特朗普与习近平北京峰会——波音订单、农业协议与AI芯片僵局

Each sub-score is normalized 0-100 with a 50 baseline. Fundamental rewards margin quality, FCF generation, and valuation gap to consensus price target. Technical rewards trend persistence, support/resistance proximity, and momentum durability while penalizing parabolic extension (Sornette log-periodic). Sentiment combines IUX24 analyzed_assets sentiment scores from the trailing 7 days with WebSearch headline tone. Smart money weights insider buy/sell ratio (last 100 transactions), institutional accumulation/distribution proxy from volume-on-balance, and 13F change momentum. Weights are 35/25/25/15 — fundamental dominates because at 5-day horizons earnings-anchored names are less subject to mean-reversion shocks.

Each article is parsed for sentiment (-1 to +1) and impact magnitude (0 to 1) using a combination of IUX24's analyzedData field where available and a separate keyword-mechanism overlay where not. Volume-weighting multiplies sentiment by impact and aggregates across the trailing 72-hour window with exponential decay (half-life 36h). Scores above +0.5 are 'strong bull' (3+ confirming direct headlines), below -0.5 'strong bear', and the (-0.2, +0.2) band is 'neutral'. We exclude analyst price-target changes from this score to avoid double-counting with Module 1's fundamental sub-score.

CAR (Cumulative Abnormal Return) measures the stock-specific return after removing market-wide drift. We compute it as: CAR_t = sum from event day 0 to t of (R_stock,t - R_benchmark,t), where R_benchmark is the S&P 500 daily return. The 5-day CAR captures the immediate market reaction and 20-day CAR captures the post-event drift. Statistically significant CAR (>1.5x daily volatility) over 5 days suggests genuine information content; over 20 days suggests potential post-earnings announcement drift (PEAD) that institutional reactor positioning has not fully arbitraged.

Momentum forecast fits an OLS log-return regression over the trailing 60 sessions and extrapolates the trend forward. R² above 0.6 indicates a stable trend regime (Jegadeesh-Titman momentum precedent). Mean-reversion uses a Bollinger-z-score: at |z| > 2.0, the historical 60-day reversion half-life is approximately 12 trading days. Names with both high momentum R² and high |z| (QCOM at R²=0.81 and z=+2.37) are momentum-extended and create the largest spread between the two forecasts — typically resolved by the mean-reversion outcome over a 5–20 day window. ATR(14) is used for position sizing to make volatility-equivalent dollar risk across very different price levels.

Insider buy/sell ratios are computed from the most recent 100 Form 4 filings per ticker (FMP insider-trading endpoint). We treat 10b5-1 routine sells as baseline noise but flag any insider buys as positively informative (Lakonishok-Lee 2001). OBV is on-balance volume — a cumulative volume measure that adds volume on up-days and subtracts on down-days; rising OBV confirms accumulation. Factor exposures use sector percentile rankings (P/B vs sector members from FMP peers endpoint). The alpha decay curve assumes a 5-day half-life consistent with event-driven catalyst trades (Bernard-Thomas 1989 PEAD generalization) — expected residual alpha decays exponentially after the announcement landing.

Position size = (Portfolio risk %) / (Entry - Stop) × Portfolio value. We target 1.5-2.5% portfolio risk per single name, biased larger toward high-conviction longs (BA, LNG, BLK at 2.0-2.5%) and smaller toward shorts (NVDA, QCOM at 1.5% each, with NVDA+QCOM treated as a correlation-adjusted single 3% allocation). ATR(14)-derived stop distances ensure equivalent dollar risk across very different price levels. The 5-day half-life on alpha decay drives the rationale for partial profit-taking at T1 and full exit by T+10 unless materially new information emerges.

Jarque-Bera tests joint null of skewness=0 and excess-kurtosis=0 (normality); p < 0.05 rejects normality. Ljung-Box tests for autocorrelation up to 10 lags; p < 0.05 suggests momentum/mean-reversion structure exploitable by directional strategies. VaR is computed via historical simulation over the trailing 60 sessions — the 95% VaR is the 5th percentile loss, the 99% is the 1st percentile. CVaR (Expected Shortfall) is the average loss conditional on exceeding the VaR threshold; it captures the magnitude of tail risk better than VaR alone. Names with non-normal distributions (BA, VG, NVDA, QCOM) require position sizing buffers above the raw VaR figures.