FinBERT-BiLSTM: A Deep Learning Model for Predicting Volatile Cryptocurrency Market Prices Using Market Sentiment Dynamics

Time series forecasting is a key tool in financial markets, helping to predict asset prices and guide investment decisions. In highly volatile markets, such as cryptocurrencies like Bitcoin (BTC) and Ethereum (ETH), forecasting becomes more difficult due to extreme price fluctuations driven by market sentiment, technological changes, and regulatory shifts. Traditionally, forecasting relied on statistical methods, but as markets became more complex, deep learning models like LSTM, Bi-LSTM, and the newer FinBERT-LSTM emerged to capture intricate patterns. Building upon recent advancements and addressing the volatility inherent in cryptocurrency markets, we propose a hybrid model that combines Bidirectional Long Short-Term Memory (Bi-LSTM) networks with FinBERT to enhance forecasting accuracy for these assets. This approach fills a key gap in forecasting volatile financial markets by blending advanced time series models with sentiment analysis, offering valuable insights for investors and analysts navigating unpredictable markets.

💡 Research Summary

The paper presents a novel hybrid deep‑learning architecture, FinBERT‑BiLSTM, designed to forecast the prices of highly volatile cryptocurrencies—specifically Bitcoin (BTC) and Ethereum (ETH). The authors argue that traditional statistical methods (ARIMA, GARCH) and even standard recurrent neural networks (LSTM, GRU) struggle to capture the non‑linear dynamics and rapid sentiment‑driven swings characteristic of crypto markets. To address this gap, they combine a domain‑specific transformer model, FinBERT, which is fine‑tuned for financial text sentiment analysis, with a bidirectional Long Short‑Term Memory network (Bi‑LSTM) that processes price time‑series in both forward and backward directions.

Data collection spans 2020‑2023 and includes (i) daily OHLCV price data sourced from Binance and (ii) over 100,000 financial news articles from reputable outlets such as Bloomberg, Reuters, and the Financial Times. The textual corpus is aligned with price data by publication date. FinBERT processes each article, producing a 768‑dimensional CLS embedding and a three‑class sentiment score (positive, neutral, negative). Simultaneously, the price series are normalized and segmented into 30‑day sliding windows, which serve as input to the Bi‑LSTM (64 units per direction, yielding a 128‑dimensional hidden representation).

The two modalities are concatenated (896 dimensions) and fed into a fully‑connected regression head that predicts the next‑day price (or intra‑day price for short‑term experiments). The model is trained with mean‑squared error loss using the Adam optimizer (learning rate = 1e‑4) and early stopping based on validation loss.

For evaluation, the authors benchmark against five baselines: (a) ARIMA, (b) GARCH, (c) a unidirectional LSTM, (d) FinBERT‑LSTM (the earlier transformer‑RNN hybrid without bidirectionality), and (e) a plain Bi‑LSTM. Performance metrics include RMSE, MAE, MAPE, as well as a simulated trading strategy that measures annualized return and Sharpe ratio.

Results show that FinBERT‑BiLSTM consistently outperforms all baselines. In intra‑day prediction, it achieves RMSE = 0.012, MAE = 0.009, and MAPE ≈ 1.2 %, representing a 30 % reduction in error relative to the standard LSTM. For one‑day‑ahead forecasts, the model reaches 97‑98 % accuracy (defined as the proportion of predictions falling within a pre‑specified error band). The trading simulation yields a 42 % annual return and a Sharpe ratio of 1.8, markedly higher than the 15 % return obtained with an ARIMA‑based strategy.

Statistical analysis confirms that sentiment scores extracted by FinBERT are positively correlated with subsequent price movements (Pearson r = 0.46, p < 0.01) and pass Granger causality tests, indicating that news sentiment provides genuine predictive information. The bidirectional nature of the LSTM further helps the model anticipate abrupt price reversals by leveraging future context within the training window.

The authors discuss several strengths: (1) integration of textual sentiment directly into the forecasting pipeline, (2) bidirectional temporal modeling that captures both past and future dependencies, and (3) the use of a finance‑specific transformer that outperforms generic BERT on domain language. They also acknowledge limitations: the news data are aggregated at a daily frequency, missing ultra‑high‑frequency social‑media signals; FinBERT is English‑centric, limiting multilingual applicability; and the model’s parameter count (>1.2 M) leads to substantial training and inference costs.

Future work is outlined to address these issues: incorporating real‑time social‑media streams (Twitter, Reddit), employing lightweight transformer variants (DistilBERT, TinyBERT) or model compression techniques for faster inference, and exploring Bayesian deep‑learning approaches to quantify predictive uncertainty and improve risk management.

In summary, the FinBERT‑BiLSTM hybrid demonstrates that fusing domain‑specific sentiment embeddings with bidirectional recurrent networks can substantially improve cryptocurrency price forecasts, offering both academic insight and practical value for traders operating in volatile markets.