Bitcoin Price Direction Prediction via Regime-Aware Multi-Modal Fusion of Social Sentiment and Technical Features

The Problem: Why Bitcoin Prediction Stays Hard

Bitcoin price prediction on sub-daily timescales—hours to a few hours ahead—remains notoriously difficult. The asset exhibits fat-tailed returns (extreme moves happen more often than traditional models expect), non-stationary dynamics (the statistical properties shift over time), and a price discovery process deeply entangled with social chatter on Reddit and Twitter. The authors' key insight is that existing methods treat all market conditions the same way. They fuse technical price data (OHLCV: open, high, low, close, volume) with social sentiment using static concatenation—essentially stacking the inputs side by side—and apply identical weights regardless of whether the market is calm or turbulent. This violates what behavioural finance already knows: retail sentiment is predictive during volatile periods but noisy and misleading during calm ones.

The Method: Regime-Aware Fusion

The authors propose Regime-Aware Multi-Modal Learning (RAML), which dynamically adjusts how much the model should trust sentiment versus price dynamics based on current market state.

The system works in three steps. First, it detects market regime using a simple, interpretable rule: rolling 24-hour volatility partitions each hour into either a "stable" or "volatile" regime. This avoids black-box regime detection and keeps the approach transparent.

Second, the model processes price features (OHLCV) and sentiment separately through neural components—a BiLSTM processes price history, while FinBERT (a BERT variant pre-trained on financial text) encodes Reddit sentiment from /r/Bitcoin. These produce two embeddings that sit side by side.

Third, a learnable sigmoid gate—a single parameterized function—adjusts the relative weight of sentiment versus price based on the detected regime. During volatile phases, the gate trusts sentiment more. During stable phases, it relies more on price dynamics. This adaptive weighting is the core innovation: instead of a static concatenation, the fusion is conditioned on market state.

Results and Why It Matters

The authors evaluate on 3,491 hourly observations spanning July 2024 to September 2025, predicting 3-hour and 6-hour ahead price direction (up or down). RAML achieves macro-F1 scores of 0.5474 (3-hour) and 0.5513 (6-hour), with the highest AUC at 3 hours (0.5084). These numbers are modest—barely above random guessing in absolute terms—but the ablation study reveals why: removing any component (sentiment branch, regime detection, or adaptive gating) degrades performance, and replacing adaptive weighting with static concatenation causes recall to collapse entirely at 6 hours (F1 drops to 0.14).

The finding is actionable: regime-conditioned fusion is not optional for multi-modal financial forecasting. Even on a hard, noisy prediction task, accounting for when sentiment matters versus when it doesn't is necessary. For practitioners, this suggests that building financial models without regime awareness—treating all market conditions identically—is likely to fail when stress-tested in-sample.