About CryptoSent

Bitcoin pricing exhibits extreme non-stationarity, heteroskedastic volatility, and decoupling from traditional valuation. CryptoSent is a high-frequency forecasting architecture that combines autoregressive price signals with natural language features. Using cloud-native infrastructure for minute-level inferencing, it deploys a hierarchical stacked ensemble. A Temporal Fusion Transformer (TFT) integrates a NeuralProphet-LSTM time-series branch with a regime-switching sentiment branch (FinBERT and ModernBERT). The approach achieves a MAPE of approximately 0.31% over a 12-hour horizon and significantly outperforms a random walk baseline, demonstrating the efficacy of attention-based mechanisms in capturing the interplay between social signals and price discovery.