Research & Methodology

Sentiment Intelligence
Methodology

How we collect, process, score, and validate market sentiment signals across 185 financial assets — with full backtesting results and open methodology.

145+
News feeds monitored continuously
185
Assets tracked across 4 classes
131k+
Articles analyzed in the database
r = 0.211
Peak BTC correlation at +2h lag (p<0.001)

Data coverage

SentiSignal aggregates news from financial news outlets, crypto publishers, commodity and energy news feeds, social media aggregators, and central bank data feeds. The 185 tracked assets span four classes:

Asset class Count Examples Sentiment pipeline
Cryptocurrency 130 BTC, ETH, SOL, XRP, DOGE, AVAX… Full — all 5 stages
Commodities 20 GOLD, SILVER, OIL_BRENT, NATURAL_GAS, WHEAT… Full — all 5 stages
FX pairs 35 EUR/USD, GBP/USD, USD/JPY, AUD/USD… LLM only — no algo prefilter
Macro indicators 13 VIX, SPX, DXY, US10Y, YIELD_SPREAD… LLM only
Why FX skips the algo prefilter: Currency-affecting news (FOMC decisions, geopolitical events, trade data) rarely names the specific pair directly. Keyword-based filtering would produce excessive false negatives, so FX articles proceed straight to LLM classification.

The 5-stage sentiment pipeline

Every ingested article passes sequentially through five stages. The two-stage prefilter eliminates irrelevant content at low cost before the more expensive deep-analysis stages are invoked.

1

Algorithmic Prefilter

3-tier keyword matching (safe identifiers / industry terms / reject) applied to title and snippet. Scope: crypto and commodity only — FX excluded by design.

2

LLM Prefilter

Relevance classification on title + snippet (≤400 chars). Resolves ambiguous cases keyword matching cannot handle.

3

Content Fetch

Full HTML extraction via headless browser crawler. TDM opt-out declarations are respected.

4

LLM Deep Analysis

Full-article scoring across 6 dimensions: generalSentiment (–1 to +1), plus quality, credibility, importance, market_impact, and influence scores (0–1 each). Per-symbol sentiment extracted for every tracked asset mentioned.

5

VADER Cross-Validation

Rule-based lexicon analysis as a deterministic consistency anchor. Compound score: s / √(s² + 15), clamped to [–1, +1]. Constants: B_INCR=0.293, negation=–0.74, ALL-CAPS=0.733.

Composite signal formula

The production signal combines two validated components at empirically derived weights, computed hourly per tracked symbol:

composite(t) = 0.585 × z(article_volume[t])
             + 0.415 × is_extreme(sentiment[t])

S1 — Article Volume Z-Score

Weight: 0.585

All-crypto article volume z-score normalized over a rolling 30-day window. A spike in article volume — regardless of direction — precedes BTC volatility increases at a +2h lag.

S2 — Sentiment Extreme Flag

Weight: 0.415

Binary flag triggered when BTC sentiment reaches the 95th or 5th percentile of the rolling 30-day distribution. Extreme sentiment episodes exhibit a contrarian property at +24h.

Position sizing: LOW (≤0.30) → 100% exposure  ·  NEUTRAL (0.30–0.70) → 90%  ·  ELEVATED (0.70–1.20) → 75%  ·  HIGH (>1.20) → 50%.

Backtesting results

Sprint 1 · n=1,393h · 90-day window

Volume–Volatility Correlation

All-crypto article volume predicts BTC realized volatility. Peak Pearson r = 0.211 at +2h lag (p<0.001, n=1,393). Statistically significant at lags +0h through +8h. ETH spillover: r=0.201 at +2h.

Sprint 1 · Contrarian signal

Extreme Sentiment Reversal

Bullish extremes (top 5%, 64 events): 72.1% reversal at 24h, avg return –1.02%. Bearish extremes: 53.3% reversal — a notable asymmetry between fear and euphoria signals.

Sprint 2 & 3 · Walk-forward

Out-of-Sample Validation

3-fold (60d train / 30d test): avg out-of-sample r = 0.181, accuracy 53.3%, Sharpe 5.81. BTC multi-asset: r=0.189, accuracy=57.3%, Sharpe=7.46 → GO status.

Disclaimer: Past backtest results do not guarantee future performance. Correlation figures are from validated backtesting windows as of April 2026. This content is for informational and research purposes only and does not constitute financial advice. Full methodology details are available in the downloadable PDF whitepaper.