A Survey of Financial AI: Architectures, Advances and Open Challenges

Junhua Liu
Forth AI
j@[Link]
arXiv:2411.12747v1 [[Link]] 1 Nov 2024

Abstract deep learning models, methodological advances in training

and optimization, and practical improvements in deployment
Financial AI empowers sophisticated approaches and scalability. These developments enable more robust pre-
to financial market forecasting, portfolio optimiza- diction under market uncertainty, more efficient portfolio op-
tion, and automated trading. This survey pro- timization under constraints, and more sophisticated trading
vides a systematic analysis of these developments strategies incorporating multiple information sources.
across three primary dimensions: predictive models
that capture complex market dynamics, decision- 1.1 Related Work
making frameworks that optimize trading and in-
vestment strategies, and knowledge augmentation Several recent surveys have explored the applications of
systems that leverage unstructured financial infor- LLMs in finance from different perspectives. [Lee et al.,
mation. We examine significant innovations in- 2024] present the first comprehensive review of financial
cluding foundation models for financial time series, LLMs (FinLLMs) including their evolution from general-
graph-based architectures for market relationship domain LMs to financial-domain LMs, and provide extensive
modeling, and hierarchical frameworks for portfo- coverage of model architectures, training data, and bench-
lio optimization. Analysis reveals crucial trade-offs marks. [Li et al., 2023] review the current approaches em-
between model sophistication and practical con- ploying LLMs in finance and propose a decision framework
straints, particularly in high-frequency trading ap- to guide their adoption, focusing on practical implementation
plications. We identify critical gaps and open chal- considerations. [Dong et al., 2024a] provide a scoping review
lenges between theoretical advances and industrial on ChatGPT and related LLMs specifically in accounting and
implementation, outlining open challenges and op- finance, highlighting use cases in these domains. [Zhao et al.,
portunities for improving both model performance 2024] comprehensively explore the integration of LLMs into
and practical applicability1 . various financial tasks and applications. [Ding et al., 2024a]
focus specifically on LLM agents in financial trading, exam-
ining architectures, data inputs, and empirical performance
1 Introduction through systematic analysis of 27 relevant papers.
Figure 1 shows the taxonomy of our work, which provides
Recent advancements in artificial intelligence, particularly greater breadth across financial applications while maintain-
large language models (LLMs), have significantly trans- ing rigorous depth in specific areas, as compared to the cur-
formed quantitative finance. These innovations span multiple rent surveys. Undoubtedly, these surveys have advanced our
domains including predictive modeling, decision making, and understanding of LLMs in finance. However, there remain
knowledge retrieval, enabling more sophisticated approaches significant gaps in addressing the broader implications for
to market analysis and trading automation. While Large Lan- financial decision-making and industry practices. Notably,
guage Models (LLMs) have garnered significant attention for many surveys lack detailed analysis of real-world deployment
their capabilities in natural language processing and reason- challenges, integration with existing systems, and compre-
ing, parallel developments in specialized architectures for fi- hensive evaluation frameworks that consider both technical
nancial applications demonstrate equal innovation and prac- and practical aspects. Our survey addresses these gaps by
tical impact. providing an expanded view of implementation challenges
Classical approaches to financial modeling face fundamen- and opportunities.
tal limitations in capturing complex market dynamics, han-
dling non-stationary distributions, and integrating diverse in- 1.2 Contributions
formation sources. Recent work addresses these challenges
through three primary directions: architectural innovations in Our survey makes four primary contributions that distinguish
it from prior work:
1 First, we provide a systematic analysis of the common for-
Full list of papers and summary slides are available at:
[Link] mulations, techniques, evaluations across the complete spec-
 Figure 1: Taxonomy of the survey paper

trum of financial AI applications. The analysis results in a enable more sophisticated approaches to financial modeling
comprehensive task-oriented categorization of the surveyed while highlighting crucial areas for future research. Our
works. analysis emphasizes practical implementation considerations
Second, we present a comprehensive review of architec- alongside theoretical advancements, providing a uniquely
tural and methodological advances beyond LLMs, including comprehensive resource for researchers and practitioners in
wide variations of Graph Neural Networks, Reinforcement computational finance.
Learning, and time series architectures, providing insights for 1.3 Survey Organization
optimal model selection based on specific financial applica-
tions. The survey is organized into eleven core sections. Section 2
establishes the mathematical preliminaries and notations for
Third, we identify critical gaps between theories and appli- both predictive and decision-making tasks in financial mar-
cations, bridging the gap between academia and industry, and kets. Section 2.4 presents a comprehensive evaluation frame-
inspiring future research investigation for both researchers work encompassing statistical accuracy metrics and trading
and practitioners in advancing the field. performance measures. Sections 3 to 5 examine predictive
Through rigorous examination of recent work across both tasks: continuous price forecasting, binary trend classifica-
academic literature and industry applications, we demon- tion, and ranking-based stock selection, analyzing architec-
strate how innovations in AI architecture and methodology tural innovations and empirical validations. Section 6 and 7
investigate decision-making tasks spanning portfolio opti- Table 1: Summary of Notations in Problem Formulations
mization and quantitative trading, with emphasis on multi-
agent frameworks and execution strategies. Section 8 ex- Symbol Description
plores knowledge retrieval and augmentation, focusing on au- S Asset pool
tomated analysis and market simulation. Section 9 surveys xti Feature vector for asset i at time t
recent financial datasets and benchmarks, while Section 10 Xi Temporal feature sequence for asset i
examines specialized time series models. Finally, Section 11 Mt Market state at time t
identifies promising research directions and open challenges G Asset relationship graph
in financial AI. Throughout, we maintain rigorous analysis of Tit Textual information for asset i at time t
theoretical foundations while emphasizing practical deploy- wt Portfolio weights at time t
ment considerations. Bt Available capital at time t
ct Transaction costs at time t
2 Preliminaries fθ Prediction function with parameters θ
Financial market applications can be broadly categorized into πϕ Policy function with parameters ϕ
predictive tasks (continuous price forecasting, binary trend hψ Encoder function with parameters ψ
classification, ranking-based selection) and decision-making pti Price of asset i at time t
tasks (portfolio optimization, quantitative trading). Despite rit Return of asset i at time t
their distinct objectives, these tasks share common mathemat- δ Threshold for binary classification
ical foundations and feature spaces. This section presents a
unified formulation framework and notations that encompass
both prediction and decision-making problems. Binary Trend Classification
2.1 Notation and Problem Setup The classification task extends the general forecasting frame-
work by discretizing price movements into directional cate-
Let S = {s1 , ..., sN } denote a pool of N assets. For each gories. The model learns a mapping function that predicts
asset si , we observe a temporal sequence of T historical fea- movement direction:
tures Xi = {xti }Tt=1 ∈ RT ×d , with d being the feature di-
mension. The feature vector xti encompasses three categories ŷit+h = fθ (xt−w+1:t
i , Mt , G, Tit ) (2)
of market information. The first category consists of price-
derived features, including open, high, low, close prices, and with target label:
trading volume, which capture direct market activities. The (
pt+h −pti
second category comprises technical indicators such as mov- 1 if i
>δ
yit+h = pti (3)
ing averages, relative strength index (RSI), and Bollinger 0 otherwise
bands, which provide insights into market momentum and
trends. The third category includes fundamental metrics like where δ represents a threshold parameter accounting for
price-to-earnings ratio, market capitalization, and liquidity transaction costs and market impact.
measures, which reflect underlying asset value and market
characteristics. Ranking-based Selection
The formulation incorporates several components to cap- The ranking task focuses on learning relative ordering of as-
ture broader market dynamics. The market state Mt ∈ Rdm sets based on their expected future performance through a
represents broader market conditions, economic indicators, scoring function:
and risk factors. The relational structure G = (V, E, R) mod-
r̂it+h = fθ (xit−w+1:t , Mt , G, Tit ) (4)
els complex relationships between assets, where V represents
the set of assets, E captures their interactions, and R defines which induces ranking:
different types of relationships. Textual information Tit from
sources such as news and financial reports provides qualita- πt = argsort({r̂it+h }N
i=1 ) (5)
tive context. The agent’s state st includes portfolio positions
wt ∈ RN , capital Bt , and transaction costs ct , essential for 2.3 Decision Making Tasks Formulation
decision-making tasks. Portfolio optimization and quantitative trading can be formu-
lated as sequential decision-making problems under uncer-
2.2 Predictive Tasks Formulations tainty, typically modeled through Markov Decision Processes
Continuous Price Forecasting (MDPs) or their variants.
The continuous forecasting task learns a function fθ mapping
historical observations to future values over horizon h: Portfolio Optimization
ŷit+1:t+h = fθ (xit−w+1:t , Mt , G, Tit ) (1) The portfolio optimization task aims to determine optimal as-
set allocations over time. Let wt = [w1t , ..., wN t
] represent
The target variable y can take several forms: raw price lev- t
portfolio weights at time t, where wi denotes the proportion
els yitp= pti , returns yit = (pti − pt−1
i )/pt−1
i , or volatility of capital allocated to asset i. The state space st ∈ S encom-
t t t 2 t
yi = E[(ri − µi ) ], where ri represents the log-return and passes market features Xt , current portfolio weights wt−1 ,
µti its mean. and available capital Bt :
 Table 2: Summary of Notations In Performance Evaluation
st = [Xt , wt−1 , Bt ] (6) Symbol Description
Rp Portfolio return
The action space at ∈ A defines target portfolio weights:
Rf Risk-free rate (3-month Treasury bill yield)
N Rm Market return (S&P 500 index)
Rb Benchmark return (relevant market index)
X
at = wt , s.t. wit = 1, wit ≥ 0 (7)
i=1
σp Portfolio volatility
σp−b Relative portfolio volatility
The reward function incorporates both returns and transac- σstrategy Strategy volatility
tion costs: σbenchmark Benchmark volatility
TP True Positive
N N
X X TN True Negative
rt = wit rit −c |wit − wit−1 | (8) FP False Positive
i=1 i=1 FN False Negative
Precision TP / (TP + FP)
The objective is to learn a policy πθ : S → A maximizing Recall TP / (TP + FN)
expected cumulative returns: E Expected value operator
" T # wit Weight of asset i at time t
max E
X
t
γ rt |πθ (9) qt Trading quantity at time t
θ Pte Expected execution price at time t
t=1
Pta Actual execution price at time t
Quantitative Trading
Trading can be formulated as a POMDP where the true mar- Statistical Metrics
ket state is partially observable. The agent’s belief state bt For continuous price forecasting, prediction quality is mea-
combines observable market features with internal estimates sured through:
of latent variables:
N
1 X
bt = [Xt , ht , pt , vt ] (10) MSE = (ŷi − yi )2 (15)
N i=1
where ht represents hidden state estimates, pt is position IC = corr(ŷ, y) (16)
size, and vt is remaining capital.
The action space includes both trade direction and size: Binary classification employs:

at = (dt , qt ), dt ∈ {−1, 0, 1}, 0 ≤ qt ≤ Qmax (11) TP · TN − FP · FN

MCC = p
(TP + FP)(TP + FN)(TN + FP)(TN + FN)
The transition function incorporates market impact:
(17)
pt+1 = pt + dt qt (12) 2 · Precision · Recall
F1 = (18)
Precision + Recall
vt+1 = vt − dt qt (pt + αqt ) (13)
Risk-Adjusted Return Metrics
where α models price impact. The reward balances profit Portfolio performance is assessed through:
against risk:
Rp − Rf
SR = (19)
rt = dt qt (pt+1 − pt ) − c|qt | − λRisk(pt , qt ) (14) σp

Both tasks can be solved through various reinforcement Rp − Rf

CR = (20)
learning approaches, with policy gradient methods being MDD
particularly suitable due to their ability to handle continu- where Maximum Drawdown (MDD) captures downside
ous action spaces and complex constraints. The choice be- risk:
tween model-based and model-free approaches depends on
the trade-off between sample efficiency and computational MDD = max{max(Vτ ) − Vt }/ max(Vτ ) (21)
t τ ≤t τ ≤t
complexity.
Alpha measures excess returns over a benchmark:
2.4 Performance Evaluation
α = Rp − [β(Rm − Rf ) + Rf ] (22)
Definitions and Notations The Information Ratio quantifies consistency of excess re-
The following notations are used throughout this section: turns:
 Baselines
Rp − Rb Performance is benchmarked against:
IR = (23)
σp−b • Traditional strategies: Equal-weight, Minimum Variance,
Long-term performance is captured through: Maximum Sharpe
T • Market indices: S&P 500, NASDAQ Composite
252 X
AR = rt (24) • Statistical models: ARIMA, GARCH
T t=1
• Machine learning baselines: LSTM, GRU, Transformer
CAGR = (1 + RT )252/T − 1 (25) All metrics are computed using walk-forward optimization with
Volatility-adjusted measures include: appropriate lookback windows to prevent forward-looking bias.
σstrategy
λVol = (26) 3 Continuous Price Forecasting
σbenchmark
Continuous price forecasting represents a fundamental challenge in
Trading Performance Metrics quantitative finance, combining elements of time series analysis,
Trading strategy evaluation incorporates: market microstructure, and behavioral finance. Recent works in AI
have led to tremendous improvements in forecasting accuracy and
T
X robustness, particularly through the integration of complex market
rt − c|qt | − αqt2


Net Return = (27) relationships, adaptive learning mechanisms, and multimodal data
t=1 sources. As summarized in Table 3, these works center around three
The Win-Loss Ratio measures trading effectiveness: primary areas, namely, relational learning approaches that capture
market structure, distribution shift modeling that addresses tempo-
Number of Profitable Trades ral dynamics, and hybrid approaches that integrate multiple infor-
WLR = (28) mation modalities. The contributions range from theoretical inno-
Number of Loss-Making Trades
vations in model architecture (e.g., MASTER’s dynamic correla-
Execution Quality Metrics tion modeling) to practical applications in performance metrics (e.g.,
Portfolio optimization strategies are evaluated on rebalancing DIFFSTOCK’s improvements in Sharpe ratios). This section exam-
efficiency: ines these developments systematically, beginning with a rigorous
mathematical formulation of the forecasting problem, followed by
T N detailed analysis of methodological innovations and their empirical
1 XX t
Turnover = |w − wit−1 | (29) validation across different market contexts.
T t=1 i=1 i
Trading execution quality is measured through Implemen- 3.1 Relational Learning Approaches
tation Shortfall: The complex interdependencies in financial markets naturally mo-
tivate graph-based and relational learning approaches, which can
T capture both explicit market relationships and implicit correlations
1X e
IS = (P − Pta )qt (30) between assets. Recent works have demonstrated the effectiveness
T t=1 t of such approaches through increasingly sophisticated architectures.
where Pte and Pta are expected and actual execution prices. MDGNN [Xu et al., 2024] established a fundamental framework
by modeling multi-relational market structure through a hierarchical
Stability and Robustness Metrics graph representation. The framework employs a transformer struc-
Strategy stability is assessed through portfolio diversification ture with ALIBI position encoding to capture temporal evolution
measures: patterns, demonstrating superior performance on China’s CSI300
index with significant improvements in Information Coefficient and
N Cumulative Return metrics.
1 X
Extending this relational modeling paradigm, MASTER [Li et al.,
Effective N = PN , wi2
HHI = (31)
w 2 2024b] introduced a more dynamic approach through a specialized
i=1 i i=1
transformer architecture. Unlike MDGNN’s static graph structure,
Robustness across market regimes is evaluated through
MASTER alternates between intra-stock and inter-stock information
conditional metrics: aggregation, enabling the capture of time-varying relationships. Its
market-guided gating mechanism dynamically selects relevant fea-
Rpk − Rf tures based on market conditions, achieving a 13% improvement in
Regime SR = , k ∈ {bull, bear, neutral} (32) ranking metrics and 47% in portfolio-based metrics on CSI300 and
σpk
CSI800 markets.
Strategy timing ability is assessed through the Treynor- DANSMP [Zhao et al., 2022] further advanced the field by incor-
Mazuy model: porating higher-order market relationships through a comprehensive
market knowledge graph (MKG). The model expands beyond tra-
Rp − Rf = α + β(Rm − Rf ) + γ(Rm − Rf )2 + ϵ (33) ditional asset-to-asset relationships by integrating executive entities
and implicit connections, processing these diverse signals through
Multi-factor performance attribution employs the Fama- dual attention mechanisms. This richer relational modeling led to
French framework: superior performance on the CSI300E dataset, achieving investment
returns of 16.97% with Sharpe ratios of 4.628, demonstrating the
Rp −Rf = α+βM (Rm −Rf )+βS SMB+βH HML+βM MOM+ϵ value of incorporating executive-level relationships in prediction
(34) models.
 Table 3: Summary of Recent Contributions in Continuous Price Forecasting
Work Key Innovation Methodology Primary Results
MDGNN Multi-relational graph incorpo- Hierarchical graph embed- IC improvement: 15% (CSI300)
[Xu et al., 2024] rating industry and institutional ding with ALIBI position Stronger performance on larger
relationships encoding datasets
MASTER Dynamic stock correlation mod- Market-guided gating mech- Ranking metrics: +13%
[Li et al., 2024b] eling through alternating atten- anism for feature selection Portfolio metrics: +47%
tion (CSI300, CSI800)
DoubleAdapt Dual adaptation for distribution Multi-head transformation Consistent improvement across
[Zhao et al., 2023] shifts in streaming data layers for feature and label LSTM, GRU, Transformer im-
adaptation plementations
Model-agnostic adaptability
DPA-STIFormer Feature-centric temporal mod- Double-path mechanism Superior performance across
[Yan and Tan, 2024] eling with adaptive gating CSI500, CSI1000, NASDAQ,
NYSE
Improved IC and Sharpe Ratio
DANSMP Executive-level market knowl- Dual attention network for Returns: 16.97%
[Zhao et al., 2022] edge graph complex entity interactions Sharpe Ratio: 4.628
(CSI300E)
GINN Integration of GARCH theory GARCH-based regulariza- Consistent outperformance
[Xu et al., 2024] with neural networks tion in LSTM loss function across 7 global indices
Strong zero-shot performance
DIFFSTOCK Diffusion models for market Adaptive noise scheduling Sharpe Ratio improvement:
[Daiya et al., 2024] prediction for stock volatility NYSE: +7.92%
NASDAQ: +6.18%
[Wang et al., 2024b] LLM integration for news pro- Text-based transformation Superior performance in event-
cessing of numerical forecasting driven market shifts
Multi-domain applicability

3.2 Distribution Shift and Temporal Dynamics (7.92% and 6.18% increases for NYSE and NASDAQ respectively),
though with increased computational requirements.
The non-stationary nature of financial markets presents a fundamen-
tal challenge to traditional machine learning approaches, spurring
the development of methods specifically designed to handle tempo-
ral distribution shifts. DoubleAdapt [Zhao et al., 2023] addressed 3.3 multimodal Fusion
this challenge through a meta-learning framework that implements
dual adaptation mechanisms. The framework’s multi-head transfor-
mation layers adapt both features and labels into locally stationary The integration of traditional financial theory with modern machine
distributions, while its model adapter learns initialization parameters learning techniques has emerged as a promising direction for im-
that enable quick adaptation to new data. This approach showed con- proving forecast robustness. GINN [Xu et al., 2024] exemplifies
sistent improvements across various time-series architectures, effec- this approach by bridging classical GARCH models with neural net-
tively handling both gradual and sudden market shifts. works. The framework incorporates GARCH predictions as a regu-
Building on these insights, DPA-STIFormer [Yan and Tan, 2024] larization term in the LSTM’s loss function, effectively combining
introduced a novel perspective on temporal modeling by treating fea- statistical finance theory with deep learning flexibility. Evaluated on
tures rather than time steps as tokens. Its double-path mechanism seven global market indices over a 30-year period, both GINN and
adaptively learns stock relationships, while a specialized decoder its variant GINN-0 demonstrated consistent outperformance over
decomposes predictions into mean and deviation components. This traditional approaches, though with a trade-off in capturing market
architecture proved particularly effective across diverse market con- volatility.
ditions, demonstrating robust performance on four major markets Taking a different approach to multimodal integration, [Wang et
(CSI500, CSI1000, NASDAQ, NYSE). al., 2024b] leveraged recent works in large language models to in-
DIFFSTOCK [Daiya et al., 2024] approached the distribution corporate unstructured news data into forecasting. The framework
shift challenge from a generative modeling perspective, leverag- transforms numerical forecasting into a text-based task, employing
ing denoising diffusion models. The framework’s Masked Re- LLM-based agents for news filtering and prediction evaluation. This
lational Transformer architecture processes different relationship approach proved particularly effective in capturing sudden market
types through separate attention heads, while its adaptive noise shifts caused by external events, demonstrating the value of incor-
schedule accounts for both individual stock volatility and intra- porating qualitative information in traditionally quantitative predic-
cluster dynamics. This approach greatly improves Sharpe ratios tions.
 Table 4: Comparison of Recent Continuous Price Forecasting Approaches

Model Temporal Relational Distribution multimodal Compute Market Coverage

Modeling Learning Shift Integration Cost (Datasets)
MDGNN ✓ ✓ - - Medium China (CSI100, CSI300)
MASTER ✓ ✓ ✓ - High China (CSI300, CSI800)
DoubleAdapt ✓ - ✓ - Low China (CSI300, CSI500)
DPA-STIFormer ✓ ✓ ✓ - High China (CSI500, CSI1000),
US (NASDAQ, NYSE)
GINN ✓ - - ✓ Low Global (7 major market indices,
30-year period)
DIFFSTOCK ✓ ✓ ✓ - Very High US (NASDAQ, NYSE, StockNet)
DANSMP ✓ ✓ - ✓ High China (CSI100E, CSI300E)

3.4 Comparative Analysis spatial relational modeling with stock dynamics modeling. The spa-
The reviewed approaches exhibit distinct characteristics in terms of tial component aggregates multimodal information across the graph
their modeling capacity, computational efficiency, and practical ap- structure, while the dynamics component captures post-earnings an-
plicability. Table 4 summarizes the key aspects of representative nouncement effects through learnable stochastic processes. When
models. evaluated on the S&P 500 and NASDAQ-100 indices from 2018-
In terms of architectural design, models exhibit a clear trade-off 2023, this architecture achieved a 2.297% increase in F1 score and
between modeling capacity and computational efficiency. MASTER 15.629% increase in MCC over baseline methods, with particu-
and DPA-STIFormer achieve superior performance through sophis- larly strong performance during earnings seasons. MGDPR [You
ticated attention mechanisms but require significant computational et al., 2024] advanced the field by introducing dynamic relation-
resources. In contrast, DoubleAdapt and GINN maintain efficiency ship modeling through information entropy and signal energy. The
through focused adaptation mechanisms and classical model inte- framework’s innovation lies in its multi-relational diffusion process,
gration, respectively. which adaptively learns and updates relationship strengths between
Market coverage and generalization capabilities also vary sig- assets. Implementation challenges, particularly in computational ef-
nificantly. While models like MDGNN and DANSMP demon- ficiency for large-scale markets, were addressed through sparse ma-
strate strong performance on Chinese markets, DPA-STIFormer and trix operations and optimized graph convolutions. This approach
DIFFSTOCK show broader applicability across both US and Chi- enabled more effective capture of market dynamics, leading to con-
nese markets. GINN stands out with its global market coverage, sistent outperformance in next-day trend forecasting across NAS-
though focusing on major indices rather than individual stocks. This DAQ, NYSE, and Shanghai markets over a seven-year test period
market specialization versus generalization represents a key consid- from 2016-2023.
eration in model selection and development.
The integration of domain knowledge presents another key dif- 4.2 Denoising and Debiasing
ferentiator. GINN’s incorporation of GARCH theory leads to more
interpretable and theoretically grounded predictions, while purely Financial markets are characterized by high noise levels and non-
data-driven approaches like DIFFSTOCK achieve higher perfor- stationary distributions, leading to the development of specialized
mance at the cost of reduced interpretability. denoising and adaptation techniques. LARA [Zeng et al., 2024] in-
troduced a comprehensive framework combining locality-aware at-
tention with iterative refinement. The framework’s two-stage ap-
4 Binary Trend/Movement Classification proach first identifies profitable trading opportunities through metric
Binary trend classification represents a critical task in quantitative learning, then iteratively refines noisy labels to improve prediction
finance that focuses on predicting directional price movements in fi- robustness. Evaluated on high-frequency data from China’s A-share
nancial markets. Unlike continuous price forecasting which aims to market (2020-2023), cryptocurrency markets (2021-2023), and ETF
predict exact values, this task addresses the fundamental question of markets (2019-2023), this architecture demonstrated remarkable re-
price movement direction, making it particularly relevant for trading silience in volatile market conditions, achieving up to 59.1% preci-
strategies and risk management. Prediction accuracy are improved sion while maintaining computational efficiency through optimized
by market structure modeling, noise handling, and information fu- attention mechanisms. MANA-Net [Wang and Ma, 2024] tackled
sion. These developments span three primary directions: graph- the critical challenge of ”aggregated sentiment homogenization” in
based relational modeling that captures complex market dependen- financial news analysis. The model employs a dynamic market-news
cies, denoising techniques that address market noise and distribution attention mechanism that weights news items based on their market
shifts, and multimodal approaches that integrate diverse information relevance, integrating news aggregation and market prediction into
sources, as summarized in Table 5. a unified framework. Validated on an extensive dataset spanning
2003-2018 with over 2.7 million news items from major financial
4.1 Graph-based Relational Learning news sources, MANA-Net achieved a 1.1% increase in Profit and
The complex interdependencies in financial markets have motivated Loss and a 0.252 increase in daily Sharpe ratio, demonstrating the
the development of sophisticated graph-based architectures that cap- value of sophisticated news aggregation in market prediction. These
ture both explicit and implicit relationships between assets. ECHO- binary trend classification approaches exhibit distinct characteristics
GL [Liu et al., 2024a] established a comprehensive framework for in terms of their modeling capacity, computational efficiency, and
modeling market relationships through heterogeneous graph learn- practical applicability. Table 6 summarizes the key aspects of repre-
ing. The model introduces a dual-mechanism approach combining sentative models.
 Table 5: Summary of Recent Contributions in Binary Trend Classification

Work Key Innovation Methodology Primary Results

ECHO-GL Heterogeneous graph learning Dual-mechanism approach F1: +2.297%
[Liu et al., 2024a] from earnings calls data combining spatial relations MCC: +15.629%
with stock dynamics (S&P500, NASDAQ-100)
MGDPR Dynamic relationship model- Multi-relational diffusion pro- Consistent outperformance
[You et al., 2024] ing through information en- cess with parallel retention across NASDAQ, NYSE,
tropy Shanghai (2016-2023)
LARA Two-stage denoising frame- Iterative refinement labeling Precision: 59.1%
[Zeng et al., 2024] work with locality-aware at- for noise reduction Improved win-loss ratio in
tention volatile markets
MANA-Net Dynamic market-news atten- Trainable sentiment aggrega- P&L: +1.1%
[Wang and Ma, 2024] tion mechanism tion optimized for prediction Daily SR: +0.252
SH-Mix Hierarchical multimodal aug- Modality-specific and span- Performance improvement: 3-
[Jain et al., 2024] mentation strategy based mixing techniques 7% across tasks
SEP Self-reflective LLM frame- Three-stage pipeline with Superior performance in both
[Koa et al., 2024a] work for autonomous learning PPO optimization prediction and explanation
quality

Table 6: Comparison of Recent Binary Trend Classification Approaches

Model Temporal Graph Distribution Multimodal Compute Market Coverage

Modeling Learning Handling Integration Cost (Datasets)
ECHO-GL ✓ ✓ - ✓ High US (S&P500, NASDAQ-100)
MGDPR ✓ ✓ ✓ - Medium Global (NASDAQ, NYSE, SSE)
LARA ✓ - ✓ - Low China (A-share), Crypto, ETFs
MANA-Net ✓ - ✓ ✓ Medium US (S&P500)
SH-Mix ✓ - - ✓ Medium US (S&P500)
SEP - - - ✓ Very High Global (Multiple Indices)

4.3 Multimodal Fusion accuracy through comprehensive market structure modeling but re-
The integration of multiple data modalities presents unique chal- quire significant computational resources. In contrast, LARA main-
lenges in feature alignment and data scarcity. SH-Mix [Jain et al., tains efficiency through focused denoising mechanisms while sacri-
2024] developed a hierarchical augmentation strategy operating at ficing some modeling capacity.
both local and global levels. The framework performs modality- Market coverage and generalization capabilities vary significantly
specific mixing based on feature importance locally, while apply- across approaches. MGDPR demonstrates strong cross-market ap-
ing span-based mixing on fused representations globally. Built plicability spanning US, Chinese, and European markets, while
upon an attention-driven fusion architecture and evaluated on earn- models like MANA-Net and SH-Mix show specialized performance
ings call datasets from S&P 500 companies (2019-2023), the ap- on US markets. SEP’s LLM-based approach offers global cover-
proach achieved 3-7% improvement over existing methods while age but with higher computational requirements and less predictable
demonstrating strong generalization capabilities across various mul- performance characteristics.
timodal tasks. The emergence of large language models has enabled The integration of domain knowledge presents another key dif-
new approaches to feature extraction and prediction. SEP [Koa et ferentiator. ECHO-GL’s incorporation of earnings call informa-
al., 2024a] implements a three-stage pipeline combining summa- tion and MANA-Net’s sophisticated news processing lead to more
rization, explanation, and prediction components. The framework, interpretable predictions, while purely data-driven approaches like
tested on market data and financial news from 2020-2023, allows LARA achieve robust performance through statistical learning. This
LLMs to autonomously learn and improve stock predictions with- trade-off between interpretability and performance represents a cru-
out human expert intervention, demonstrating superior performance cial consideration in model selection.
over both traditional deep learning and existing LLM approaches in
both prediction accuracy and explanation quality.
5 Ranking-based Stock Selection
4.4 Comparative Analysis Ranking-based stock selection represents a fundamental task in
The architectural approaches demonstrate clear trade-offs between quantitative finance that aims to order a universe of assets based
modeling sophistication and computational efficiency. Graph-based on their expected performance. Unlike continuous price forecasting
models like ECHO-GL and MGDPR achieve superior prediction which predicts exact values or binary classification which focuses
 Table 7: Summary of Recent Contributions in Ranking-based Stock Selection
Work Key Innovation Methodological Contribu- Primary Results
tion
CI-STHPAN Channel-independent pre- Self-supervised framework IR: +21.3%, IC: +15.7%,
[Xia et al., 2024b] training with dynamic hyper- with reversible normalization Outperforms SOTA on NYSE,
graph learning NASDAQ
ADB-TRM Meta-learning framework for Temporal-relational adversar- Returns: +28.41%
[Chen et al., 2024] dual-level bias mitigation ial training with global distri- Risk-adjusted Returns: +9.53%
bution adaptation (NYSE, NASDAQ, TSE)
RSAP-DFM Dual regime-shifting mecha- Gradient-based posterior fac- Factor Returns: +18.2%
[Xiang et al., 2024] nism for factor modeling tors with adversarial learning Robust macro-state adaptation
(A-share market)
RT-GCN Pure convolutional temporal- Three-strategy information Returns: +40.4%
[Zheng et al., 2023] relational modeling propagation with time- Training Time: 13.4× faster
sensitive weighting (NASDAQ, NYSE, CSI)

Table 8: Comparison of Recent Ranking-based Stock Selection Approaches

Model Pre-training Factor Distribution Compute Market Coverage

Integration Modeling Adaptation Cost (Datasets)
CI-STHPAN ✓ - ✓ High US (NYSE, NASDAQ)
RT-GCN - - ✓ Low Global (NYSE, NASDAQ, CSI)
ADB-TRM - - ✓ Medium Global (NYSE, NASDAQ, TSE)
RSAP-DFM - ✓ ✓ High China (A-share)

on directional movement, ranking approaches learn relative order- accuracy, achieving up to 40.4% improvement in investment returns
ings that directly inform portfolio construction decisions. Rank- while reducing computational time by 13.4× compared to existing
ing accuracy and robustness are optimized through innovations in methods across NASDAQ, NYSE, and CSI markets.
self-supervised learning, bias mitigation, and integration with fac-
tor models. As summarized in Table 7, these works provide three 5.2 Denoising and Debiasing
primary directions: pre-training approaches that leverage unlabeled
data, debiasing techniques that address various market biases, and Financial markets exhibit various biases and non-stationary dis-
hybrid approaches that combine deep learning with traditional fi- tributions that challenge traditional learning approaches. ADB-
nance theory. TRM [Chen et al., 2024] introduced a comprehensive framework
addressing both micro-level biases in stock data and macro-level dis-
5.1 Self-supervised Pre-training tribution shifts. The model employs temporal adversarial training to
The abundance of unlabeled financial data has motivated the devel- handle inherent noise while using relational adversarial training to
opment of sophisticated pre-training approaches that learn mean- mitigate momentum spillover effects. Its meta-learning framework
ingful representations before fine-tuning for ranking tasks. CI- enables adaptation to changing market conditions through invariant
STHPAN [Xia et al., 2024b] established a comprehensive frame- feature extraction. Evaluated across NYSE, NASDAQ, and Tokyo
work combining channel-independent processing with dynamic hy- Stock Exchange, this approach demonstrated noticeable improve-
pergraph learning. The model’s innovation lies in constructing adap- ments of 28.41% in cumulative returns and 9.53% in risk-adjusted
tive hypergraphs based on time series similarities using Dynamic returns while maintaining computational efficiency.
Time Warping, moving beyond predefined relationships. To address
distribution shifts, the framework incorporates reversible instance 5.3 Factor Models
normalization and employs a two-stage training process. When eval-
uated on NASDAQ and NYSE markets over five years, this architec- The incorporation of traditional financial theory with modern ma-
ture achieved 21.3% improvement in Information Ratio and 15.7% chine learning has emerged as a promising direction for improv-
in Information Coefficient, with particularly strong performance in ing ranking robustness. RSAP-DFM [Xiang et al., 2024] pioneered
capturing complex market dependencies. this approach through a dual regime-shifting mechanism that con-
RT-GCN [Zheng et al., 2023] advanced the field through a pure tinuously captures macroeconomic states and their impact on fac-
convolutional approach to temporal-relational modeling. The frame- tor dynamics. The framework employs multi-head attention for dy-
work’s key innovation lies in its unified treatment of temporal namic factor generation while introducing gradient-based posterior
patterns and stock relationships through a relation-temporal graph factors through adversarial learning. A novel bilevel optimization al-
structure. The model employs three relation-aware strategies for gorithm separates factor construction from model optimization, en-
information propagation - uniform, weighted, and time-sensitive - abling more efficient training. Tested on China’s A-share market, the
with adaptive weighting based on temporal dynamics. This archi- model achieved 18.2% improvement in factor returns while provid-
tecture enables significantly faster training while maintaining high ing explicit interpretability through macroeconomic state mapping.
5.4 Comparative Analysis patterns at different frequencies, processing these through a Fre-
These ranking approaches exhibit distinct characteristics in terms of quency State Encoder for multi-granular asset representation. This
their modeling capacity, computational efficiency, and practical ap- approach achieved up to 2.1× improvement in Annualized Rate of
plicability. Table 8 summarizes key aspects of representative mod- Return across U.S., Korean, and cryptocurrency markets, with par-
els. ticular strength in handling market regime shifts.
The architectural approaches demonstrate clear trade-offs be- TrendTrader [Ding et al., 2024b] complemented frequency analy-
tween modeling sophistication and computational efficiency. Pre- sis with multimodal integration, combining price patterns with news
training models like CI-STHPAN achieve superior ranking accuracy sentiment through a spatial-temporal backbone network. Its incre-
through comprehensive market structure modeling but require sig- mental learning approach for portfolio weights demonstrated robust
nificant computational resources. In contrast, RT-GCN maintains performance across DJIA and SSE-50 indices, particularly in cap-
efficiency through focused convolutional processing while achiev- turing sentiment-driven market movements.
ing competitive performance.
Market coverage and generalization capabilities vary significantly 6.4 Graph Models
across approaches. RT-GCN and ADB-TRM demonstrate strong
The complex dependencies in financial markets have motivated
cross-market applicability spanning multiple global markets, while
network-based approaches to risk management. [Hui and Wang,
RSAP-DFM shows specialized performance on Chinese markets
2024] developed a framework using Graph Theory and Extreme
through its integration with factor models.
Value Theory to mitigate extremal risks. The approach constructs
graphs based on extremal dependencies between stock returns, us-
6 Portfolio Optimization ing maximum independent sets for diversification. Tested on CSI
Portfolio optimization represents a fundamental challenge in quan- 300 components, this method outperformed traditional sector-based
titative finance that involves allocating capital across multiple as- approaches, especially during market downturns.
sets to maximize risk-adjusted returns. Recent work in Financial AI [Yamagata and Ono, 2024] advanced this direction by replacing
have significantly improved portfolio management through innova- sector-based diversification with market graph-based clustering. The
tions in multi-agent systems, frequency domain analysis, and risk framework incorporates turnover sparsity regularization to manage
modeling. As summarized in Table 9, these works centered around transaction costs while ensuring cluster-based diversification. Ex-
three primary directions: agent-based approaches that enable adap- periments on S&P500 demonstrated superior Sharpe ratios com-
tive management, frequency-based methods that capture market dy- pared to traditional methods, particularly in challenging market con-
namics across different time scales, and network-centric techniques ditions.
that model complex dependencies for risk management. The architectural approaches demonstrate clear trade-offs be-
tween modeling sophistication and computational efficiency. Multi-
6.1 Comparative Analysis agent systems like MASA achieve superior performance through
comprehensive market modeling but require significant computa-
These portfolio optimization approaches exhibit distinct character-
tional resources. In contrast, EarnMore maintains efficiency through
istics in terms of their modeling capacity, computational efficiency,
focused token-based processing while achieving competitive perfor-
and practical applicability. Table 10 summarizes key aspects of rep-
mance.
resentative models.

6.2 Agentic Management 7 Quantitative Trading

The complexity of portfolio management has motivated the develop-
ment of sophisticated multi-agent architectures that decompose the Quantitative trading represents a systematic approach to financial
optimization problem into specialized components. MASA [Li et markets that leverages mathematical models and computational
al., 2024e] established a comprehensive framework using three co- methods to develop and execute trading strategies. Recent works
operative agents: a trend observer for market monitoring, a return in Financial AI have significantly enhanced quantitative trading
optimizer for portfolio maximization, and a risk manager for risk through innovations in three key areas: strategy development in-
minimization. The framework’s key innovation lies in its reward- corporating predictive signals and market dynamics, execution op-
based guiding mechanism that combines return and action rewards timization through hierarchical control, and high-frequency trading
to maintain strategy diversity while adapting to market conditions. systems that adapt to market microstructure. Unlike traditional ap-
When evaluated on major indices (CSI300, DJIA, S&P500), this proaches that rely on fixed rules or simple statistical arbitrage, mod-
architecture demonstrated superior performance in both stable and ern quantitative trading systems employ sophisticated deep learning
volatile markets. architectures to capture complex market patterns while maintain-
EarnMore [Zhang et al., 2024d] advanced the field through a cus- ing computational efficiency for real-time deployment. As summa-
tomizable approach to portfolio management. The framework in- rized in Table 11, these works propose predictive modeling frame-
troduces a maskable token system to represent unfavorable stocks works that extract meaningful signals from noisy market data, exe-
and employs self-supervised learning for relationship modeling. Its cution optimization methods that decompose complex trading deci-
one-shot training capability enables efficient adaptation to different sions into manageable components, and adaptive systems that han-
stock pools while maintaining performance, achieving 40% profit dle the unique challenges of high-frequency market environments.
improvement across diverse market conditions. This section examines these developments systematically, analyzing
their theoretical foundations, architectural innovations, and empiri-
6.3 Frequency-based Models cal validation across different market contexts.
The multi-scale nature of market patterns has inspired approaches
that explicitly model different frequency components. Fre-
7.1 Signal Detection
Quant [Jeon et al., 2024] pioneered this direction through a rein- Signal Detection primarily focuses on extracting robust signals from
forcement learning framework operating in the frequency domain. complex market data. StockFormer [Gao et al., 2023] pioneered a
The model employs Discrete Fourier Transform to identify market three-branch transformer architecture that captures long-term trends,
 Table 9: Summary of Recent Contributions in Portfolio Optimization

Work Key Innovation Methodological Primary Results

MASA Multi-agent framework with Reward-based guiding mech- Superior risk-adjusted re-
[Li et al., 2024e] specialized risk and return anism for strategy diversity turns across CSI300, DJIA,
agents S&P500
FreQuant Frequency domain analysis Multi-granular asset represen- ARR: 2.1× improvement, En-
[Jeon et al., 2024] for pattern identification tation through DFT hanced stability in market
shifts
EarnMore Customizable portfolio pools Self-supervised masking and Profit: +40%, Comparable
[Zhang et al., with maskable tokens reconstruction risk levels
2024d]
TrendTrader Multimodal fusion of price Spatial-temporal RL frame- Superior ARR, ASR, MDD in
[Ding et al., 2024b] and sentiment work DJIA, SSE-50
Network-EDM Extremal risk mitigation Maximum independent sets Outperformance in market
[Hui and Wang, through network theory for diversification downturns on CSI 300
2024]
Market-Graph Market graph-based cluster- Turnover sparsity regulariza- Higher Sharpe ratios on
[Yamagata and ing for index tracking tion with primal-dual splitting S&P500
Ono, 2024]
LLM-Alpha LLM-based alpha factor min- Multi-agent system with dy- Return: +53.17% vs -11.73%
[Kou et al., 2024] ing namic weight-gating market (China, 2023)

Table 10: Comparison of Recent Portfolio Optimization Approaches

Model Adaptive Risk Training Compute Market Geographic/Asset

Management Paradigm Cost Coverage (Datasets)
MASA ✓ ✓ MARL High Global CSI300, DJIA, S&P500
FreQuant ✓ - RL (DFT) Medium Global + Crypto Global + Crypto
EarnMore ✓ ✓ RL + SSL Low Multiple Markets Multiple Markets
TrendTrader ✓ - RL High Regional DJIA, SSE-50
Network-EDM - ✓ Optimization Medium Regional China (CSI 300)
Market-Graph - ✓ Graph Learning Medium Regional US (S&P500)
LLM-Alpha ✓ ✓ LLM + RL Very High Regional China (A-shares)

short-term movements, and inter-asset relationships through spe- markets.

cialized attention mechanisms. Its diversified multi-head atten- IMM [Niu et al., 2024b] automates market making through multi-
tion design maintains pattern diversity across concurrent time se- price level strategies, employing temporal-spatial attention networks
ries, demonstrating particular effectiveness in volatile cryptocur- to mitigate adverse selection risk. Its imitation learning approach
rency markets with 40.3% improvement in portfolio returns. from expert strategies facilitates efficient exploration of complex ac-
TRR [Zhang et al., 2024c] represents a novel direction in in- tion spaces, demonstrating improved risk-adjusted returns in futures
corporating qualitative data through LLMs, enabling zero-shot rea- markets.
soning for unprecedented market events. Its four-phase framework HRT [Zhao and Welsch, 2024] tackles the joint problem of stock
combines impact graph generation, temporal context management, selection and execution through a hierarchical PPO-DDPG frame-
and relational reasoning, showing superior performance in detecting work. Its phased alternating training ensures coordinated learning
market crashes across multiple historical crisis periods. between controllers, achieving superior performance across both
bullish and bearish markets while maintaining portfolio diversifica-
7.2 Execution Optimization tion.
Execution optimization has evolved toward hierarchical frameworks
that decompose complex trading decisions. MacMic [Niu et al., 7.3 High-Frequency Trading
2024a] introduces a two-level architecture where a high-level agent High-frequency trading systems have advanced through specialized
handles volume scheduling while a low-level agent manages precise architectures addressing market microstructure and computational
order placement. The framework’s stacking Hidden Markov Model efficiency. EarnHFT [Qin et al., 2024] introduces a three-stage
enables unsupervised extraction of multi-granular market represen- framework combining Q-learning with specialized agent pools, ef-
tations, achieving superior price execution across US and Chinese fectively handling extended trading trajectories while maintaining
 Table 11: Summary of Recent Contributions in Quantitative Trading

Work Key Innovation Methodology Primary Results

StockFormer Predictive coding with multi- Three-branch transformer for Returns: +40.3%, SR: +22.7%
aspect modeling market dynamics vs SAC
MacMic Order execution decomposition Hierarchical RL with stacking Superior execution across
HMM US/CN markets
IMM Multi-price market making Adversarial state representation Lower adverse selection in fu-
strategy learning tures
EarnHFT Adaptive crypto trading system Hierarchical Q-learning with +30% profit vs SOTA in crypto
router
MacroHFT Market regime-based decompo- Memory-enhanced policy inte- Superior risk-adjusted crypto
sition gration returns
DRPO Direct portfolio weight opti- Probabilistic state decomposi- 415ms latency in production
mization tion
CPPI-MADDPG Portfolio insurance integration Multi-agent RL with protection AR: 9.68%, SR: 2.18 in SZSE
strategies
HRT Integrated selection-execution PPO-DDPG hierarchy for trad- Improved diversification in
ing S&P500
TRR LLM-based crash detection Four-phase temporal reasoning Superior detection of market
crashes

30% higher profitability across market conditions. texts. This specialization-generalization trade-off suggests that op-
MacroHFT [Zong et al., 2024] employs a two-phase approach timal deployment strategies might involve ensemble approaches tai-
with specialized sub-agents for different market regimes, using a lored to specific market conditions and trading objectives.
memory-enhanced hyper-agent for rapid adaptation. DRPO [Han et
al., 2023] achieves practical deployment success through state space
decomposition and probabilistic dynamic programming, maintain- 8 Knowledge Retrieval and Augmentation
ing 415ms latency in production environments. Knowledge retrieval and augmentation in financial markets involves
CPPI-MADDPG [Zhang et al., 2024b] integrates classical port- extracting, processing, and leveraging structured and unstructured
folio insurance with multi-agent reinforcement learning, achiev- information for improved decision-making. This section focuses on
ing 9.68% annual returns while maintaining downside protection discussing three primary directions: financial information retrieval
through adaptive strategy selection. pipelines that processes complex financial texts, intelligent report
generation that synthesizes multiple data sources, and agent-based
7.4 Comparative Analysis market simulation that enables scenario analysis. As summarized
in Table 13, these works range from specialized architectures for
The surveyed approaches demonstrate distinct characteristics in financial text processing to comprehensive frameworks for market
their architectural design, computational requirements, and market simulation.
applicability. Table 12 summarizes key aspects across different
frameworks:
The architectural approaches demonstrate clear trade-offs be- 8.1 Information Retrieval
tween modeling sophistication and practical deployment constraints. Information retrieval (IR) focused on processing complex financial
Predictive frameworks like StockFormer achieve superior perfor- documents through specialized models. MACK [Huang et al., 2024]
mance through comprehensive market modeling but require signif- established a comprehensive framework for Chinese financial event
icant computational resources. In contrast, execution-focused sys- extraction through matrix chunking. Unlike previous approaches
tems like DRPO maintain efficiency through focused optimization requiring pre-identified entities, MACK processes raw text directly
while sacrificing broader strategy integration. through a two-dimensional annotation method that visualizes com-
Furthermore, compute requirements vary significantly, from ponent interactions. When evaluated on the FINEED dataset with
DRPO’s production-ready 415ms latency to TRR’s resource- 5,000 annotated events, this architecture achieved 81.33% F1-score
intensive LLM processing. This latency-sophistication trade-off in event extraction while maintaining 96.89% accuracy in word seg-
critically influences deployment scenarios, with HFT systems pri- mentation.
oritizing speed while signal generation approaches emphasize mod- The emergence of large language models has enabled new ap-
eling capacity. proaches to financial annotation. LLM-Annotator [Aguda et al.,
Market coverage and generalization capabilities also differ sub- 2024] demonstrated that models like GPT-4 and PaLM 2 can out-
stantially. Models like StockFormer and TRR demonstrate cross- perform crowdworkers by 29% in accuracy while maintaining cost
market applicability, while specialized frameworks like IMM and efficiency. Using the REFinD dataset with 28,676 SEC filing re-
EarnHFT show superior performance within specific market con- lations, the framework’s reliability index successfully identified in-
 Table 12: Comparison of Quantitative Trading Approaches

Model Strategy Execution Market Compute Key Strengths Key Limitations

StockFormer Predictive Indirect Multi-asset High Robust pattern extraction; High compute overhead;
Volatile market performance Complex training
MacMic Execution Direct Equities Medium Superior price execution; Limited signal generation;
Hierarchical control Single market focus
IMM Market Making Direct Futures Medium Low adverse selection; Market-specific design;
Queue priority preservation Limited asset coverage
EarnHFT HFT Direct Crypto High Strong profitability; Crypto-specific;
Adaptive routing Resource intensive
MacroHFT HFT Direct Crypto High Regime adaptation; Complex training process;
Memory-enhanced decisions Market-specific
DRPO HFT Direct Multi-asset Low Production-ready latency; Limited strategy scope;
Efficient computation Basic signals
CPPI-MADDPG Portfolio Indirect Equities Medium Downside protection; Coordination overhead;
Strategy integration Slower adaptation
HRT Hybrid Direct Equities High Balanced execution; Complex training;
Portfolio diversification High resource usage
TRR Signal Indirect Multi-asset Very High Novel signal sources; High latency;
Crisis detection Resource intensive

Table 13: Summary of Recent Contributions in Knowledge Retrieval and Augmentation

Work Key Innovation Methodology Primary Results
MACK Matrix chunking for financial Two-dimensional annotation F1: 81.33% (Event), 96.89%
event extraction method (Word)
FinReport Automated investment analy- Three-module system with Return: 57.76%, Accuracy:
sis news factorization 75.40%
LLM-Annotator LLM-based financial relation Reliability index for annota- 29% improvement over
extraction tion quality crowdworkers
TRR Temporal reasoning for crash Four-phase news analysis Superior crisis detection
detection framework across multiple periods
StockAgent Multi-LLM trading simula- Dynamic agent behavior Quantified impact of informa-
tion modeling tion sources
EconAgent LLM-powered economic sim- Perception-memory modules Reproduction of key eco-
ulation for agents nomic phenomena

stances requiring expert review, enabling reliable automation of ap- crashes across multiple historical crises, including the 2007 financial
proximately 65% of annotation tasks. crisis and 2020 COVID-19 crash.

8.2 Report Generation 8.3 Agentic Simulation

The synthesis of multiple information sources for automated analy- The complexity of market interactions has motivated the develop-
sis has advanced through sophisticated architectures. FinReport [Li ment of sophisticated simulation frameworks. StockAgent [Zhang
et al., 2024d] pioneered this direction through a three-module sys- et al., 2024a] established a comprehensive multi-agent system using
tem combining news factorization, return forecasting, and risk as- different LLMs to simulate realistic trading behaviors. The frame-
sessment. The framework integrates semantic role labeling with work incorporates external factors including macroeconomic condi-
dependency parsing for news understanding, while employing en- tions and market sentiment, revealing distinct trading patterns be-
hanced Fama-French models for return prediction. This approach tween models like GPT-3.5 and Gemini Pro. Experimental results
achieved 75.40% accuracy in news factorization and 57.76% annu- quantified the impact of different information sources, with interest
alized returns in backtesting. rate information significantly affecting trading frequency.
TRR [Koa et al., 2024b] advanced the field through temporal re- EconAgent [Li et al., 2024a] advanced this direction through
lational reasoning for market event detection. The framework’s four- LLM-powered agents for macroeconomic simulation. The frame-
phase approach combines impact chain generation, temporal context work’s innovation lies in its perception and memory modules that
management, and PageRank-based attention mechanisms. This ar- enable heterogeneous decision-making while considering historical
chitecture demonstrated particular effectiveness in detecting market market dynamics. This approach successfully reproduced key eco-
nomic phenomena like the Phillips Curve and demonstrated effec- aligned news records. The dataset covers 4,775 S&P500 compa-
tiveness in simulating crisis impacts, particularly during the COVID- nies from 1999-2023, demonstrating that increased dataset scale sig-
19 period. nificantly improves prediction accuracy. Experiments with various
deep learning architectures revealed that transformer-based mod-
8.4 Comparative Analysis els achieve optimal performance (R² = 0.988) when leveraging the
dataset’s full scale, while sentiment information provides modest but
Table ?? summarizes key aspects of representative models The three consistent improvements in prediction accuracy.
research directions in knowledge retrieval and augmentation offer
complementary approaches to market understanding, each with dis-
tinct theoretical limitations. Information extraction methods face 9.3 Sentiment and Emotion
fundamental challenges in handling context-dependent financial ter- AlphaFin [Li et al., 2024c] addressed the interpretability gap in
minology and evolving market jargon. While MACK demonstrates financial analysis through a dataset combining traditional market
strong performance on Chinese texts, the generalization to multiple data with chain-of-thought annotations. The framework’s integra-
languages and domains remains challenging due to linguistic com- tion with retrieval-augmented generation enables real-time mar-
plexity and domain-specific variations. ket awareness while maintaining reasoning capabilities, achieving
Report generation frameworks encounter theoretical limitations 30.8% annualized returns in experimental validation. StockEmo-
in causal reasoning and temporal dependency modeling. Although tions [Lee et al., 2023] complemented this direction through fine-
FinReport successfully combines multiple information sources, es- grained emotion analysis, providing 10,000 annotated StockTwits
tablishing robust causal relationships between news events and mar- comments across 12 emotion classes. The dataset’s multi-step anno-
ket movements remains an open challenge. The inherent delay be- tation pipeline, combining pre-trained language models with expert
tween information release and market impact further complicates validation, demonstrated that incorporating emotional features sig-
temporal modeling. nificantly improves market prediction accuracy compared to purely
Simulation approaches face fundamental limitations in agent be- numerical approaches.
havior modeling and market complexity. While StockAgent and
EconAgent demonstrate impressive capabilities, they necessarily 9.4 Comparative Analysis
simplify market microstructure and agent interactions. The chal-
These financial datasets exhibit distinct characteristics in their cov-
lenge of calibrating agent behaviors to realistic market conditions
erage, scale, and applicability to different financial tasks. As
while maintaining computational tractability represents a key area
shown in Table 15, coverage ranges from focused sentiment anal-
for future research.
ysis (StockEmotions) to comprehensive market-wide data integra-
tion (FNSPID). Scale varies significantly, with FNSPID providing
9 Financial Datasets millions of aligned price-news records while StockEmotions offers
deeper but narrower emotional annotation.
Financial datasets play a crucial role in advancing artificial intelli- Market coverage and temporal resolution reveal complementary
gence applications for market analysis and trading. Recent develop- strengths across datasets. Market-GAN focuses on high-fidelity syn-
ments in dataset creation have focused on three primary directions: thetic data for a single major index (DJIA) with emphasis on tempo-
synthetic data generation for enhanced model training, multimodal ral consistency. In contrast, FNSPID provides broader market cov-
integration combining market data with textual information, and erage across the S&P500 with emphasis on news-price alignment.
specialized datasets for reasoning and sentiment analysis. Unlike AlphaFin bridges multiple markets while prioritizing reasoning an-
traditional financial datasets that focus solely on price and volume notation quality, and StockEmotions offers specialized coverage of
information, modern datasets incorporate diverse data types includ- retail investor sentiment through StockTwits data.
ing news sentiment, investor emotions, and chain-of-thought reason-
ing annotations. As summarized in Table 14, these works contribute
to both the creation of novel datasets and the development of sophis- 10 Time Series Models
ticated data generation techniques, addressing critical challenges in
data availability, quality, and comprehensiveness for financial appli- Time series modeling for financial applications focuses on three pri-
cations. mary directions: foundation models that leverage large-scale pre-
training, efficient architectures that handle temporal complexity, and
unified frameworks that address multiple time series tasks. Unlike
9.1 Synthetic Data Generation traditional statistical approaches, modern time series models incor-
Market-GAN [Xia et al., 2024a] established a framework for gener- porate transformer architectures, multiscale analysis, and transfer
ating high-fidelity financial data with controllable semantic context learning capabilities. As summarized in Table 16, these foundation
through a novel two-stage training approach. The framework com- models are trained on massive datasets, specialized architectures
bines GAN architecture with autoencoder and supervisory networks for temporal modeling, and unified frameworks supporting multiple
to maintain data fidelity while ensuring alignment with given market tasks.
contexts. Its C-TimesBlock innovation effectively captures temporal
dependencies while preventing mode collapse, a common challenge 10.1 Foundation Models
in financial time series generation. When evaluated on Dow Jones
Industrial Average data spanning 2000-2023, this approach demon- The emergence of foundation models has transformed time se-
strated superior performance in context alignment and downstream ries analysis through large-scale pre-training and transfer learning.
task fidelity compared to existing generation methods. Timer established a comprehensive framework using a decoder-
only transformer architecture trained on 1 billion time points. The
model’s innovation lies in its unified sequence format that han-
9.2 Multimodal Fusion dles heterogeneous time series data while enabling few-shot learn-
FNSPID [Dong et al., 2024b] pioneered large-scale integration of ing capabilities. When evaluated across forecasting, imputation,
quantitative and qualitative financial data through a comprehen- and anomaly detection tasks, this architecture achieved state-of-the-
sive dataset combining 29.7 million stock prices with 15.7 million art performance using only 1-5MOMENT advanced this direction
 Table 14: Summary of Recent Contributions in Financial Datasets
Dataset Key Innovation Data Characteristics Primary Results
Market-GAN Controllable synthetic data Two-stage GAN with C-TimesBlock Superior fidelity in DJIA sim-
with semantic context for temporal consistency ulation (2000-2023)
FNSPID Large-scale price-news inte- 29.7M prices, 15.7M news records Transformer accuracy R² =
gration framework for 4,775 companies 0.988, Reproducible updates
AlphaFin Chain-of-thought financial Market data with expert reasoning 30.8% annualized returns,
reasoning annotations Enhanced interpretability
StockEmotions Fine-grained investor psy- 10k annotated comments, 12 emo- Improved prediction with
chology tion classes, emoji features emotional features

Table 15: Comparison of Financial Dataset Characteristics

Dataset Time Text Market Scale Primary

Series Data Coverage Application
Market-GAN ✓ - DJIA Medium Market Simulation
FNSPID ✓ ✓ S&P500 Very Large Predictive Modeling
AlphaFin ✓ ✓ Multiple Large Interpretable Analysis
StockEmotions ✓ ✓ StockTwits Medium Sentiment Analysis

through an open-source family of foundation models for general- 10.4 Discussion

purpose time series analysis. The framework addresses key chal-
lenges in multi-dataset training through specialized techniques for The three research directions in time series modeling offer comple-
handling varying sampling rates and amplitudes. Its contribution mentary approaches to temporal analysis, each with distinct theo-
of Time Series Pile, a diverse collection of public datasets, en- retical limitations. Table 17 summarizes key aspects across differ-
ables robust pre-training while maintaining reproducibility. This ap- ent frameworks. Foundation models face fundamental challenges in
proach demonstrated particular strength in limited-supervision set- capturing rare temporal patterns and handling non-stationary distri-
tings across financial applications. butions, though they excel at extracting general temporal dependen-
cies. While large-scale pre-training improves robustness, the inher-
ent challenge of temporal causality and regime changes remains.
10.2 Efficient Architectures
Efficient architectures encounter trade-offs between model capac-
The complexity of temporal patterns has motivated development of ity and computational complexity. While patch-based approaches
specialized architectures balancing modeling capacity with compu- and multiscale decomposition provide practical solutions, they nec-
tational efficiency. TimeMixer pioneered this direction through a essarily introduce approximations in temporal modeling. The bal-
multiscale MLP-based architecture that decomposes temporal vari- ance between local and global temporal dependencies remains a key
ations at different sampling scales. The model employs Past- theoretical challenge.
Decomposable-Mixing blocks for separate processing of seasonal Unified frameworks face fundamental limitations in optimal pa-
and trend components, while Future-Multipredictor-Mixing blocks rameter sharing across diverse tasks. While architectures like Times-
combine predictions across scales. This architecture demonstrated Net demonstrate impressive multi-task capability, the theoretical
state-of-the-art performance across 18 benchmarks while maintain- foundations for optimal architecture design across different temporal
ing computational efficiency. PatchTST advanced efficient process- modeling objectives remain incomplete. These limitations suggest
ing through a patch-based approach to time series modeling. The future research directions in developing more theoretically grounded
framework’s key innovations lie in subseries-level patching for lo- approaches to temporal modeling while maintaining practical appli-
cal information capture and channel-independent processing for uni- cability.
variate sequences. This architecture achieved 21% reduction in The integration of domain knowledge remains crucial for devel-
Mean Squared Error for long-term forecasting while effectively han- oping more robust and interpretable temporal models, particularly in
dling extended historical sequences without memory constraints. financial applications where model reliability and theoretical sound-
ness are paramount.
10.3 Unified Frameworks
The diverse requirements of time series analysis have inspired 11 Open Challenges
development of unified frameworks supporting multiple tasks.
TimesNet established a comprehensive architecture through multi- Having comprehensively reviewed the recent works on Financial Ai,
periodicity analysis that transforms one-dimensional sequences into we identify several promising research directions that warrant fur-
two-dimensional representations. The framework captures both in- ther investigation. In this section, We organize these opportunities
traperiod and interperiod variations through parameter-efficient in- into architectural innovations, methodological advancements, and
ception blocks, demonstrating superior performance across forecast- practical considerations to potentially extend the impact to further
ing, imputation, classification, and anomaly detection tasks. advancement of the field.
 Table 16: Summary of Recent Contributions in Time Series Models
Model Key Innovation Methodology Primary Results
Timer Pre-trained decoder-only Unified sequence format for Strong few-shot performance
transformer heterogeneous data with 1-5% data
MOMENT Open-source foundation Multi-dataset training strate- Superior limited-supervision
model family gies performance
TimeMixer Multiscale MLP architecture Decomposable mixing blocks SOTA across 18 benchmarks
for temporal patterns
TimesNet Multi-periodicity analysis 1D to 2D transformation for Unified performance across 5
temporal patterns major tasks
PatchTST Patch-based time series pro- Channel-independent trans- 21% MSE reduction in long-
cessing former term forecasting

Table 17: Comparison of Time Series Model Characteristics

Model Pre-training Multi-task Compute Scalability Primary Application

Timer ✓ ✓ High Linear Few-shot Learning
MOMENT ✓ ✓ High Linear Limited Supervision
TimeMixer - - Medium Sublinear Multiscale Patterns
TimesNet - ✓ Medium Linear Task Flexibility
PatchTST - - Low Sublinear Long Sequences

11.1 Architectural Innovations particular attention. Current approaches [Hui and Wang, 2024;
The evolution of financial AI architectures suggests several key di- Yamagata and Ono, 2024] often rely on simple volatility estimates
rections for improvement. Foundation models pre-trained on mas- or Sharpe ratios, failing to capture tail risks and regime changes ade-
sive financial datasets show promise [Liu et al., 2024b; Goswami et quately. Research into neural architectures specifically designed for
al., 2024] but currently lack domain-specific inductive biases cru- extreme value theory and systemic risk modeling could significantly
cial for finance. Future research should explore specialized pre- improve portfolio management and risk assessment.
training objectives that incorporate market microstructure theory
and regulatory constraints. The development of modular architec- 11.3 Multimodal Foundation Models
tures that combine pre-trained components with task-specific adap- The emergence of multimodal foundation models presents excit-
tors could enable more efficient transfer learning while maintaining ing opportunities for financial applications. While current work [Li
model interpretability. Multi-agent architectures for portfolio opti- et al., 2024c; Zhang et al., 2024a] demonstrates promise in com-
mization and market making demonstrate strong performance [Li et bining textual and numerical data, future research should explore
al., 2024e; Zhang et al., 2024b] but face challenges in convergence more sophisticated fusion mechanisms. This includes developing
guarantees and Nash equilibrium stability. Research into theoretical specialized architectures for processing earnings calls, satellite im-
frameworks for multi-agent learning under market non-stationarity agery, and alternative data sources simultaneously. The adaptation
could lead to more robust trading systems. The integration of market of foundation models for real-time market analysis remains chal-
impact models into agent training frameworks, building upon work lenging. Research into streaming architectures that can efficiently
like [Niu et al., 2024a], remains crucial for bridging the gap between update their knowledge base and adapt to new market conditions
simulation and real-world deployment. could enable more responsive trading systems. Additionally, explor-
ing techniques for maintaining temporal coherence across different
11.2 Methodological Advancements data modalities could improve prediction accuracy during market
Several methodological challenges require attention from the re- regime changes.
search community. The treatment of temporal dependencies in fi-
nancial data remains suboptimal, with current approaches [Wang 11.4 Hardware-Accelerated Trading
et al., 2024a; Nie et al., 2023] struggling to capture long-range The advancement of hardware-accelerated trading systems presents
dependencies while maintaining computational efficiency. Future several research opportunities. Current approaches [Han et al.,
work should explore adaptive attention mechanisms that dynami- 2023; Qin et al., 2024] demonstrate promising potential for low-
cally adjust their receptive field based on market conditions. The latency trading. Research into specialized neural network archi-
integration of classical financial theory with deep learning frame- tectures optimized for FPGA implementation could further reduce
works presents another promising direction. While works like [Xu latency while maintaining model sophistication. The development
et al., 2024] incorporate GARCH models and factor analysis, a more of hardware-aware training algorithms represents another promis-
systematic approach to theory-guided neural network design could ing direction. Future work should explore techniques for jointly
improve both performance and interpretability. This includes devel- optimizing model architecture and hardware implementation, poten-
oping loss functions that explicitly encode financial principles and tially leading to more efficient high-frequency trading systems. This
constraints. Risk modeling in deep learning frameworks requires includes developing specialized attention mechanisms and network
pruning techniques that consider hardware constraints during train- could help bridge the gap between academic innovation and practi-
ing. The integration of hardware acceleration with market making cal deployment. This includes developing standardized interfaces
systems presents unique challenges. Research into architectures that between trading systems and regulatory monitoring frameworks.
can efficiently process order book data and execute trades with mi- These practical considerations highlight the need for research that
crosecond latency could significantly improve market liquidity. This explicitly addresses production deployment challenges while main-
includes developing specialized circuits for order matching and risk taining the theoretical rigor expected in academic work. Success in
calculation that maintain accuracy while minimizing latency. these areas could accelerate the adoption of AI innovations in real-
world financial systems.
11.5 Research-Industry Collaborations
A notable limitation in current research is the absence of real-world 12 Conclusion
deployment studies and industry validation. While works like [Han
et al., 2023] report production latency metrics, and [Qin et al., 2024; This survey has systematically analyzed recent advancements in ar-
Zong et al., 2024] demonstrate strong backtesting performance, tificial intelligence for financial applications, revealing significant
none of the surveyed papers provide comprehensive evidence of suc- progress across predictive modeling, decision-making, and knowl-
cessful industrial deployment. This gap manifests in several critical edge retrieval tasks. The examined innovations span architectural
aspects: First, most research evaluations rely solely on historical developments in foundation models, specialized network designs for
data backtesting, which fails to capture crucial real-world challenges temporal and relational learning, and practical frameworks for pro-
such as market impact, execution slippage, and microstructure ef- duction deployment. While these advances demonstrate substantial
fects. The lack of live trading results or paper trading validation improvements in model performance and capability, several crucial
raises questions about the practical viability of proposed methods. challenges remain unresolved.
Second, computational requirements and system integration receive The theoretical foundations for financial AI systems still require
limited attention. While papers like [Han et al., 2023] discuss la- strengthening, particularly in establishing convergence guarantees
tency constraints, few address critical production concerns such as for multi-agent trading systems and developing robust frameworks
system reliability, fault tolerance, and integration with existing trad- for handling market non-stationarity. The integration of classical fi-
ing infrastructure. The absence of deployment case studies or per- nancial theory with deep learning architectures presents another crit-
formance analysis under real market conditions represents a signif- ical area for development, as current approaches often fail to fully
icant limitation in current research. Third, regulatory compliance incorporate market microstructure effects and regulatory constraints.
and risk management frameworks remain largely theoretical. De- Production deployment remains a significant challenge, with few
spite works like [Li et al., 2024d] addressing interpretability, none studies demonstrating successful real-world implementation or com-
of the surveyed papers demonstrate compliance with specific reg- prehensive compliance with regulatory requirements. The gap be-
ulatory requirements or integration with institutional risk manage- tween academic innovation and industrial application suggests the
ment systems. This gap between academic innovation and regula- need for closer collaboration between researchers and practitioners,
tory reality hinders industrial adoption. Future research would ben- particularly in developing standardized evaluation frameworks that
efit significantly from collaborations between academic institutions better approximate real-world trading conditions.
and financial firms to validate proposed methods under real market Future research directions should focus on developing more ro-
conditions. Studies documenting deployment challenges, system ar- bust theoretical frameworks for financial AI, improving the integra-
chitecture decisions, and practical performance metrics would pro- tion of domain knowledge in model architecture design, and ad-
vide valuable insights for both researchers and practitioners. Addi- dressing the practical challenges of system deployment and regu-
tionally, research into robust evaluation frameworks that better ap- latory compliance. The advancement of hardware-accelerated trad-
proximate real-world trading conditions could help bridge the gap ing systems and the development of multimodal foundation models
between academic research and industrial applications. present particularly promising opportunities for innovation. Success
in these areas could significantly advance the field while improving
11.6 Practical Considerations the reliability and effectiveness of AI-driven financial systems.
Several critical implementation challenges require focused research
attention. While synthetic data generation techniques [Xia et al., References
2024a] help address data limitations, research is needed on frame-
works that can rigorously validate model robustness under real mar- [Aguda et al., 2024] Toyin Aguda, Suchetha Siddagangappa, Elena
ket conditions. This includes developing standardized stress test- Kochkina, Simerjot Kaur, Dongsheng Wang, Charese Smiley,
ing methodologies that simulate market crises, liquidity shocks, and and Sameena Shah. Large language models as financial data an-
extreme events. System resilience and failsafe mechanisms repre- notators: A study on effectiveness and efficiency. arXiv preprint
sent another crucial area. Research into graceful degradation strate- arXiv:2403.18152, 2024.
gies, automated circuit breakers, and robust fallback mechanisms [Chen et al., 2024] Weijun Chen, Shun Li, Xipu Yu, Heyuan Wang,
could improve the safety of automated trading systems. This in- Wei Chen, and Tengjiao Wang. Automatic de-biased temporal-
cludes developing methods to automatically detect and respond to relational modeling for stock investment recommendation. In
anomalous market conditions or model behavior. The challenge Kate Larson, editor, Proceedings of the Thirty-Third Interna-
of continuous learning and model updating in production environ- tional Joint Conference on Artificial Intelligence, IJCAI-24,
ments remains largely unexplored. Future research should investi- pages 1999–2008. International Joint Conferences on Artificial
gate techniques for safely updating models in live trading systems, Intelligence Organization, 8 2024. Main Track.
including methods for gradual deployment, A/B testing, and per-
formance monitoring. This includes developing frameworks for de- [Daiya et al., 2024] Divyanshu Daiya, Monika Yadav, and
tecting model degradation and safely reverting changes if necessary. Harshit Singh Rao. Diffstock: Probabilistic relational stock mar-
Cross-jurisdictional compliance presents another significant chal- ket predictions using diffusion models. In ICASSP 2024-2024
lenge, particularly for global trading operations. Research into au- IEEE International Conference on Acoustics, Speech and Signal
tomated compliance verification and real-time regulatory reporting Processing (ICASSP), pages 7335–7339. IEEE, 2024.
[Ding et al., 2024a] Han Ding, Yinheng Li, Junhao Wang, and [Lee et al., 2023] Jean Lee, Hoyoul Luis Youn, Josiah Poon, and
Hang Chen. Large language model agent in financial trading: Soyeon Caren Han. Stockemotions: Discover investor emo-
A survey. arXiv preprint arXiv:2408.06361, 2024. tions for financial sentiment analysis and multivariate time series.
[Ding et al., 2024b] Wei Ding, Zhennan Chen, Hanpeng Jiang, arXiv preprint arXiv:2301.09279, 2023.
Yuanguo Lin, and Fan Lin. Trend-heuristic reinforcement learn- [Lee et al., 2024] Jean Lee, Nicholas Stevens, Soyeon Caren Han,
ing framework for news-oriented stock portfolio management. and Minseok Song. A survey of large language models in finance
In ICASSP 2024-2024 IEEE International Conference on Acous- (finllms). arXiv preprint arXiv:2402.02315, 2024.
tics, Speech and Signal Processing (ICASSP), pages 5120–5124.
[Li et al., 2023] Yinheng Li, Shaofei Wang, Han Ding, and Hang
IEEE, 2024.
Chen. Large language models in finance: A survey. In Proceed-
[Dong et al., 2024a] Michael Ming Dong, Theophanis C Stratopou- ings of the fourth ACM international conference on AI in finance,
los, and Victor Xiaoqi Wang. A scoping review of chatgpt re- pages 374–382, 2023.
search in accounting and finance. SSRN, 2024.
[Li et al., 2024a] Nian Li, Chen Gao, Mingyu Li, Yong Li, and
[Dong et al., 2024b] Zihan Dong, Xinyu Fan, and Zhiyuan Peng. Qingmin Liao. Econagent: large language model-empowered
Fnspid: A comprehensive financial news dataset in time series. agents for simulating macroeconomic activities. In Proceedings
In Proceedings of the 30th ACM SIGKDD Conference on Knowl- of the 62nd Annual Meeting of the Association for Computational
edge Discovery and Data Mining, pages 4918–4927, 2024. Linguistics (Volume 1: Long Papers), pages 15523–15536, 2024.
[Gao et al., 2023] Siyu Gao, Yunbo Wang, and Xiaokang Yang.
[Li et al., 2024b] Tong Li, Zhaoyang Liu, Yanyan Shen, Xue Wang,
Stockformer: Learning hybrid trading machines with predictive
Haokun Chen, and Sen Huang. Master: Market-guided stock
coding. In IJCAI, pages 4766–4774, 2023.
transformer for stock price forecasting. In Proceedings of the
[Goswami et al., 2024] Mononito Goswami, Konrad Szafer, Arjun AAAI Conference on Artificial Intelligence, volume 38, pages
Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. Moment: A 162–170, 2024.
family of open time-series foundation models. In International
[Li et al., 2024c] Xiang Li, Zhenyu Li, Chen Shi, Yong Xu, Qing
Conference on Machine Learning, 2024.
Du, Mingkui Tan, Jun Huang, and Wei Lin. Alphafin: Bench-
[Han et al., 2023] Li Han, Nan Ding, Guoxuan Wang, Dawei marking financial analysis with retrieval-augmented stock-chain
Cheng, and Yuqi Liang. Efficient continuous space policy op- framework. arXiv preprint arXiv:2403.12582, 2024.
timization for high-frequency trading. In Proceedings of the 29th
ACM SIGKDD Conference on Knowledge Discovery and Data [Li et al., 2024d] Xiangyu Li, Xinjie Shen, Yawen Zeng, Xiaofen
Mining, pages 4112–4122, 2023. Xing, and Jin Xu. Finreport: Explainable stock earnings forecast-
ing via news factor analyzing model. In Companion Proceedings
[Huang et al., 2024] Yusheng Huang, Ning Hu, Kunping Li, Nan of the ACM on Web Conference 2024, pages 319–327, 2024.
Wang, and Zhouhan Lin. Extracting financial events from raw
texts via matrix chunking. In Proceedings of the 2024 Joint In- [Li et al., 2024e] Zhenglong Li, Vincent Tam, and Kwan L. Ye-
ternational Conference on Computational Linguistics, Language ung. Developing a multi-agent and self-adaptive framework with
Resources and Evaluation (LREC-COLING 2024), pages 7035– deep reinforcement learning for dynamic portfolio risk manage-
7044, 2024. ment. In Proceedings of the 23rd International Conference on
Autonomous Agents and Multiagent Systems, AAMAS ’24, page
[Hui and Wang, 2024] Qian Hui and Tiandong Wang. Mitigat-
1174–1182, Richland, SC, 2024. International Foundation for
ing extremal risks: A network-based portfolio strategy. arXiv Autonomous Agents and Multiagent Systems.
preprint arXiv:2409.12208, 2024.
[Liu et al., 2024a] Mengpu Liu, Mengying Zhu, Xiuyuan Wang,
[Jain et al., 2024] Samyak Jain, Parth Chhabra, Atula Tejaswi
Guofang Ma, Jianwei Yin, and Xiaolin Zheng. Echo-gl: Earnings
Neerkaje, Puneet Mathur, Ramit Sawhney, Shivam Agarwal,
calls-driven heterogeneous graph learning for stock movement
Preslav Nakov, Sudheer Chava, and Dinesh Manocha. Saliency-
prediction. In Proceedings of the AAAI Conference on Artificial
aware interpolative augmentation for multimodal financial pre-
Intelligence, volume 38, pages 13972–13980, 2024.
diction. In Proceedings of the 2024 Joint International Con-
ference on Computational Linguistics, Language Resources and [Liu et al., 2024b] Yong Liu, Haoran Zhang, Chenyu Li, Xiang-
Evaluation (LREC-COLING 2024), pages 14285–14297, 2024. dong Huang, Jianmin Wang, and Mingsheng Long. Timer: Gen-
[Jeon et al., 2024] Jihyeong Jeon, Jiwon Park, Chanhee Park, and erative pre-trained transformers are large time series models. In
U Kang. Frequant: A reinforcement-learning based adaptive Forty-first International Conference on Machine Learning, 2024.
portfolio optimization with multi-frequency decomposition. In [Nie et al., 2023] Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong,
Proceedings of the 30th ACM SIGKDD Conference on Knowl- and Jayant Kalagnanam. A time series is worth 64 words: Long-
edge Discovery and Data Mining, pages 1211–1221, 2024. term forecasting with transformers. In International Conference
[Koa et al., 2024a] Kelvin JL Koa, Yunshan Ma, Ritchie Ng, and on Learning Representations, 2023.
Tat-Seng Chua. Learning to generate explainable stock predic- [Niu et al., 2024a] Hui Niu, Siyuan Li, and Jian Li. Macmic: Ex-
tions using self-reflective large language models. In Proceedings ecuting iceberg orders via hierarchical reinforcement learning.
of the ACM on Web Conference 2024, pages 4304–4315, 2024. In Kate Larson, editor, Proceedings of the Thirty-Third Inter-
[Koa et al., 2024b] Kelvin JL Koa, Yunshan Ma, Ritchie Ng, Huan- national Joint Conference on Artificial Intelligence, IJCAI-24,
huan Zheng, and Tat-Seng Chua. Temporal relational reasoning pages 6008–6016. International Joint Conferences on Artificial
of large language models for detecting stock portfolio crashes. Intelligence Organization, 8 2024. Main Track.
arXiv preprint arXiv:2410.17266, 2024. [Niu et al., 2024b] Hui Niu, Siyuan Li, Jiahao Zheng, Zhouchi Lin,
[Kou et al., 2024] Zhizhuo Kou, Holam Yu, Jingshu Peng, and Lei Bo An, Jian Li, and Jian Guo. Imm: An imitative reinforcement
Chen. Automate strategy finding with llm in quant investment. learning approach with predictive representation learning for au-
arXiv preprint arXiv:2409.06289, 2024. tomatic market making. In Kate Larson, editor, Proceedings of
 the Thirty-Third International Joint Conference on Artificial In- [Zeng et al., 2024] Liang Zeng, Lei Wang, Hui Niu, Ruchen Zhang,
telligence, IJCAI-24, pages 5999–6007. International Joint Con- Ling Wang, and Jian Li. Trade when opportunity comes: Price
ferences on Artificial Intelligence Organization, 8 2024. Main movement forecasting via locality-aware attention and iterative
Track. refinement labeling. In Kate Larson, editor, Proceedings of the
Thirty-Third International Joint Conference on Artificial Intel-
[Qin et al., 2024] Molei Qin, Shuo Sun, Wentao Zhang, Haochong
ligence, IJCAI-24, pages 6134–6142. International Joint Con-
Xia, Xinrun Wang, and Bo An. Earnhft: Efficient hierarchical
ferences on Artificial Intelligence Organization, 8 2024. Main
reinforcement learning for high frequency trading. In Proceed-
Track.
ings of the AAAI Conference on Artificial Intelligence, volume 38,
pages 14669–14676, 2024. [Zhang et al., 2024a] Chong Zhang, Xinyi Liu, Mingyu Jin, Zhong-
mou Zhang, Lingyao Li, Zhengting Wang, Wenyue Hua, Dong
[Wang and Ma, 2024] Mengyu Wang and Tiejun Ma. Mana- Shu, Suiyuan Zhu, Xiaobo Jin, et al. When ai meets finance
net: Mitigating aggregated sentiment homogenization with news (stockagent): Large language model-based stock trading in simu-
weighting for enhanced market prediction. In Proceedings of the lated real-world environments. arXiv preprint arXiv:2407.18957,
33rd ACM International Conference on Information and Knowl- 2024.
edge Management, pages 2379–2389, 2024.
[Zhang et al., 2024b] Hengxi Zhang, Zhendong Shi, Yuanquan Hu,
[Wang et al., 2024a] Shiyu Wang, Haixu Wu, Xiaoming Shi, Wenbo Ding, Ercan E Kuruoğlu, and Xiao-Ping Zhang. Optimiz-
Tengge Hu, Huakun Luo, Lintao Ma, James Y Zhang, and JUN ing trading strategies in quantitative markets using multi-agent
ZHOU. Timemixer: Decomposable multiscale mixing for time reinforcement learning. In ICASSP 2024-2024 IEEE Interna-
series forecasting. In International Conference on Learning Rep- tional Conference on Acoustics, Speech and Signal Processing
resentations (ICLR), 2024. (ICASSP), pages 136–140. IEEE, 2024.
[Wang et al., 2024b] Xinlei Wang, Maike Feng, Jing Qiu, Jinjin [Zhang et al., 2024c] Wentao Zhang, Lingxuan Zhao, Haochong
Gu, and Junhua Zhao. From news to forecast: Integrating event Xia, Shuo Sun, Jiaze Sun, Molei Qin, Xinyi Li, Yuqing Zhao,
analysis in llm-based time series forecasting with reflection. Ad- Yilei Zhao, Xinyu Cai, et al. A multimodal foundation agent for
vances in Neural Information Processing Systems, 2024. financial trading: Tool-augmented, diversified, and generalist. In
Proceedings of the 30th ACM SIGKDD Conference on Knowl-
[Xia et al., 2024a] Haochong Xia, Shuo Sun, Xinrun Wang, and
edge Discovery and Data Mining, pages 4314–4325, 2024.
Bo An. Market-gan: Adding control to financial market data
generation with semantic context. In Proceedings of the AAAI [Zhang et al., 2024d] Wentao Zhang, Yilei Zhao, Shuo Sun, Jie
Conference on Artificial Intelligence, volume 38, pages 15996– Ying, Yonggang Xie, Zitao Song, Xinrun Wang, and Bo An.
16004, 2024. Reinforcement learning with maskable stock representation for
portfolio management in customizable stock pools. In Proceed-
[Xia et al., 2024b] Hongjie Xia, Huijie Ao, Long Li, Yu Liu, Sen ings of the ACM on Web Conference 2024, pages 187–198, 2024.
Liu, Guangnan Ye, and Hongfeng Chai. Ci-sthpan: Pre-trained
attention network for stock selection with channel-independent [Zhao and Welsch, 2024] Zijie Zhao and Roy E Welsch. Hierarchi-
spatio-temporal hypergraph. In Proceedings of the AAAI Con- cal reinforced trader (hrt): A bi-level approach for optimizing
ference on Artificial Intelligence, volume 38, pages 9187–9195, stock selection and execution. arXiv preprint arXiv:2410.14927,
2024. 2024.
[Xiang et al., 2024] Quanzhou Xiang, Zhan Chen, Qi Sun, and Ru- [Zhao et al., 2022] Yu Zhao, Huaming Du, Ying Liu, Shaopeng
jun Jiang. Rsap-dfm: Regime-shifting adaptive posterior dy- Wei, Xingyan Chen, Fuzhen Zhuang, Qing Li, and Gang Kou.
namic factor model for stock returns prediction. In Kate Larson, Stock movement prediction based on bi-typed hybrid-relational
editor, Proceedings of the Thirty-Third International Joint Con- market knowledge graph via dual attention networks. IEEE
ference on Artificial Intelligence, IJCAI-24, pages 6116–6124. Transactions on Knowledge and Data Engineering, 35(8):8559–
International Joint Conferences on Artificial Intelligence Organi- 8571, 2022.
zation, 8 2024. Main Track. [Zhao et al., 2023] Lifan Zhao, Shuming Kong, and Yanyan Shen.
Doubleadapt: A meta-learning approach to incremental learning
[Xu et al., 2024] Zeda Xu, John Liechty, Sebastian Benthall,
for stock trend forecasting. In Proceedings of the 29th ACM
Nicholas Skar-Gislinge, and Christopher McComb. Garch- SIGKDD Conference on Knowledge Discovery and Data Mining,
informed neural networks for volatility prediction in financial pages 3492–3503, 2023.
markets. arXiv preprint arXiv:2410.00288, 2024.
[Zhao et al., 2024] Huaqin Zhao, Zhengliang Liu, Zihao Wu, Yi-
[Yamagata and Ono, 2024] Eisuke Yamagata and Shunsuke Ono. wei Li, Tianze Yang, Peng Shu, Shaochen Xu, Haixing Dai,
Risk-managed sparse index tracking via market graph clustering. Lin Zhao, Gengchen Mai, et al. Revolutionizing finance with
In ICASSP 2024-2024 IEEE International Conference on Acous- llms: An overview of applications and insights. arXiv preprint
tics, Speech and Signal Processing (ICASSP), pages 9796–9800. arXiv:2401.11641, 2024.
IEEE, 2024.
[Zheng et al., 2023] Zetao Zheng, Jie Shao, Jia Zhu, and Heng Tao
[Yan and Tan, 2024] Wenbo Yan and Ying Tan. Double-path Shen. Relational temporal graph convolutional networks for
adaptive-correlation spatial-temporal inverted transformer for ranking-based stock prediction. In 2023 IEEE 39th International
stock time series forecasting. arXiv preprint arXiv:2409.15662, Conference on Data Engineering (ICDE), pages 123–136, 2023.
2024. [Zong et al., 2024] Chuqiao Zong, Chaojie Wang, Molei Qin, Lei
[You et al., 2024] Zinuo You, Pengju Zhang, Jin Zheng, and John Feng, Xinrun Wang, and Bo An. Macrohft: Memory augmented
Cartlidge. Multi-relational graph diffusion neural network with context-aware reinforcement learning on high frequency trading.
parallel retention for stock trends classification. In ICASSP 2024- In Proceedings of the 30th ACM SIGKDD Conference on Knowl-
2024 IEEE International Conference on Acoustics, Speech and edge Discovery and Data Mining, pages 4712–4721, 2024.
Signal Processing (ICASSP), pages 6545–6549. IEEE, 2024.