discovering multi-frequency trading patterns. In Proceedings of the 23rd ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, pages
ticated weight deep learning mechanism that dynamically selects 2141–2149, Halifax, NS, Canada, 2017. ACM.
and assigns weights to the most relevant agents based on real-time [24] Wentao Xu, Weiqing Liu, Chang Xu, Jiang Bian, Jian Yin, and Tie-Yan Liu. Rest:
Relational event-driven stock trend forecasting. In Jure Leskovec, Marko Grobel-
market conditions. Through extensive experiments on Chinese nik, Marc Najork, Jie Tang, and Leila Zia, editors, WWW ’21: The Web Conference
stock markets, we demonstrate that our framework significantly 2021, pages 1–10, Virtual Event / Ljubljana, Slovenia, 2021. ACM / IW3C2.
outperforms tranditional alpha generating SOTA works, human [25] Wentao Xu, Weiqing Liu, Lewen Wang, Yingce Xia, Jiang Bian, Jian Yin, and
Tie-Yan Liu. Hist: A graph-based framework for stock trend forecasting via
traders performance & index across multiple financial metrics. mining concept-oriented shared information, 2021.
[26] Robert Novy-Marx. Fundamentally, momentum is fundamental momentum.
Technical Report w20984, NBER, 2015.
References [27] Y. Yang, Y. Tang, and K. Y. Tam. Investlm: A large language model for investment
[1] Vladi Nikolov. 3 quantitative strategies based on alternative data. 2024. using financial domain instruction tuning, 2023.
[2] I. Tulchinsky. Finding Alphas: A Quantitative Approach to Building Trading [28] S. Zhang, L. Dong, X. Li, S. Zhang, X. Sun, S. Wang, and et al. Instruction tuning
Strategies. 2015. for large language models: A survey, 2023.
[3] Eugene F Fama. Efficient capital markets: A review of theory and empirical work. [29] W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y. Hou, Y. Min, B. Zhang, J. Zhang,
The Journal of Finance, 25(2):383–417, 1970. Z. Dong, and et al. A survey of large language models, 2023.
[4] Franklin Allen and Risto Karjalainen. Using genetic algorithms to find technical [30] H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière,
trading rules. Journal of Financial Economics, 51(2):245–271, 1999. N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample.
[5] Can Cui, Wei Wang, Meihui Zhang, Gang Chen, Zhaojing Luo, and Beng Chin Llama: Open and efficient foundation language models, 2023.
Ooi. Alphaevolve: A learning framework to discover novel alphas in quantitative [31] B. W. and et al. Bloom: A 176b-parameter open-access multilingual language
investment. In Proceedings of the 2021 International Conference on Management model, 2023.
of Data (SIGMOD ’21), pages 2208–2216, Virtual Event, China, 2021. ACM. [32] H. W. and et al. Scaling instruction-finetuned language models, 2022.
[6] Shuo Yu, Hongyan Xue, Xiang Ao, Feiyang Pan, Jia He, Dandan Tu, and Qing He. [33] P. Lewis and et al. Retrieval-augmented generation for knowledge-intensive nlp
Generating synergistic formulaic alpha collections via reinforcement learning. tasks. In NeurIPS, volume 33, pages 9459–9474, 2020.
In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and [34] K. J. Koa, Y. Ma, R. Ng, and T.-S. Chua. Learning to generate explainable stock
Data Mining (KDD ’23), pages 5476–5486, New York, NY, USA, 2023. Association predictions using self-reflective large language models. In Proceedings of the
for Computing Machinery. ACM on Web Conference 2024, pages 4304–4315, 2024.
[7] Yumo Xu and Shay B Cohen. Stock movement prediction from tweets and [35] Li L, Chang T Y, and Wang H. Multimodal gen-ai for fundamental investment
historical prices. In Proceedings of the 56th Annual Meeting of the Association for research, 2023.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al.

Document name Author

Financial Time Series Prediction Using Mixture of Experts M. Serdar Yumlu, Fikret S. Gurgen, and Nesrin Okay
Model-Free Implied Dependence and the Cross-Section of Returns Koen Inghelbrecht, Gertjan Verdickt, Daniël Linders, and Yong Xie
Dynamic Graph-based Deep Reinforcement Learning with Long Pengqian Yu, Joon Sern Lee, Ilya Kulyatin, Zekun Shi, Sakyasingha
and Short-term Relation Modeling for Portfolio Optimization Dasgupta
Can Large Language Models Beat Wall Street? Unveiling the Poten- Georgios Fatouros, Konstantinos Metaxas, John Soldatos, Dimos-
tial of AI in Stock Selection thenis Kyriazis
Mastering Stock Markets with Efficient Mixture of Diversified Trad- Shuo Sun, Xinrun Wang, Wanqi Xue, Xiaoxuan Lou, and Bo An
ing Experts
101 Formulaic Alphas Zura Kakushadze
A Multimodal Foundation Agent for Financial Trading: Tool- Wentao Zhang, Lingxuan Zhao, Haochong Xia, Shuo Sun, Jiaze Sun,
Augmented, Diversified, and Generalist Molei Qin, Xinyi Li, Yuqing Zhao, Yilei Zhao, Xinyu Cai, Longtao
Zheng, Xinrun Wang, Bo An
Causal Factor Investing: Can Factor Investing Become Scientific? Marcos Lopez de Prado
Alpha-GPT: Human-AI Interactive Alpha Mining for Quantitative Saizhuo Wang, Hang Yuan, Leon Zhou, Lionel M. Ni, Heung-Yeung
Investment Shum, Jian Guo
AlphaPortfolio: Direct Construction Through Deep Reinforcement Lin William Cong, Ke Tang, Jingyuan Wang, Yang Zhang
Learning and Interpretable AI
FinGPT: Open-Source Financial Large Language Models Hongyang (Bruce) Yang, Xiao-Yang Liu, Christina Dan Wang

[36] Ouyang K, Liu Y, Li S, and et al. Modal-adaptive knowledge-enhanced graph- B Generate Seed Alpha factory
based financial prediction from monetary policy conference calls with llm, 2024.
[37] H. Zhao, Z. Liu, Z. Wu, and et al. Revolutionizing finance with llms: An overview C Sample prompt for Seed Alpha selection
of applications and insights, 2024.
[38] W. Zhang, L. Zhao, H. Xia, and et al. A multimodal foundation agent for financial 1. The training set includes:
trading: Tool-augmented, diversified, and generalist. Technical report, 2024. • Financial Reports: 4 quarters (from December 31, 2021, to
[39] J. Li, X. Wang, S. Zhu, and et al. Cumo: Scaling multimodal llm with co-upcycled
mixture-of-experts, 2024. September 30, 2022) for 50 companies listed in the SSE 50.
[40] A. Luz. Enhancing the interpretability and explainability of ai-driven risk models • Factor Analysis Data: 37 factors (from September 30, 2022,
using llm capabilities. Technical report, EasyChair, 2024. to December 30, 2022) divided into groups: Momentum,
[41] Y. Talebirad and A. Nadiri. Multi-agent collaboration: Harnessing the power of
intelligent llm agents, 2023. Mean Reversion, Volatility, Fundamental, Quality, Growth,
[42] Q. Wang, Z. Wang, Y. Su, and et al. Rethinking the bounds of llm reasoning: Are Technical, Macro Economics. The metric used is the Infor-
multi-agent discussions the key?, 2024.
[43] Yangyang Yu and et al. Finmem: A performance-enhanced llm trading agent with
mation Coefficient (IC).
layered memory and character design. In Proceedings of the AAAI Symposium 2. Objective: Learn the relationship between the performance
Series, volume 3, 2024.
[44] T. Nathan Mundhenk, Mikel Landajuela, Ruben Glatt, Daniel M. Faissol, and Bren-
of financial reports for the first four quarters from 2022 to 2023 and
den K. Petersen. Symbolic regression via neural-guided genetic programming the factor analysis data (IC) for each of the 37 factors in the last
population seeding. In 9th International Conference on Learning Representations quarter of 2022.
(ICLR), Virtual Event, Austria, 2021.
[45] Chirag Deb, Fan Zhang, Junjing Yang, Siew Eang Lee, and Kwok Wei Shah. A 3. When provided with the test set (performance of the
review on time series forecasting techniques for building energy consumption. first 4 quarters of 2023):
Renewable and Sustainable Energy Reviews, 74:902–924, 2017.
[46] Aditya Ashok, Manimaran Govindarasu, and Venkataramana Ajjarapu. Online • Select the factors that will perform best in the last quarter
detection of stealthy false data injection attacks in power system state estimation. of the SSE 50.
IEEE Transactions on Smart Grid, 9(3):1636–1646, 2018. • Provide a confidence score & the risk preference for your
[47] Suppawong Tuarob, Conrad S. Tucker, Soundar R. T. Kumara, C. Lee Giles,
Aaron L. Pincus, David E. Conroy, and Nilam Ram. How are you feeling: A selection for each selected Alpha factor.
personalized methodology for predicting mental states from temporally observ- 4. Selection Criteria:
able physical and behavioral information. Journal of Biomedical Informatics,
68:1–19, 2017. • If no relationship between financial reports and IC can be
[48] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. found, select the Alpha factor with the highest IC value in
Artificial Intelligence Review, 42:275–293, 2014.
[49] Liping Wang, Jiawei Li, Lifan Zhao, Zhizhuo Kou, Xiaohan Wang, Xinyi Zhu, each group.
Hao Wang, Yanyan Shen, and Lei Chen. Methods for acquiring and incorporating • For verification of market information differences, if no re-
knowledge into stock price prediction: A survey. arXiv preprint arXiv:2308.04947,
2023.
lationship between financial reports and IR can be found,
select the Alpha factor with the highest IR value.
A Seed Alphas document lists for GPTs Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009
Automate Strategy Finding with LLM in Quant investment Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

Category Name Short Code

Momentum Price Momentum (CLOSE - DELAY(CLOSE, 14))
Volume Momentum (VOLUME - DELAY(VOLUME, 14))
RSI Momentum (RSI - DELAY(RSI, 14))
Rate of Change (ROC) ((CLOSE / DELAY(CLOSE, 14)) - 1)
MACD Momentum (MACD - DELAY(MACD, 14))
Momentum Oscillator ((CLOSE - DELAY(CLOSE, 14)) / DELAY(CLOSE, 14))
Chande Momentum Oscillator (SUM(IF(CLOSE - DELAY(CLOSE, 1) > 0, CLOSE - DELAY(CLOSE, 1),
(CMO) 0), 14) - SUM(IF(CLOSE - DELAY(CLOSE, 1) < 0, DELAY(CLOSE, 1)
- CLOSE, 0), 14)) / (SUM(IF(CLOSE - DELAY(CLOSE, 1) > 0, CLOSE -
DELAY(CLOSE, 1), 0), 14) + SUM(IF(CLOSE - DELAY(CLOSE, 1) < 0,
DELAY(CLOSE, 1) - CLOSE, 0), 14)) * 100
Stochastic Momentum Index (SMI) ((CLOSE - MIN(LOW, 14)) - (MAX(HIGH, 14) - CLOSE)) / (MAX(HIGH,
14) - MIN(LOW, 14))
ATR Momentum (ATR - DELAY(ATR, 14))
Detrended Price Oscillator (DPO) (CLOSE - DELAY(SMA(CLOSE, 14), 7))
Average Directional Index (ADX) (ADX - DELAY(ADX, 14))
Momentum
Mean Reversion Mean Reversion (MEAN(CLOSE, 20) - CLOSE)
Z-Score Mean Reversion (CLOSE - MEAN(CLOSE, 20)) / STD(CLOSE, 20)
Bollinger Bands (CLOSE - LOWER_BAND) / (UPPER_BAND - LOWER_BAND)
Keltner Channel (CLOSE - LOWER_CHANNEL) / (UPPER_CHANNEL -
LOWER_CHANNEL)
Moving Average Reversion (SMA(CLOSE, 20) - CLOSE)
Exponential Moving Average (EMA(CLOSE, 20) - CLOSE)
(EMA) Reversion
Distance from High (MAX(HIGH, 20) - CLOSE)
Distance from Low (CLOSE - MIN(LOW, 20))
Relative Strength Index (RSI) Rever- (100 - RSI)
sion
Percent B ((CLOSE - LOWER_BAND) / (UPPER_BAND - LOWER_BAND)) * 100
Volatility Standard Deviation STD(CLOSE, 20)
Average True Range (ATR) ATR(14)
Bollinger Band Width (UPPER_BAND - LOWER_BAND) / SMA(CLOSE, 20)
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al.

Historical Volatility STD(RETURNS, 20) * SQRT(252)

Volatility Ratio STD(CLOSE, 10) / STD(CLOSE, 50)
Chaikin Volatility (EMA(HIGH - LOW, 10) / DELAY(EMA(HIGH - LOW, 10), 10)) - 1
Garman-Klass Volatility SQRT(0.5 * LOG(HIGH / LOW)2 - (2 * LOG(2) - 1) * LOG(CLOSE /
OPEN)2 )
Parkinson Volatility SQRT((1 / (4 * N * LOG(2))) * SUM(LOG(HIGH / LOW)2 , 20))
Yang-Zhang Volatility SQRT(VAR(LOG(CLOSE / OPEN)) + 0.5 * VAR(LOG(HIGH / OPEN) -
LOG(LOW / OPEN)) + 0.25 * VAR(LOG(CLOSE / DELAY(OPEN, 1))))
Ulcer Index SQRT(MEAN(DRAWDOWN2 , 14))
Fundamental Price-to-Earnings Ratio (P/E) (CLOSE / EPS)
Price-to-Book Ratio (P/B) (CLOSE / BOOK_VALUE)
Dividend Yield (DIVIDENDS / CLOSE)
Earnings Yield (EPS / CLOSE)
Sales-to-Price Ratio (SALES / CLOSE)
Cash Flow Yield (OPERATING_CASH_FLOW / CLOSE)
Liquidity Trading Volume VOLUME
Average Trading Volume MEAN(VOLUME, 20)
Volume Rate of Change (VROC) (VOLUME - DELAY(VOLUME, 14)) / DELAY(VOLUME, 14)
On-Balance Volume (OBV) SUM(VOLUME * SIGN(CLOSE - DELAY(CLOSE, 1)))
Liquidity Ratio VOLUME / MARKET_CAP
Turnover Rate VOLUME / SHARES_OUTSTANDING
Amihud Illiquidity Ratio ABS(RETURN) / VOLUME
High-Low Spread (HIGH - LOW) / CLOSE
Dollar Volume VOLUME * CLOSE
Debt-to-Equity Ratio (TOTAL_DEBT / TOTAL_EQUITY)
Return on Equity (ROE) (NET_INCOME / EQUITY)
Return on Assets (ROA) (NET_INCOME / TOTAL_ASSETS)
Gross Profit Margin (GROSS_PROFIT / REVENUE)
Price-to-Sales Ratio (P/S) (CLOSE / SALES)
Price-to-Cash Flow Ratio (P/CF) (CLOSE / OPERATING_CASH_FLOW)
Book-to-Market Ratio (B/M) (BOOK_VALUE / CLOSE)
Enterprise Value to EBITDA (ENTERPRISE_VALUE / EBITDA)
(EV/EBITDA)
Bid-Ask Spread (ASK_PRICE - BID_PRICE) / MID_PRICE
High-Low Spread (HIGH - LOW) / CLOSE
Dollar Volume VOLUME * CLOSE
Quality Gross Profit Margin (GROSS_PROFIT / REVENUE)
Operating Profit Margin (OPERATING_INCOME / REVENUE)
Net Profit Margin (NET_INCOME / REVENUE)
Earnings Stability STD(EPS, 5) / MEAN(EPS, 5)
Debt to Equity Ratio (TOTAL_DEBT / TOTAL_EQUITY)
Interest Coverage Ratio (EBIT / INTEREST_EXPENSE)
Cash Conversion Cycle (DIO + DSO - DPO)
Asset Turnover Ratio (REVENUE / TOTAL_ASSETS)
Growth Earnings Growth Rate (EPS / DELAY(EPS, 1) - 1)
Revenue Growth Rate (REVENUE / DELAY(REVENUE, 1) - 1)
EBITDA Growth Rate (EBITDA / DELAY(EBITDA, 1) - 1)
Cash Flow Growth Rate (CASH_FLOW / DELAY(CASH_FLOW, 1) - 1)
Dividends Growth Rate (DIVIDENDS / DELAY(DIVIDENDS, 1) - 1)
Book Value Growth Rate (BOOK_VALUE / DELAY(BOOK_VALUE, 1) - 1)
Sales Growth Rate (SALES / DELAY(SALES, 1) - 1)
Asset Growth Rate (ASSETS / DELAY(ASSETS, 1) - 1)
Equity Growth Rate (EQUITY / DELAY(EQUITY, 1) - 1)
Retained Earnings Growth Rate (RETAINED_EARNINGS / DELAY(RETAINED_EARNINGS, 1) - 1)
Automate Strategy Finding with LLM in Quant investment Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

Technical Moving Average (MA) SMA(CLOSE, 20)

Exponential Moving Average EMA(CLOSE, 20)
(EMA)
Relative Strength Index (RSI) RSI(14)
Moving Average Convergence Di- (EMA(CLOSE, 12) - EMA(CLOSE, 26))
vergence (MACD)
Bollinger Bands UPPER_BAND - LOWER_BAND
Stochastic Oscillator ((CLOSE - MIN(LOW, 14)) / (MAX(HIGH, 14) - MIN(LOW, 14))) * 100
Average True Range (ATR) ATR(14)
Commodity Channel Index (CCI) (TYPICAL_PRICE - SMA(TYPICAL_PRICE, 20)) / (0.015 *
MEAN_DEV(TYPICAL_PRICE, 20))
Williams %R ((MAX(HIGH, 14) - CLOSE) / (MAX(HIGH, 14) - MIN(LOW, 14))) * -100
Macro Economics GDP Growth Rate GDP - DELAY(GDP, n)
Inflation Rate CPI - DELAY(CPI, n)
Unemployment Rate UNEMPLOYMENT_RATE - DELAY(UNEMPLOYMENT_RATE, n)
Interest Rate INTEREST_RATE - DELAY(INTEREST_RATE, n)
Industrial Production Index IPI - DELAY(IPI, n)
Retail Sales Growth RETAIL_SALES - DELAY(RETAIL_SALES, n)
Housing Starts Growth HOUSING_STARTS - DELAY(HOUSING_STARTS, n)
Consumer Confidence Index (CCI) CCI - DELAY(CCI, n)
Trade Balance EXPORTS - IMPORTS
Foreign Exchange Reserves FX_RESERVES - DELAY(FX_RESERVES, n)