What is quantitative stock analysis?

Quantitative stock analysis uses mathematical models, statistical methods, and machine learning algorithms to evaluate securities. StoQuant combines 93 features including technical indicators, fundamental ratios, sentiment signals, and macro data to generate alpha scores for 822 stocks.

How does the hidden gem screener work?

The hidden gem screener uses the Benjamin Graham intrinsic value formula (V = EPS x (8.5 + 2g)) to identify undervalued small-cap stocks. It filters for P/E under 15, P/B under 1.5, PEG under 1.0, debt-to-equity under 0.5, and flags stocks trading at least 30% below intrinsic value with fewer than 8 analyst coverage.

What machine learning models power the predictions?

StoQuant uses a 3-model stacking ensemble: LightGBM, CatBoost, and XGBoost with a logistic regression meta-learner. Predictions are calibrated with Venn-ABERS conformal prediction. The system uses 93 features spanning technical, fundamental, sentiment, macro, and alternative data signals.

How does portfolio optimization work?

StoQuant integrates ML price forecasts with the Black-Litterman portfolio optimization model. ML predictions serve as investor views (Q vector) with model confidence intervals mapped to uncertainty (omega matrix). This produces optimal portfolio allocations that blend market equilibrium with AI-driven insights.

Is my portfolio data secure?

Yes. StoQuant uses JWT authentication with timing-safe comparisons, rate limiting (300 req/min per IP), secure headers (HSTS, X-Frame-Options, X-Content-Type-Options), and portfolio ownership verification on every request. No sourcemaps are exposed in production.

What data sources does StoQuant use?

StoQuant aggregates data from SEC EDGAR filings, FMP financial data, social media sentiment (Reddit, StockTwits, Twitter), Google Trends, insider trading records, institutional holdings, options flow, and macro economic indicators from FRED. All data is stored as append-only time series.

Does StoQuant work with Claude Desktop or Claude Code?

Yes. StoQuant ships an official Model Context Protocol (MCP) server, stoquant-mcp, available on npm. One command — npx -y stoquant-mcp install — wires Claude Desktop, Claude Code, or any MCP-compatible client into roughly 30 read-only stock-research tools: prices, Q-Score, ML predictions, hidden-gem screener, insider trades, market regime, and macro dashboards. Power-tier subscription required.

Methodology: How the StoQuant Alpha Score Is Built

The StoQuant alpha score is a calibrated probability that a stock will outperform its sector over a defined forward horizon. It is constructed from 93 engineered features and a three-model stacking ensemble.

Nine alpha dimensions

Valuation — P/E, P/B, PEG, EV/EBITDA, Graham intrinsic-value discount.
Quality — ROE, ROIC, gross margin stability, Piotroski F-score.
Momentum — 1-, 3-, 6-, and 12-month risk-adjusted returns.
Growth — revenue, EPS, and FCF CAGR with stability penalties.
Sentiment — credibility-weighted social signal (Reddit, StockTwits, Twitter) plus FinBERT on news and earnings calls.
Macro — HMM regime probabilities, yield curve slope, credit spreads, FRED indicators.
Alt-data — Google Trends, insider filings, institutional flow, options skew, short interest.
Forensic — Beneish M-score, accrual ratios, cash-flow-to-earnings divergence.
Cross-sectional — sector- and size-adjusted ranks on every above signal.

Model ensemble

Three gradient-boosted learners — LightGBM, CatBoost, and XGBoost — are trained on disjoint time windows. A logistic-regression meta-learner stacks their out-of-fold predictions. Raw probabilities are calibrated with Venn-ABERS (Vovk, 2015) to produce conformal confidence intervals.

Validation protocol

Strict walk-forward: training window slides forward, no look-ahead, no peek at future targets. Transaction costs 0.1–0.5% per trade, slippage 1–2¢ on small-caps. Benchmark: Russell 2000 buy-and-hold. Out-of-sample results are appended — we never overwrite history.

Citations

Graham, Security Analysis (1934); Black & Litterman, Global Portfolio Optimization (1991); Vovk, Venn-Abers Predictors (2015); Loughran & McDonald, When Is a Liability Not a Liability? (2011).