PreBit - A multimodal model with Twitter FinBERT embeddings for extreme price movement prediction of Bitcoin

In one sentence

Summary

This paper addresses the challenge of predicting extreme price fluctuations in Bitcoin, a highly volatile asset influenced by speculative retail trading and social media sentiment. The authors argue that existing models relying solely on sentiment scores or traditional technical indicators fail to capture the full informational content of social media discussions. The research aims to determine if embedding raw Twitter text using financial-domain language models provides superior predictive power for significant market movements compared to baseline methods. The methodology introduces the PreBit dataset, comprising 9.4 million tweets from 2015 to 2021 alongside daily OHLCV data, correlated asset prices (Ethereum, Gold), and 13 technical indicators. The core model is a hybrid architecture: a Support Vector Machine processes the technical data, while a Convolutional Neural Network processes sentence-level FinBERT embeddings of concatenated daily tweets. These modalities are fused to classify next-day extreme price movements (defined as up/down 2% or 5%). The study includes an ablation analysis to isolate the contribution of social media data and backtests a trading strategy based on model predictions with varying confidence thresholds. Results indicate that adding FinBERT-embedded Twitter data significantly improves prediction accuracy over models using only technical indicators. The ablation study confirms that social media content adds unique predictive value. Backtesting reveals that a trading strategy based on the multimodal model’s predictions achieves higher returns with reduced risk exposure compared to simple buy-and-hold or moving average strategies. The authors highlight that adjusting the prediction threshold allows traders to balance profitability against risk, offering a practical tool for managing downside risk in cryptocurrency markets.

Research problem

The paper investigates whether social media discussions from the general public on Bitcoin have predictive power for extreme price movements, specifically aiming to improve upon existing models that rely on simplified sentiment scores. It addresses the gap in using full-text financial embeddings for cryptocurrency price prediction and the lack of backtested trading strategies that utilize such multimodal data to manage risk.

Core contributions

• Proposes PreBit, a multimodal deep learning model for predicting extreme Bitcoin price movements.
• Introduces a novel approach using sentence-level FinBERT embeddings of raw Twitter text instead of sentiment scores.
• Releases the PreBit dataset, containing 9.4 million tweets and corresponding market data from 2015-2021.
• Conducts an ablation study to quantify the marginal contribution of social media data to prediction accuracy.
• Demonstrates a backtested trading strategy that uses model predictions to reduce risk exposure compared to baseline strategies.

Method

• Data collection: 5,000 tweets per day using GetOldTweets-python library.
• Text preprocessing: Lowercasing, URL removal, stopword filtering, and tokenization.
• Embedding: Sentence-level FinBERT embeddings (768-dimensional) for tweet slices.
• Text modeling: CNN architectures (sequential and parallel) to process embedded tweet blocks.
• Technical modeling: SVM with RBF kernel on normalized technical indicators and asset prices.
• Fusion: Combining predictions from text and technical models using a fusion layer (SVM/Logistic Regression).
• Labeling: Binary classification based on daily high/low price changes exceeding 2% or 5% thresholds.

Results & findings

• The multimodal model outperforms models using only technical indicators or sentiment scores.
• Ablation studies confirm that Twitter data provides significant incremental predictive value.
• FinBERT embeddings capture financial context better than general-purpose sentiment models.
• The trading strategy based on model predictions shows reduced risk compared to buy-and-hold.
• Adjusting prediction thresholds allows for optimization of the risk-return trade-off.
• The study highlights the influence of retail investor sentiment on Bitcoin's extreme volatility.

Data & experiments

• PreBit dataset: 9,435,437 tweets containing 'Bitcoin' collected daily from 2015 to 2021.
• Technical data: Daily OHLCV, Ethereum and Gold prices, and 13 technical indicators (e.g., MACD, Bollinger Bands).
• Prediction tasks: Binary classification of next-day extreme price movements (up/down 2% or 5%).
• Experimental design: Hybrid model fusing SVM (technical data) and CNN (text data).
• Evaluation: Accuracy metrics and backtesting of a trading strategy with varying prediction thresholds.

Limitations & caveats

• The model is specific to Bitcoin and may not generalize to other cryptocurrencies without retraining.
• Twitter data may contain noise, bots, or non-financial discussions that are hard to filter completely.
• The study uses daily data, missing intraday dynamics and high-frequency trading opportunities.
• FinBERT embeddings are static and may not capture real-time context shifts effectively.
• The trading strategy backtest does not account for transaction costs or slippage in detail.

Curator notes

Discovery status: round4_promoted. Source links: Large Language Models and Sentiment Analysis in Financial Markets: A Review, Datasets, and Case Study