Documentation
Methodology
How Silicon Analysis calculates Alpha scores, tracks signals, and validates model performance.
Data Sources
News & Sentiment
News is ingested from NewsAPI and GNews, filtered for relevance to tracked companies. Each article is processed by Claude LLM to extract sentiment scores (-1 to +1), impact scores (0-100), severity classification, and supporting evidence snippets. All claims require citations from the source article.
Stock Prices
Real-time and daily prices are fetched from Finnhub (primary) and AlphaVantage (fallback). Price staleness is tracked and surfaced in the UI. Prices older than 15 minutes during market hours are marked as stale.
SEC Financials
Quarterly financials are pulled from SEC EDGAR via the company facts API. We track revenue (TTM), gross profit, operating income, R&D expense, and free cash flow proxy (operating cash flow - capex). CIK numbers are resolved automatically from tickers.
Alpha Score (v3.0 Quantamental)
Alpha v3 is a quantamental signal combining technical momentum, news sentiment, and fundamentals. It uses risk-adjusted returns, evidence-gated news, and valuation circuit breakers to identify quality opportunities.
News Signal (35%)
Time-decayed sentiment with evidence gating. News older than 72 hours receives exponentially less weight. Low-confidence or low-evidence sentiment is damped toward neutral.
Risk-Adjusted Momentum (30%)
20-day trailing return divided by 20-day annualized volatility (Sharpe-like). Category-relative percentile ranking rewards consistent, low-volatility gains.
Fundamentals (FSI) (20%)
Financial Strength Index (FSI) based on operating margin, FCF margin, revenue growth, and R&D intensity. Percentile-ranked within category.
Data Confidence (15%)
Data quality score penalizing stale financials (>180d), stale prices (>24h), and low news coverage. Ensures high-confidence signals.
Valuation Circuit Breaker
Companies in the top 25% of P/S ratio (category-relative) receive a 0-20 point penalty. Companies with FCF yield >5% and reasonable valuations receive up to 10 bonus points. This prevents chasing momentum into extreme valuations.
Composite Score (v2.0 Evidence-Only)
Composite v2 removes static/manual fields from v1 and uses only fact-based, evidence-backed components. It combines Alpha score with financial strength and evidence quality for a cleaner overall signal.
Alpha Score (50%)
The v3 Alpha score captures momentum and sentiment signals.
Financial Strength Index (FSI) (30%)
SEC-derived fundamentals: margins, FCF, R&D intensity.
Evidence Quality (20%)
News coverage, average LLM confidence, and evidence snippet count.
Valuation Score & Financial Strength Index
Valuation Score
Category-relative percentile ranking based on P/S (TTM) and FCF Yield. A score of 80 means the company is cheaper than 80% of peers in its category. Categories (AI_HARDWARE, CLOUD_INFRA, FOUNDRY, MODEL_LAB, INFRA_SAAS) have different valuation norms, so we compare within-category only.
Financial Strength Index (FSI)
Composite of profitability metrics: gross margin, operating margin, R&D intensity, and FCF generation. Higher FSI indicates stronger fundamentals. Used as a quality filter alongside the Alpha signal.
Signal Health & Backtesting
We continuously validate signal quality through daily backtests:
- Information Coefficient (IC): Spearman rank correlation between signal values and forward returns. IC > 0.05 indicates predictive power.
- Spread: Return difference between top and bottom decile by signal. Positive spread means the signal separates winners from losers.
- Net Spread: Spread minus estimated transaction costs (turnover × spread cost). Must be positive for the signal to be tradeable.
- Turnover: Fraction of names changing in top decile day-to-day. Lower turnover means more stable signal and lower trading costs.
Signal health metrics are displayed on report pages and the system performance dashboard.
Known Limitations
- Survivorship bias: Our universe only includes currently-tracked companies. Failed or delisted companies are not in historical backtests.
- News lag: News processing has ~15 minute latency from publication to score update. Intraday signals may be stale.
- Small universe: With ~50-60 companies, statistical significance of backtest results requires caution. We report sample sizes alongside metrics.
- LLM variance: Sentiment extraction via LLM may have day-to-day variance. We use low temperature (0.3) and structured output validation to reduce this.
- No transaction costs in Alpha: The Alpha score itself does not account for trading costs. Net Spread in backtests does.