Analyzing News Sentiment Impact on BTC Futures: A 3-State HMM Approach

I’ve been working on a pipeline to map Tier-1 crypto news (CoinDesk) to 1-minute Binance Futures microstructure data, and I wanted to share some findings regarding news impact decay and market regimes.

I built a pipeline that aligns news timestamps with price action at T0, T+5m, T+15m, and T+1h, while enriching it with pre-market volume anomalies and funding rate data. After processing ~35,000 events, I applied a 3-State Gaussian Hidden Markov Model (HMM) to classify market regimes. Here is what the data suggests:

1. Regime-Dependent Decay: The market’s reaction is not universal. In a "Flat" regime (State 2), I’m observing a classic "Spike & Revert" pattern—prices move violently in the first 5 minutes post-headline but almost always mean-revert within 15-20 minutes. Trading breakouts here is a trap.
2. The Altcoin Inertia: While BTC absorbs macro news shocks within ~5 minutes, assets like SOL and LINK show a consistent 15-to-30 minute lag in absorption. There seems to be a reliable statistical arbitrage window here for momentum-based altcoin strategies.
3. Volume Anomaly as a Predictor: Using a 1-hour pre-market volume anomaly metric (comparing current volume vs. rolling baseline), I’ve found that events with a >1.5x anomaly significantly correlate with higher magnitude moves post-publication.

Methodology:

• Source: CoinDesk headlines + Binance Futures (/fapi/v1/).
• Alignment: No-look-ahead script (matching news to the exact minute-candle close).
• Classification: 3-State Gaussian HMM (trained on rolling returns/volatility).

I’ve uploaded a sample of this data to Kaggle along with a Jupyter notebook that visualizes these decay curves. I’m curious if anyone here has experimented with HMM for news classification, or if there are other microstructure features (like order book imbalance at the moment of news) that you've found to improve predictive accuracy?