Hybrid LSTM and GRU for Cryptocurrency Price Forecasting Based on Social Network Sentiment Analysis Using FinBERT

In one sentence

Summary

This study addresses the challenge of predicting highly volatile cryptocurrency prices by integrating historical market data with social media sentiment. The authors argue that traditional models relying solely on price history miss critical information contained in public opinion, which significantly influences short-term price movements in the crypto market. The research focuses on Ethereum and Solana, targeting the need for more robust forecasting tools for traders and investors dealing with high-frequency, sentiment-driven assets. The methodology combines daily OHLC (Open, High, Low, Close) price data from Binance with daily sentiment scores extracted from Twitter. Sentiment analysis is performed using FinBERT, a financial-domain pre-trained BERT model, which processes cleaned tweets to generate a daily average sentiment score. This sentiment feature is concatenated with the normalized market data and fed into a hybrid LSTM-GRU neural network. The hybrid architecture leverages LSTM's long-term memory and GRU's computational efficiency to mitigate vanishing gradient issues. The model is evaluated against standalone LSTM and GRU models, both with and without sentiment inputs, using MSE, RMSE, MAE, and MAPE metrics over a one-year period. Results indicate that the hybrid LSTM-GRU model outperforms all benchmarks, achieving the lowest error rates. The inclusion of FinBERT-derived sentiment consistently improves prediction accuracy across all models, reducing MAPE by approximately 0.5% to 1%. For instance, the hybrid model with sentiment achieved a MAPE of 4.11% for Ethereum and 4.13% for Solana, compared to higher errors for models without sentiment. The study concludes that combining deep learning sequence models with financial NLP sentiment analysis provides a significant edge in cryptocurrency forecasting, though it notes limitations regarding data source diversity and computational cost.

Research problem

Cryptocurrency price prediction is difficult due to high volatility and the influence of non-market factors like public sentiment. Traditional models often ignore social media data, leading to suboptimal forecasts. This paper addresses the gap by integrating social sentiment into deep learning time-series models to improve prediction accuracy for specific altcoins.

Core contributions

• Proposes a hybrid LSTM-GRU architecture that balances memory capacity and computational efficiency for time-series forecasting.
• Integrates FinBERT-extracted social media sentiment as a novel feature for cryptocurrency price prediction.
• Demonstrates that adding sentiment data improves prediction accuracy for both LSTM and GRU models.
• Provides empirical evidence that hybrid RNN models outperform standalone LSTM or GRU models in crypto markets.
• Validates the approach on two major altcoins, Ethereum and Solana, showing consistent performance gains.

Method

• Data collection: Daily OHLC prices and Twitter tweets for ETH and SOL from 2021.
• Preprocessing: MinMax scaling for prices; cleaning tweets (removing URLs, symbols, RTs).
• Sentiment Extraction: FinBERT model classifies tweet sentiment into scores (0-1).
• Feature Engineering: Daily average sentiment score combined with OHLC features.
• Model Architecture: Parallel LSTM (50 neurons) and GRU (30 neurons) layers with dropout.
• Hybridization: Outputs from LSTM and GRU are concatenated and passed through dense layers.
• Training: Adam optimizer, MSE loss, batch size 16, 100 epochs, 7-day lookback window.
• Evaluation: Comparison of MAE, RMSE, MAPE across different model configurations.

Results & findings

• Hybrid LSTM-GRU with sentiment achieved the lowest MAPE (4.11% for ETH, 4.13% for SOL).
• Sentiment integration reduced MAPE by 0.5% to 1% across all model types compared to no sentiment.
• Hybrid LSTM-GRU outperformed standalone LSTM and GRU models in all error metrics.
• GRU generally outperformed LSTM, but the hybrid combination yielded the best results.
• Models without sentiment showed higher error rates, confirming the value of social data.
• The study confirms that social media sentiment is a significant predictor for crypto price movements.

Data & experiments

• Historical daily OHLC data for Ethereum and Solana from Binance (via CryptoDataDownload) for 2021.
• Twitter data collected using SNScrape, filtered for relevant keywords and high-engagement tweets.
• FinBERT model used for sentiment classification, outputting daily average sentiment scores.
• Experiments compare LSTM, GRU, and Hybrid LSTM-GRU models with and without sentiment features.
• Evaluation metrics include MAE, RMSE, MAPE, and MSE across five repeated trials for stability.
• Data split into 80% training and 20% testing sets with a 7-day window period for sequence input.

Limitations & caveats

• Sentiment data is limited to Twitter, ignoring other sources like news or forums.
• Computation cost is not reported, though hybrid models are likely more expensive.
• Sentiment extraction may be noisy due to outliers or irrelevant tweets despite filtering.
• Study period is limited to 2021, which may not generalize to different market conditions.
• No backtesting or trading strategy implementation, only price point prediction.
• FinBERT performance may vary for non-English tweets or slang-heavy crypto communities.

Curator notes

Discovery status: round2_promoted. Source links: FinBERT: A Pretrained Language Model for Financial Communications; FinBERT: Financial Sentiment Analysis with Pre-trained Language Models