Attention-augmented hybrid CNN-LSTM model for social media sentiment analysis in cryptocurrency investment decision-making.

Résumé

Cryptocurrencies have emerged miraculously all over the globe due to their legitimacy, transparency, immutability, and the traceability that blockchain technology provides. However, the benefits it provides are dwarfed by how unpredictable and extremely price-volatile the cryptocurrencies are. That makes it really tough for investors to find their profitable opportunities in such volatile markets. Social media sources, like Twitter and Reddit, have evolved as crucial tools of sentiment estimation above the explosively volatile price movements of decentralized currencies. Here we introduce an attention-based hybrid CNN-LSTM model optimized for social media sentiment analysis to use them towards investment decisions in a broad portfolio of cryptocurrencies. The existing Convolutional Neural Network (CNN) effectively extracts the essential features, and Long Short-Term Memory (LSTM) has the potential to capture the long dependencies between phrases. Although these models can process massive textual data, they limit treating all the features equally important. Therefore, the proposed model induces the attention mechanism into hybrid CNN-LSTM for emphasizing more or fewer weights on different words according to their contributions and optimizes the parameters of employed neural networks using grid search. In our pipeline, the attention-augmented CNN-LSTM first transforms each tweet/review into a 512-dimensional task-specific embedding; a calibrated radial-basis SVM then serves as the final decision layer, refining the margin for classes that the neural network alone tends to blur. This sequential ('deep-features-plus-SVM') architecture boosts F1 by 3.2 pp over a pure Softmax head while adding only 0.4 ms of inference time. Extensive experiments conducted on cryptocurrency-related tweets and Reddit reviews reveal the outperformance of the proposed model over existing Deep Neural Networks (DNNs) and state-of-the-art models. Trained on 9.9 k crypto-tweets and 33 k Reddit comments, AEH attains 98.7% accuracy, 0.987 F1, and κ = 0.94, outperforming strong baselines (pure LSTM + 8.3 pp; pure CNN + 19.3 pp) and the widely-used VADER toolkit (+ 11.8 pp). On the forecasting side, a complementary GRU regressor trained on eight-year price series yielded MAE = 0.0315, MAPE = 5.95%, and MSE = 0.0022 for Bitcoin, beating an ARIMA benchmark at p