Paper: arXiv 2502.14897

Authors: Hamid Moradi-Kamali, Mohammad-Hossein Rajabi-Ghozlou, Mahdi Ghazavi, Ali Soltani, Amirreza Sattarzadeh, Reza Entezari-Maleki

Abstract

Financial Sentiment Analysis (FSA) traditionally relies on human-annotated sentiment labels to infer investor sentiment and forecast market movements. However, inferring the potential market impact of words based on their human-perceived intentions is inherently challenging. We hypothesize that the historical market reactions to words, offer a more reliable indicator of their potential impact on markets than subjective sentiment interpretations by human annotators. To test this hypothesis, a market-derived labeling approach is proposed to assign tweet labels based on ensuing short-term price trends, enabling the language model to capture the relationship between textual signals and market dynamics directly. A domain-specific language model was fine-tuned on these labels, achieving up to an 11% improvement in short-term trend prediction accuracy over traditional sentiment-based benchmarks. Moreover, by incorporating market and temporal context through prompt-tuning, the proposed context-aware language model demonstrated an accuracy of 89.6% on a curated dataset of 227 impactful Bitcoin-related news events with significant market impacts. Aggregating daily tweet predictions into trading signals, our method outperformed traditional fusion models (which combine sentiment-based and price-based predictions). It challenged the assumption that sentiment-based signals are inferior to price-based predictions in forecasting market movements. Backtesting these signals across three distinct market regimes yielded robust Sharpe ratios of up to 5.07 in trending markets and 3.73 in neutral markets. Our findings demonstrate that language models can serve as effective short-term market predictors. This paradigm shift underscores the untapped capabilities of language models in financial decision-making and opens new avenues for market prediction applications.

Complexity vs Empirical Score

  • Math Complexity: 2.5/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; reported Sharpe ratio of 5.07 is implausibly high for daily or lower-frequency data; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 8.0)

Why this score: The paper focuses on practical implementation with a detailed backtesting methodology (Sharpe ratios, market regimes) and real data application, but relies on established NLP/ML techniques without novel mathematical derivations or heavy statistical theory.

Research Flowchart

  flowchart TD
  A["Research Goal: <br>Can market-derived sentiment <br>outperform human-annotated sentiment?"] --> B["Data Preparation"]
  
  subgraph B ["Data Preparation"]
      B1["Input: Historical Crypto Tweets"]
      B2["Input: Bitcoin Price Data"]
      C1["Market-Derived Labeling: <br>Assign labels based on <br>subsequent price trends"]
  end
  
  C1 --> D["Model Development"]
  
  subgraph D ["Model Development"]
      D1["Fine-tune Language Model <br>on market-derived labels"]
      D2["Prompt-Tuning: <br>Inject market & temporal context"]
  end
  
  D --> E["Validation & Backtesting"]
  
  subgraph E ["Validation & Backtesting"]
      E1["Short-term trend prediction"]
      E2["Test on news events dataset"]
      E3["Backtest trading signals <br>across market regimes"]
  end
  
  E --> F["Key Outcomes"]
  
  subgraph F ["Key Outcomes"]
      F1["+11% accuracy <br>vs sentiment benchmarks"]
      F2["89.6% accuracy <br>on news event prediction"]
      F3["Sharpe ratios up to 5.07 <br>(trending markets)"]
      F4["Outperforms traditional <br>fusion models"]
  end

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 market-derived labeling approach for financial sentiment analysis and a context‑aware language model fine‑tuned on these labels to predict short‑term Bitcoin market movements.

Source sentence
To address these limitations, we propose a novel approach to FSA that focuses on measurable market reactions rather than inferred intentions.

Assumes 2

  • Each tweet has a direct short‑term market impact that can be captured by the labeling scheme. market
    Source sentence
    Our approach assumes a direct short-term market impact for each tweet, which may oversimplify the intricate and often indirect relationship between social media activity and market movements.
  • The model can learn to map tweet text to market movements without needing to model underlying intent. model
    Source sentence
    By retaining the tweet’s representation, the model can identify nuanced relationships between specific words and their effect on market behavior, bridging the gap between textual sentiment and market outcomes.

Data

Backtest sample: BTC/USDT · crypto · daily

  • Bitcoin Historical Events · Kaggle text-news public
    2009-2024 · significant Bitcoin market events
    Source sentence
    This dataset contains 227 significant Bitcoin market events from 2009 to 2024, focusing on benchmarking model performance during influential
  • Tweets · PreBit and Kaggle alternative public
    2015-early 2023 · Bitcoin‑related tweets
    Source sentence
    Tweet data from two sources, spanning 2015 to early 2023, was utilized. The first dataset
  • Technical Indicators (RSI, ROC) market-prices public
    January 2015 to March 2023 · OHLCV data for Bitcoin
    Source sentence
    The Relative Strength Index (RSI) and Rate of Change (ROC) are calculated from OHLCV

Costs

Not modeled A trading or execution result is reported without costs.

Finds 5

Reported performance (context-aware language model with TBL signal generation): Sharpe 5.07 (benchmark 3.12)

  • Fine‑tuning the language model on market‑derived labels improves short‑term trend prediction accuracy by up to 11% over traditional sentiment benchmarks. performance
    Source sentence
    A domain-specific language model was fine-tuned on these labels, achieving up to an 11% improvement in short-term trend prediction accuracy over traditional sentiment-based benchmarks.
  • The context‑aware language model attains 89.6% accuracy on a curated set of 227 impactful Bitcoin news events. predictive-accuracy
    Source sentence
    the proposed context-aware language model demonstrated an accuracy of 89.6% on a curated dataset of 227 impactful Bitcoin-related news events with significant market impacts.
  • Aggregated trading signals from the model yield Sharpe ratios up to 5.07 in trending markets and 3.73 in neutral markets. performance
    Source sentence
    Backtesting these signals across three distinct market regimes yielded robust Sharpe ratios of up to 5.07 in trending markets and 3.73 in neutral markets.
  • Using market‑derived labeling and prompt‑tuning raises the F1‑score from baseline to 80.3%, a 47% relative improvement. predictive-accuracy
    Source sentence
    Fine-tuning language models with these market-derived labels improved the F1-score by 15.3%, and further integrating market context through prompt‑tuning boosted performance by a staggering 47%, reaching an F1-score of 80.3%.
  • The Majority TBL strategy achieves the highest Sharpe ratio of 5.07 in bullish markets and 4.70 in bearish markets, outperforming buy‑and‑hold benchmarks. performance
    Source sentence
    Notably, the Majority TBL strategy achieved the highest Sharpe ratio in both the Bullish and Bearish markets, with a Sharpe ratio of 5.07 in the former and 4.70 in the latter.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 33 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive (1 by independent authors)

Code

https://github.com/hamidm21/Revisit_FSA — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (69)

  1. Prompt-Learning for Short Text Classification (2024)
  2. Natural language based financial forecasting: a survey (2018)
  3. Exploiting Topic based Twitter Sentiment for Stock Prediction (2013)
  4. Sentiment Classification of Cryptocurrency-Related Social Media Posts (2023)
  5. Twitter mood predicts the stock market (2011)
  6. Giving Content to Investor Sentiment: The Role of Media in the Stock Market (2007)
  7. Domain Adaptive Multi-Modality Neural Attention Network for Financial Forecasting (2020)
  8. KryptoOracle: A Real-Time Cryptocurrency Price Prediction Platform Using Twitter Sentiments (2020)
and 61 more
  1. Fineas: Financial embedding analysis of sentiment (2021) — link
  2. Deep learning for event-driven stock prediction (2015)
  3. Whether and when did bitcoin sentiment matter for investors? Before and during the COVID-19 pandemic (2023)
  4. Is All that Talk Just Noise? The Information Content of Internet Stock Message Boards (2001)
  5. Bitcoin price change and trend prediction through twitter sentiment and data volume (2022)
  6. Predicting Stock Market Indicators Through Twitter “I hope it is not as bad as I fear” (2011)
  7. The predictive power of public Twitter sentiment for forecasting cryptocurrency prices (2020)
  8. Beyond Polarity: Interpretable Financial Sentiment Analysis with Hierarchical Query-driven Attention (2018)
  9. Enhancing financial sentiment analysis via retrieval augmented large language models — → 2310.04027
  10. Temporal effects on pre-trained models for language processing tasks (2022)
  11. Does twitter affect stock market decisions? financial sentiment analysis during pandemics: A comparative study of the h1n1 and the covid-19 periods (2022)
  12. PreBit—A multimodal model with Twitter FinBERT embeddings for extreme price movement prediction of Bitcoin (2023)
  13. Cryptocurrency price prediction using news and social media sentiment (2017)
  14. A Cryptocurrency Price Prediction Model Based on Twitter Sentiment Indicators (2022)
  15. On the predictive power of tweet sentiments and attention on bitcoin (2022)
  16. Advanced social media sentiment analysis for short-term cryptocurrency price prediction (2020)
  17. Sentiment-Driven Price Prediction of the Bitcoin based on Statistical and Deep Learning Approaches (2020)
  18. Forecasting Bitcoin Price Fluctuation by Twitter Sentiment Analysis (2020)
  19. Technical analysis and sentiment embeddings for market trend prediction (2019)
  20. A Novel Twitter Sentiment Analysis Model with Baseline Correlation for Financial Market Prediction with Improved Efficiency (2020)
  21. Twitter-Based uncertainty and cryptocurrency returns (2022)
  22. LSTM based stock prediction using weighted and categorized financial news (2023)
  23. Ask “Who”, Not “What”: Bitcoin Volatility Forecasting with Twitter Data (2021)
  24. Time-aware language models as temporal knowledge bases (2022)
  25. ParsBERT Post-Training for Sentiment Analysis of Tweets Concerning Stock Market (2021)
  26. Stock Market Trend Forecasting Based on Multiple Textual Features: A Deep Learning Method (2021)
  27. Advances in financial machine learning (2018)
  28. CADEN: A Context-Aware Deep Embedding Network for Financial Opinions Mining (2018)
  29. FAST: Financial News and Tweet Based Time Aware Network for Stock Trading (2021)
  30. Context-aware Embeddings for Stock Prediction with Visual Clues (2022)
  31. Cryptocurrency Bubble Detection: A New Stock Market Dataset, Financial Task & Hyperbolic Models (2022)
  32. Time-Weighted LSTM Model with Redefined Labeling for Stock Trend Prediction (2017)
  33. The Adaptive Markets Hypothesis: Market Efficiency from an Evolutionary Perspective (2004)
  34. Fingpt: Open-source financial large language models — → 2306.06031
  35. Efficient Capital Markets: A Review of Theory and Empirical Work (1970)
  36. Do President Trump’s tweets affect financial markets? (2021)
  37. How Elon Musk’s twitter activity moves cryptocurrency markets (2023)
  38. Analysis on Enhancing Financial Decision-making Through Prompt Engineering (2023)
  39. Leveraging Large Language Models and Prompt Settings for Context-Aware Financial Sentiment Analysis (2024)
  40. A unified approach to interpreting model predictions (2017) — link
  41. User2Vec: A Novel Representation for the Information of the Social Networks for Stock Market Prediction Using Convolutional and Recurrent Neural Networks (2023)
  42. Financial sentiment analysis: Techniques and applications (2024)
  43. Financial sentiment analysis (FSA): A survey (2019)
  44. The power of scale for parameter-efficient prompt tuning (2021) — link
  45. Bert: Pre-training of deep bidirectional transformers for language understanding (2018) — link
  46. FinBERT: A large language model for extracting information from financial text (2023)
  47. BERTweet: A pre-trained language model for English Tweets (2020) — link
  48. Time series generation with masked autoencoder (2022) — link
  49. Advanced financial data processing and labeling methods for machine learning (2022)
  50. Multi-task meta label correction for time series prediction (2024)
  51. Enhanced Genetic-Algorithm-Driven Triple Barrier Labeling Method and Machine Learning Approach for Pair Trading Strategy in Cryptocurrency Markets (2024)
  52. Simmtm: A simple pre-training framework for masked time-series modeling (2023)
  53. Cryptocurrency Price Prediction Algorithms: A Survey and Future Directions (2024)
  54. Explainable assessment of financial experts’ credibility by classifying social media forecasts and checking the predictions with actual market data (2024)
  55. News-based intelligent prediction of financial markets using text mining and machine learning: A systematic literature review (2023)
  56. FinSenticNet: A concept-level lexicon for financial sentiment analysis (2023)
  57. Sentiment analysis: a review and comparative analysis over social media (2020)
  58. Stock market trend forecasting based on multiple textual features: a deep learning method (2021)
  59. Large language models for newspaper sentiment analysis during COVID-19: The Guardian (2025)
  60. BERT-based combination of convolutional and recurrent neural network for indonesian sentiment analysis (2024)
  61. BERT-Driven stock price trend prediction utilizing tokenized stock data and multi-step optimization approach (2025)

2 of 69 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.