Paper: arXiv 2306.02136
Authors: Tingsong Jiang, Qingyun Zeng
Abstract
In this study, we integrate sentiment analysis within a financial framework by leveraging FinBERT, a fine-tuned BERT model specialized for financial text, to construct an advanced deep learning model based on Long Short-Term Memory (LSTM) networks. Our objective is to forecast financial market trends with greater accuracy. To evaluate our model’s predictive capabilities, we apply it to a comprehensive dataset of stock market news and perform a comparative analysis against standard BERT, standalone LSTM, and the traditional ARIMA models. Our findings indicate that incorporating sentiment analysis significantly enhances the model’s ability to anticipate market fluctuations. Furthermore, we propose a suite of optimization techniques aimed at refining the model’s performance, paving the way for more robust and reliable market prediction tools in the field of AI-driven finance.
Complexity vs Empirical Score
- Math Complexity: 4.0/10
- Empirical Rigor: 7.5/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper employs a relatively straightforward application of established deep learning architectures (FinBERT and LSTM) without advanced theoretical derivations, but it is heavily implementation-focused with a comprehensive dataset, detailed preprocessing steps, and performance comparisons against baselines, making it more practical and empirical.
Research Flowchart
flowchart TD Start["Research Goal: Forecast stock market trends<br>using financial sentiment analysis"] --> Data["Dataset: Comprehensive stock market news"] Data --> Process1["Text Preprocessing &<br>Sentiment Extraction (FinBERT)"] Process1 --> Process2["Feature Integration &<br>Model Construction (LSTM)"] Process2 --> Eval["Comparative Evaluation<br>vs. BERT, LSTM, ARIMA"] Eval --> Outcome["Key Findings:<br>1. Sentiment integration enhances accuracy<br>2. FinBERT-LSTM outperforms baselines<br>3. Proposed optimization techniques<br>achieved robust predictions"]
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 a model that combines the financial domain‑adapted FinBERT transformer with an LSTM network to predict stock market movements using sentiment extracted from news.
Source sentence
We propose a model which takes advantage of Transformer based contextual language model for analyzing the sentiment in financial markets, and then using LSTM‑based deep learning model in the prediction of the exact movement of the financial market by the sentiment.
Assumes 5
- Financial sentiment from news, social media, etc. has strong correlations to market movements, based on the Efficient Market Hypothesis. market
Source sentence
Our basic hypothesis is that the financial sentiment by news, social media, etc. have strong correlations to the movement of financial markets (stock market, foreign exchange, bond, etc.). This hypothesis is based on Efficient Market Hypothesis (EMH), which states that share prices reflect all information.
- Statistical models like ARIMA do not capture the nonlinearity in financial data. model
Source sentence
Traditional statistical method such as ARIMA has been used widely in the modelling of financial markets. However, statistical models like ARIMA does not capture well the nonlinearity in the financial data.
- The relationship studied is correlation only; no causal analysis is performed. other
Source sentence
One thing need to be pointed our that, we are not doing casual analysis here, as market could influence news and news might also influence the market.
- Sentiment scores are aggregated daily per stock ticker by averaging probabilities of positive and negative classes. data
Source sentence
These probabilities were averaged daily per stock ticker, resulting in a daily aggregated sentiment score computed as: SentimentScore = P(positive) −P(negative)
Show 1 more
- Missing records are dropped and non‑trading days are handled by removing incomplete data. data
Source sentence
We handled missing values by dropping incomplete records.
Data
- Historical Financial News Archive text-news public
Source sentence
This dataset, publicly available on Kaggle.com, comprises news archives for over 800 US companies, spanning 12 years up to 2019.
- Daily Financial News for 6000+ Stocks text-news public
Source sentence
Also sourced from Kaggle.com, this dataset offers a broader range of news items, albeit with less detailed information per entry (e.g., primarily headers).
- Stock Market Data (price, volume, etc.) market-prices public
Source sentence
We retrieve historical stock market data (e.g., open price, close price, percentage change) for the tickers identified in our news datasets using the yfinance API.
- FinBERT ('ProsusAI/finbert') text-news public
Source sentence
FinBERT (‘ProsusAI/finbert‘) obtained from the Hugging Face Model Hub.
- Base BERT uncased text-news public
Source sentence
We employ base BERT uncased and FinBERT (‘ProsusAI/finbert‘) obtained from the Hugging Face Model Hub.
Costs
Not applicable
Finds 4
- FinBERT+LSTM achieves the lowest validation loss among all tested models. performance
Source sentence
FINBERT + LSTM (PROPOSED) 9.1701E-05 3.1975E-04
- LSTM alone performs much better than ARIMA on the forecasting task. predictive-accuracy
Source sentence
We find the the LSTM model in general perform much better than ARIMA model
- Sentiment analysis using FinBERT outperforms LSTM‑only models, and the combined FinBERT+LSTM model outperforms both. predictive-accuracy
Source sentence
sentiment analysis using FinBERT out perform LSTM, whereras FinBERT combined with LSTM acheived the best performance.
- ARIMA captures long‑run trends well but has higher MSE (0.00975) compared to the proposed model. performance
Source sentence
ARIMA actually capture the long run trend well, whereas our model is able to fit the short run trend
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 46 related papers). Coverage is partial, so this is neither replication nor refutation.
2 papers in the archive : FinBERT-BiLSTM: A Deep Learning Model for Predicting …
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.