Crypto sentiment scoring vs price-action regimes: combining FinBERT NLP with HMM returns
Relying on crypto news sentiment alone whipsaws you on sensationalist headlines while trends quietly persist; relying on price returns alone misses macro and narrative context until volatility already spikes. Market Brief AI lets you compare and combine both on iPhone and iPad—running a pure price-return HMM (Hera) alongside input-aware and factorial HMMs (Metis and Athena) fed by local Core ML FinBERT GDELT sentiment and macro agents, strictly for educational simulation, not financial advice.
1. Strengths and blind spots: news sentiment vs return-series regimes
Financial headlines capture narrative tone instantly—regulatory lawsuits, ETF filings, or hawkish central-bank remarks register in text before multi-day moving averages budge. Yet media coverage exhibits strong negativity bias and short half-life: alarming headlines routinely appear during healthy bull-regime pullbacks without breaking positive return drift.
Conversely, pure price-action Hidden Markov Models (like Hera on Binance USDT spot bars) measure objective market clearing returns and volatility dispersion, immune to clickbait—at the cost of reacting only once price bars reflect the shift.
| Analytical property | Headline Sentiment Scoring (FinBERT / GDELT) | Price-Action Regime Modeling (Return HMMs) |
|---|---|---|
| Primary signal measured | Linguistic polarity (positive / neutral / negative financial tone) | Latent return mean (mu), variance (sigma) & state transition odds |
| Key strength | Captures narrative, regulatory & macro news tone immediately | Measures actual capital-weighted market clearing behaviour |
| Primary failure mode | Media clickbait & scary headlines during intact uptrends | Purely backward-looking until candle returns/variance shift |
| Time-series continuity | Bursty—some days produce 50 headlines, quiet weekends produce few | Continuous—every daily, 1h, or 5m Binance USDT candle closes on schedule |
| How Market Brief AI runs it | Bundled Core ML FinBERT on device + interactive sentence scorer | Hera, Metis, Hestia & Athena fitted locally on iPhone and iPad |
2. How Market Brief AI fuses FinBERT, macro agents, and HMMs
Instead of forcing you into an ideological camp (“news doesn’t matter” vs “charts don’t matter”), Market Brief AI treats sentiment and macro releases as testable inputs on one screen. Public GDELT headlines are scored locally via Core ML FinBERT, while eight named agents pull public BLS CPI (Kronos), S&P 500 (Zeus), Fed tone (Demeter), USGS earthquakes (Poseidon), NOAA tornadoes (Aiolos), and seasonal/calendar effects (Tyche, Hermes, Helios).
You can run Hera (returns-only 3-state HMM) as your control group, then run Metis (input-aware HMM in Plus) or Athena (factorial HMM in Pro) on the exact same Binance USDT window to see whether external inputs sharpened or degraded the closed-bar backtest against buy-and-hold.
| Experiment setup | Inputs active | What you learn on the backtest chart |
|---|---|---|
| Pure price-action baseline | Hera 3-state HMM on Binance USDT bars only | Baseline purple strategy curve vs blue passive buy-and-hold |
| News + inflation context | Metis input-aware HMM + FinBERT GDELT + Kronos (CPI) | Whether headline polarity & CPI releases reduced whipsaws |
| Cross-market & Fed policy | Athena factorial HMM + Zeus (S&P 500) + Demeter (Fed tone) | How equity-index correlation & Fed tone interact with crypto volatility |
| Structure + crash guards | Add Pro auction signals (POC/LVN) + CUSUM/EMA crash guards | Inspect the Pro execution log to see which rule drove simulated exits |
3. Keeping experimentation honest and private
Adding more inputs does not automatically beat a simpler model—extra covariates can overfit short windows if not checked across multiple start dates. Having Hera, Metis, Hestia, and Athena in one offline-capable iPad laboratory encourages scientific comparison rather than blind faith in either news or price.
And because Core ML FinBERT runs entirely on your device, even custom sentences you type into the sentiment analyser never leave your phone.
| Capability | Free ($0) | Plus ($7.99/mo) / Pro ($12.99/mo) |
|---|---|---|
| Core ML FinBERT on GDELT & custom sentences | Included free on device | Included |
| Hera 3-state return HMM & closed-bar backtest | Included free | Included |
| Metis input-aware & Hestia duration HMMs | — | Unlocked in Plus (7-day free trial) |
| Athena factorial HMM, Zeus/Demeter & full trade log | — | Unlocked in Pro (3-day free trial) |
What is FinBERT and why does Market Brief AI use it instead of a word list?
FinBERT is a transformer language model fine-tuned specifically on financial text so phrases like “cutting rates” or “narrowing losses” are scored in financial context rather than naive dictionary counts—running locally via Core ML.
Which Hidden Markov Model incorporates external inputs like sentiment and macro data?
Metis (unlocked in Plus) is specifically designed as an input-aware HMM, while Athena (unlocked in Pro) tracks layered factorial regimes alongside Pro macro agents.
Can I score my own custom headlines offline?
Yes. Open the built-in sentiment analyser inside Market Brief AI and score any sentence on demand with zero internet or cloud AI.
Does combining sentiment and price regimes guarantee profitable trades?
No. Market Brief AI is strictly an educational strategy-simulation tool; historical backtests never guarantee future performance and nothing shown is financial advice.
