Paper: arXiv 2403.00782
Authors: Hanshuang Tong, Jun Li, Ning Wu, Ming Gong, Dongmei Zhang, Qi Zhang
Abstract
Recent advancements in large language models (LLMs) have opened new pathways for many domains. However, the full potential of LLMs in financial investments remains largely untapped. There are two main challenges for typical deep learning-based methods for quantitative finance. First, they struggle to fuse textual and numerical information flexibly for stock movement prediction. Second, traditional methods lack clarity and interpretability, which impedes their application in scenarios where the justification for predictions is essential. To solve the above challenges, we propose Ploutos, a novel financial LLM framework that consists of PloutosGen and PloutosGPT. The PloutosGen contains multiple primary experts that can analyze different modal data, such as text and numbers, and provide quantitative strategies from different perspectives. Then PloutosGPT combines their insights and predictions and generates interpretable rationales. To generate accurate and faithful rationales, the training strategy of PloutosGPT leverage rearview-mirror prompting mechanism to guide GPT-4 to generate rationales, and a dynamic token weighting mechanism to finetune LLM by increasing key tokens weight. Extensive experiments show our framework outperforms the state-of-the-art methods on both prediction accuracy and interpretability.
Complexity vs Empirical Score
- Math Complexity: 2.0/10
- Empirical Rigor: 8.5/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper focuses on implementing and integrating existing deep learning and LLM techniques with novel prompting strategies for stock prediction, relying heavily on data processing, model fine-tuning, and backtesting results rather than presenting new theoretical mathematics.
Research Flowchart
flowchart TD A["Research Goal<br>Interpretable Stock Prediction"] --> B["Input Data<br>Financial Text & Numerical Data"] B --> C["PloutosGen<br>Multimodal Expert Fusion"] C --> D["PloutosGPT<br>LLM Rationale Generation"] D --> E["Training Mechanisms<br>Rearview-Mirror Prompting<br>Dynamic Token Weighting"] E --> F["Output<br>Accurate Prediction + Interpretable Rationale"] F --> G["Outcome<br>Outperforms SOTA<br>in Accuracy & Interpretability"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes Ploutos, a novel financial large language model framework comprising PloutosGen for multimodal analysis and PloutosGPT for interpretable stock prediction.
Source sentence
We propose Ploutos, a novel financial LLM framework that consists of PloutosGen and PloutosGPT.
Assumes 6
- Stock movement direction is determined by comparing adjusted closing prices over two consecutive trading days. market
Source sentence
The direction of a stock’s movement is determined by comparing its adjusted closing prices over two consecutive trading days.
- Positive label requires a price increase of at least 0.55%, negative label requires a decrease of at least 0.5%. data
Source sentence
Samples are labeled as either positive or negative based on the percentage change in closing price, with increases of 0.55% or more labeled as positive and decreases of 0.5% or more labeled as negative
- Tweets are extracted using regular expression queries based on NASDAQ ticker symbols for each day, excluding weekends and public holidays. data
Source sentence
Each of stock contains price data obtained from Yahoo Finance and related tweets data extracted using regular expression queries based on NASDAQ ticker symbols for each day (except for weekends and public holidays).
- A rolling window of T=5 trading days is used to generate samples. data
Source sentence
Then a T-day (T=5) rolling window is utilized to produce data candidates along the series of trading days.
Show 2 more
- Training, validation, and testing splits follow a 7:1:2 ratio over specified calendar periods. data
Source sentence
we split the dataset into distinct sets for training (01/01/2014 to 31/07/2015), validation (01/08/2015 to 30/09/2015), and testing (01/10/2015 to 01/01/2016) in a 7:1:2 ratio.
- Evaluation uses accuracy, F1 score, and Matthews Correlation Coefficient (MCC) for binary classification. statistical
Source sentence
We adopt accuracy, F1 score, and the Matthews correlation Coefficient (MCC, sourced from sklearn’s implementations)
Data
- ACL18 market-prices public
Source sentence
... we split the dataset into distinct sets for training (01/01/2014 to 31/07/2015), validation (01/08/2015 to 30/09/2015), and testing (01/10/2015 to 01/01/2016) in a 7:1:2 ratio.
- CIKM18 market-prices public
Source sentence
CIKM18 consists of 47 stocks with sufficient tweets and price data ranging from January 2017 to November 2017. Similarly, we leverage the same data processing method as ACL18 to process CIKM18 dataset
Yahoo Finance
Costs
Not applicable
Finds 3
- Ploutos outperforms state-of-the-art methods on both prediction accuracy and interpretability. performance
Source sentence
Our experiments demonstrate the effectiveness of our framework, which outperforms the state-of-the-art method on both prediction accuracy and
- Dynamic token weighting improves accuracy and faithfulness, with optimal temperature at 0.5 yielding maximum performance. predictive-accuracy
Source sentence
When temperature equals to 0.5, both faithfulness and accuracy reach their maximum, suggesting a synergistic rather than a mutually exclusive relationship between predictive accuracy and interpretability, which also demonstrates the effectiveness of dynamic token weighting.
- Ploutos shows superior faithfulness and informativeness compared to other LLM-based baselines. performance
Source sentence
Compared to other LLM methods, Ploutos demonstrates superior faithfulness and informativeness, which indicates that large language model can be trained through specific methods to generate informative and faithful rationales, thereby enhancing model’s interpretability.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 19 related papers). Coverage is partial, so this is neither replication nor refutation.
11 papers in the archive : Exploring the Synergy of Quantitative Factors and Newsflow …
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (146)
- Structure and Interpretation of Computer Programs (1985)
- Visual Information Extraction with Lixto (2001)
- An overview of the KL-ONE knowledge representation system (1985)
- Complexity results for nonmonotonic logics (1992)
- Hypertree Decompositions and Tractable Queries (2002)
- Foundations of a functional approach to knowledge representation (1984)
- A logic of implicit and explicit belief (1984)
- On the compilability and expressive power of propositional planning formalisms (2000)
and 138 more
- IJCAI Camera Ready Submission
- FinGPT: Open-Source Financial Large Language Models — → 2306.06031
- Hybrid deep sequential modeling for social text-driven stock prediction (2018)
- Halueval: A large-scale hallucination evaluation benchmark for large language models (2023)
- Truthfulqa: Measuring how models mimic human falsehoods (2021) — link
- Causality-Guided Multi-Memory Interaction Network for Multivariate Stock Price Movement Prediction (2023)
- Enhancing stock movement prediction with adversarial training (2018) — link
- Explainable text-driven neural network for stock prediction (2018)
- A Novel ensemble neuro-fuzzy model for financial time series forecasting (2019)
- An interpretable neuro-fuzzy approach to stock price forecasting (2019)
- Applications of Explainable Artificial Intelligence in Finance—a systematic review of Finance, Information Systems, and Computer Science literature (2023)
- Large Language Models in Finance: A Survey — → 2311.10723
- Bloomberggpt: A large language model for finance (2023) — link
- Instruct-FinGPT: Financial Sentiment Analysis by Instruction Tuning of General-Purpose Large Language Models — → 2306.12659
- Promptcast: A new prompt-based learning paradigm for time series forecasting (2023)
- Meta-transformer: A unified framework for multimodal learning (2023) — link
- One Fits All: Power General Time Series Analysis by Pretrained LM (2023) — link
- ChatGPT Informed Graph Neural Network for Stock Movement Prediction — → 2306.03763
- Alpha-GPT: Human-AI Interactive Alpha Mining for Quantitative Investment — → 2308.00016
- Gpt-4 technical report (2023) — link
- Llama 2: Open foundation and fine-tuned chat models (2023) — link
- Listening to chaotic whispers: A deep learning framework for news-oriented stock trend prediction (2018)
- Topic modeling based sentiment analysis on social media for stock market prediction (2015)
- Stock price prediction using LSTM, RNN and CNN-sliding window model (2017)
- Smoothing, forecasting and prediction of discrete time series (2004)
- The Wall Street Neophyte: A Zero-Shot Analysis of ChatGPT Over MultiModal Stock Movement Prediction Challenges — → 2304.05351
- Goat: Fine-tuned LLaMA Outperforms GPT-4 on Arithmetic Tasks (2023) — link
- Rest: Relational event-driven stock trend forecasting (2021)
- Twitter Financial News Sentiment (2022)
- Bitcoin Tweets Sentiment Analysis (Accessed: 2023)
- Data-Centric Financial Large Language Models (2023) — link
- Www'18 open challenge: financial opinion mining and question answering (2018)
- Exploiting social relations and sentiment for stock prediction (2014)
- Exploiting topic based twitter sentiment for stock prediction (2013)
- FAST: Financial news and tweet based time aware network for stock trading (2021)
- Stock movement prediction from tweets and historical prices (2018)
- Hats: A hierarchical graph attention network for stock movement prediction (2019) — link
- Spatiotemporal hypergraph convolution network for stock movement forecasting (2020)
- StockFormer: learning hybrid trading machines with predictive coding (2023)
- A CNN-BiLSTM-AM method for stock price prediction (2021)
- CNNpred: CNN-based stock market prediction using a diverse set of variables (2019)
- Stock market’s price movement prediction with LSTM neural networks (2017)
- A dual-stage attention-based recurrent neural network for time series prediction (2017) — link
- Stock price prediction using attention-based multi-input LSTM (2018)
- Large language models are zero-shot time series forecasters (2023) — link
- Letter to Shareholders (1991). (1991)
- A survey on large language model based autonomous agents (2023) — link
- Autogen: Enabling next-gen llm applications via multi-agent conversation framework (2023) — link
- PIXIU: A Large Language Model, Instruction Data and Evaluation Benchmark for Finance (2023) — link
- Can chatgpt forecast stock price movements? return predictability and large language models — → 2304.07619
- Deep attentive learning for stock movement prediction from social media text and company correlations (2020)
- An empirical analysis of data requirements for financial forecasting with neural networks (2001)
- Do measures of investor sentiment predict returns? (1998)
- Stock price prediction using prior knowledge and neural networks (1997)
- Association between accounting performance measures and stock prices: A test of the life cycle hypothesis (1992)
- Stock market volatility and learning (2016)
- Integrating knowledge tracing and item response theory: A tale of two frameworks (2014)
- Knowledge tracing: Modeling the acquisition of procedural knowledge (1994)
- Question Difficulty Prediction for READING Problems in Standard Tests (2017)
- EKT: Exercise-aware Knowledge Tracing for Student Performance Prediction (2019)
- Deep knowledge tracing (2015)
- Dynamic key-value memory networks for knowledge tracing (2017)
- Graph-based Knowledge Tracing: Modeling Student Proficiency Using Graph Neural Network (2019)
- Bert: Pre-training of deep bidirectional transformers for language understanding (2018) — link
- Convolutional neural networks for sentence classification (2014) — link
- What Does BERT Look At? An Analysis of BERT’s Attention (2019) — link
- Long short-term memory (1997)
- Interpretable Cognitive Diagnosis with Neural Network for Intelligent Educational Systems (2019) — link
- Hierarchical clustering schemes (1967)
- The choice of item difficulty in self-adapted testing. (2000)
- Xgboost: A scalable tree boosting system (2016)
- node2vec: Scalable feature learning for networks (2016)
- What Does a TextCNN Learn? (2018) — link
- Root mean square error (RMSE) or mean absolute error (MAE)?–Arguments against avoiding RMSE in the literature (2014)
- The effect of analogy instruction on young children’s metaphor comprehension (1998)
- The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems (2011)
- Efficient graph representations (2003)
- Using AUC and accuracy in evaluating learning algorithms (2005)
- Addressing two problems in deep knowledge tracing via prediction-consistent regularization (2018)
- Exploring hierarchical structures for recommender systems (2018)
- Hierarchical lifelong learning by sharing representations and integrating hypothesis (2019)
- Performance Factors Analysis–A New Alternative to Knowledge Tracing. (2009)
- Engaging with massive online courses (2014)
- Ubiquitous learning environment: An adaptive teaching system using ubiquitous technology (2004)
- Intelligent tutoring systems: Evolutions in design (2014)
- Hierarchical graph representation learning with differentiable pooling (2018)
- Incorporating rich features into deep knowledge tracing (2017)
- Student diagnosis in practice; bridging a gap (1995)
- Finding similar exercises in online education systems (2018)
- Retaining Minority Students in Higher Education: A Framework for Success. ASHE-ERIC Higher Education Report. Jossey-Bass Higher and Adult Education Series. (2003)
- Exam question recommender system (2005)
- Similarity Measure of Test Questions Based on Ontology and VSM (2014)
- A reinforcement learning approach to personalized learning recommendation systems (2019)
- Recommendation system for adaptive learning (2018)
- Enhancing mathematical problem solving among third-grade students with schema-based instruction. (2004)
- word2vec Explained: deriving Mikolov et al.’s negative-sampling word-embedding method (2014) — link
- Deep contextualized word representations (2018) — link
- Disan: Directional self-attention network for rnn/cnn-free language understanding (2018)
- MAM-RNN: Multi-level Attention Model Based RNN for Video Captioning. (2017)
- Adam: A method for stochastic optimization (2015)
- The graph neural network model (2008)
- Graph neural networks: A review of methods and applications (2018) — link
- Interoperability in personalized adaptive learning (2006)
- Textrank: Bringing order into text (2004)
- Silhouettes: a graphical aid to the interpretation and validation of cluster analysis (1987)
- Some methods for classification and analysis of multivariate observations (1967)
- A density-based algorithm for discovering clusters in large spatial databases with noise. (1996)
- A survey of text clustering algorithms (2012)
- Data clustering and similarity (2013)
- A comprehensive survey on graph neural networks (2020)
- Semi-supervised classification with graph convolutional networks (2017) — link
- Personalized recommendation of learning material using sequential pattern mining and attribute based collaborative filtering (2014)
- Exploiting cognitive structure for adaptive learning (2019)
- Automatically learning topics and difficulty levels of problems in online judge systems (2018)
- Cognitive modelling for predicting examinee performance (2015)
- Cognitive diagnostic assessment for education: Theory and applications (2007)
- DINA model and parameter estimation: A didactic (2009)
- Probabilistic-input, noisy conjunctive models for cognitive diagnosis (2018)
- The gap of semantic parsing: A survey on automatic math word problem solvers (2019)
- Natural language input for a computer problem solving system (1964)
- Experiments with a deductive question-answering program (1965)
- Understanding and solving arithmetic word problems: A computer simulation (1985)
- Robust understanding of word problems with extraneous information (2007)
- Frame-based calculus of solving arithmetic multi-step addition and subtraction word problems (2010)
- Neural message passing for quantum chemistry (2017) — link
- Inductive representation learning on large graphs (2017)
- Graph attention networks (2017) — link
- Knowledge questions from knowledge graphs (2017)
- Transition-based knowledge graph embedding with relational mapping properties (2014)
- Long-tail relation extraction via knowledge graph embeddings and graph convolution networks (2019) — link
- Case studies in the use of ROC curve analysis for sensor-based estimates in human computer interaction (2005)
- Stable prediction across unknown environments (2018)
- A comparison of document clustering techniques (2000)
- Empirical and theoretical comparisons of selected criterion functions for document clustering (2004)
- Evaluation of hierarchical clustering algorithms for document datasets (2002)
- Fast and Intuitive Clustering of Web Documents. (1997)
- Memory: A contribution to experimental psychology (2013)
- Self-attention graph pooling (2019) — link
7 of 146 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.