Intelligence Network Methodology

• 172 curated RSS feeds across 13 content lanes, shuffled randomly each run for source diversity
• NewsAPI.ai (Event Registry) — short keyword queries per lane, returning global wire/agency articles (Reuters, Bloomberg, AP, etc.)
• Mediastack — business news API covering the 3 highest-value lanes (Healthcare, Finance, Industrials)
• 13 content lanes: Healthcare M&A, Industrials & Defense, GCC/LATAM Geopolitics, Healthcare Marketing, World/Geopolitical, Technology, AI/ML, Finance, Government & Regulatory, Intelligence & Security, Deep-Dive (Foreign Policy, Intelligence & Defense)
• Source diversity cap: max 3 items per domain to prevent any single outlet from dominating

• Named entity extraction via regex on signal titles (capitalized multi-word phrases)
• Lane classification per topic (Healthcare, Industrials, GCC/LATAM)
• Coverage gap detection against published article corpus
• Auto-pilot: max 2 pipeline triggers per scan for high-acceleration uncovered topics

• First 4,900 characters processed per signal (NER API limit)
• HTML stripped to plain text before analysis
• Each signal gets a unique hash-based key so co-occurrence reflects real editorial relationships
• Entities deduplicated and upserted with mention counts and signal/article references
• Currently tracking 1,800+ entities across 200+ unique signal sources

• BICS Level 1 (Sector): Health Care, Industrials, Financials, Technology, Energy, Materials, Communications, Consumer Discretionary, Utilities, Government
• BICS Level 2-4: Industry Group, Industry, Sub-Industry (e.g., Health Care → HC Equipment & Services → HC Facilities & Services → Hospitals)
• Deal context: acquirer, target, regulator, competitor, partner, investor, subsidiary
• Geography: headquarters city/country
• Market cap tier: mega-cap through private/government
• Relationships: up to 5 related entities with relationship type
• Re-enrichment: weekly with latest signal titles for temporal accuracy

• Entity Momentum Scorer: Weighted composite of 9 features — mention velocity, article breadth, relationship density, deal context presence, recency, market cap, enrichment completeness, mention concentration, and entity type. Produces a 0–100 momentum score per entity. Drives node sizing in the visualization.
• Sector Clustering (KMeans): Unsupervised clustering on entity features (BICS sector, geography, market cap, deal activity, mention volume). Optimal K selected via silhouette score. Reveals data-driven entity groupings — e.g., “resource geography” clusters (Permian Basin, Haynesville Shale) or “deal-active heavyweights” (UPMC, Boeing, GE Aerospace) that span multiple BICS sectors.
• Retrain schedule: Weekly (Sundays 05:00 UTC). Auto-triggers chart re-render on completion.
• Rising Entities panel: Top 10 entities by momentum score surfaced on the intelligence page, giving investors an at-a-glance view of where attention is accelerating.

• Momentum-weighted topic selection: The Trend Acceleration Engine weights topics using a blended score: 60% signal acceleration + 40% ML momentum. An entity at momentum 85 with 3× acceleration outranks one at momentum 10 with 5× acceleration — ensuring the pipeline prioritizes structurally important topics over noise spikes.
• Article scoring bonus: After AI evaluates an article draft on 6 quality dimensions, the scoring Lambda checks whether the article covers high-momentum entities. Articles mentioning 2+ top-20 momentum entities receive a +1 point bonus; coverage of 1 top-20 or 3+ top-50 entities yields +0.5. This makes articles about rising entities less likely to be rejected.
• Social post enrichment: Every social post generation prompt is injected with a momentum context block identifying which high-momentum entities appear in the source article. The model is instructed to prioritize naming these entities. When entities from different ML clusters co-occur, the prompt flags this as a cluster-crossing signal — directing the model to emphasize the cross-sector convergence.
• Closed-loop retraining: As new articles are published and entities accumulate mentions, the weekly retrain captures updated momentum patterns. Topics that were rising last week but plateau will naturally decay in momentum, while newly accelerating entities surface automatically.

• Vertices: enriched entity profiles (name, BICS classification, deal context, geography)
• Edges: co-occurrence relationships with signal/article-count weights
• Enrichment-derived edges: relationships inferred by AI during BICS enrichment are also used as graph edges, pulling related entities into the network even without co-occurrence
• Degree centrality computed per node (connection count)
• Centrality movers: rising and falling centrality tracked day-over-day with tooltips showing degree change
• New edge detection: daily diff against previous graph state
• Currently: ~877 nodes, ~3,800 edges

• Daily graph.json snapshots archived to S3 for historical comparison
• New edge detection: set difference between today's and yesterday's edge sets
• Centrality drift: tracked via degree count changes across snapshots
• Entity staleness: last_enriched timestamp ensures profiles stay current (<7 day TTL)
• Future: temporal edge decay weighting, rolling 30-day network animation

• Input: top 10 entities by centrality, new edges in 24h, sector distribution
• Output: 3 paragraphs — dominance, cross-sector, momentum
• Tone: Bloomberg terminal style, no jargon
• Audience: institutional investors, PE professionals, corporate strategists

• Force simulation: forceLink (distance by weight), forceManyBody (charge), forceCollide (collision detection)
• Node sizing: ML momentum score (0–100), square root scale for perceptual accuracy
• Node coloring: ML cluster assignment (KMeans), sector fallback for unclustered
• Node styling: 3-stop radial gradients, cluster-colored borders, feGaussianBlur glow on high-centrality
• Interactions: hover highlighting (connected nodes stay, others fade), drag, zoom/pan
• Sector filter pills: remove/add nodes from simulation dynamically
• Label pills: dark background behind text for readability on complex graphs
• Rising Entities panel: top 10 by momentum score with hover tooltips explaining the momentum formula
• Centrality Movers panel: rising and falling entities with degree-change tooltips showing connection gains/losses
• Entity Watchlists: users can subscribe to watch specific entities with 4 alert triggers (momentum threshold, momentum delta, cluster change, new connection)
• Cross-cluster article trigger: when new edges cross ML cluster boundaries with momentum > 50, the system auto-generates articles (max 2/day)
• Pre-rendered: zero compute at request time, 1-hour cache TTL

Five orthogonal input signals:

• Attention Cascade Intensity — Hawkes process residuals. Measures surprising coverage volume after accounting for self-excitation (one story triggers more stories). The residual is the unexpected component — the attention that cannot be explained by prior attention.
• Narrative Embedding Drift — Vector embeddings cosine distance between this week’s entity-level text centroid and last week’s. Measures what is being said changing, independent of how much is being said.
• Graph Spectral Shift — Eigenvalues of the graph Laplacian tracked daily. Captures structural community reorganization — merging clusters, splitting groups, new bridge entities. Orthogonal to node-level metrics by construction.
• Sentiment-Momentum Divergence — Residual of targeted sentiment regressed on momentum score. The part of sentiment not explained by attention volume. An entity getting quietly negative while staying prominent is a contrarian signal.
• Source Concentration — Herfindahl-Hirschman Index (HHI) across source domains per entity. High HHI means one source is driving the narrative. Low HHI means consensus is forming across multiple outlets.

Signal processing pipeline:

• Orthogonalization: PCA decomposition on the 5-signal matrix. Components with eigenvalue > 1 retained. VIF screening rejects any input with VIF > 5 — ensuring no signal is a linear combination of the others.
• Regime Detection: Bayesian Online Changepoint Detection (BOCPD, Adams & MacKay 2007). Maintains a posterior distribution over run length. On changepoint detection, PSI enters a 7-day transition window with widened confidence bands and a visible “regime shift detected” flag.
• Composite: Inverse-variance weighted sum of orthogonal components. More precise signals receive higher weight automatically — no manual tuning.
• Smoothing: Kalman filter separates three components: trend (reported as PSI), mean-reverting deviations (trigger alerts), and noise (discarded). Produces native confidence intervals on every estimate.

Output & calibration:

• Z-score with plain-English labels: < −1.0 “Unusually Quiet” · −1.0 to +1.0 “Normal” · +1.0 to +2.0 “Elevated” · +2.0 to +3.0 “High Activity” · > +3.0 “Extreme”
• Calibration: First 30 days use percentile ranking (non-parametric). Day 30: full model trains and validation suite runs. If tests fail, PSI stays in calibration mode until passing.
• Audit trail: Every daily computation writes raw signals, PCA loadings, regime state, Kalman estimate + confidence interval, feature attribution, and driving entities to the database. Every PSI value traces back to source feeds and timestamps.

Card metrics:

• → Momentum Score — Existing 0–100 weighted composite. Full weight at 30+ observations, linear ramp below.
• → Sentiment Trajectory — 7-day sparkline from targeted sentiment EMA. Suppressed below 5 datapoints (direction arrow only).
• → Network Position — Personalized PageRank conditioned on sector (not raw degree centrality — PPR captures influence relative to sector context). Requires 5+ edges.
• → Narrative Drift — Vector embedding cosine distance week-over-week. Labeled “emerging” below 4 weeks of data.
• → Sentiment-Momentum Divergence — The contrarian indicator. When sentiment and momentum decouple, something is changing beneath the surface.

Eligibility & confidence:

• → Eligibility thresholds: 5+ co-occurrence edges, 3+ signal appearances in 14 days, momentum > 0
• → Confidence indicator: Each metric shows data sufficiency — “dense data” vs. “emerging” — so readers can assess reliability at a glance
• → PSI watermark: Every card displays the current PSI value in the corner, connecting the entity-level view to the market-level signal

Five validation gates:

• Permutation test (p < 0.01, 1,000 shuffles) — Proves the signal is not random noise. If shuffled inputs produce equivalent output, the signal has no information content.
• Subsample stability — Consistent direction across 3 non-overlapping time periods. A signal that works only in one sub-period is overfit to that regime, not generalizable.
• Walk-forward cross-validation — Train on months 1–N, test on month N+1. Signal must show positive information coefficient in >60% of out-of-sample windows. No look-ahead bias.
• Factor regression — Significant alpha after controlling for raw news volume and sector momentum. Proves PSI contains proprietary information, not repackaged public data.
• Decay analysis — Characterized half-life. Narrative signals typically decay 1–5 days for breaking news, 2–8 weeks for structural shifts. PSI must have a well-defined decay profile to be actionable.

Cold start & confidence:

• Hierarchical Bayesian priors: New entities borrow baseline distributions from their BICS sector cohort. As data accumulates, the entity’s posterior migrates toward its own data (James-Stein shrinkage). No entity starts from zero.
• Confidence ramps: Below 30 observations, signal weight scales linearly (n/30). Below 10 observations, the signal is suppressed entirely. No thin-data artifacts.
• No black boxes: Every PSI value traces back to source signals, feed URLs, and timestamps. The audit trail is the proof. If a value cannot be explained, it is not published.

• Composite of 5 orthogonal signals, each measuring a distinct dimension of entity activity
• Six processing layers: raw signals → PCA orthogonalization → BOCPD regime detection → inverse-variance weighting → Kalman smoothing → z-score output
• Published as a labeled score on every article and entity intelligence card
• Regime classification from QUIET to EXTREME provides immediate interpretive context
• Auto-validates at day 30 via five independent statistical gates; weekly thereafter

• Models coverage volume as a Hawkes self-exciting process: each mention increases the probability of subsequent mentions
• Computes the residual between observed coverage volume and the volume predicted by the entity’s own historical excitation pattern
• A high residual indicates the entity is attracting more attention than its own momentum predicts — an external catalyst, not self-sustaining noise
• Interpretation: positive residual = structural attention surge; negative = momentum decay

• Generates weekly entity description centroids using text embeddings (1536-dimensional)
• Computes cosine distance between consecutive weekly centroids
• High drift signals that the story surrounding the entity is changing direction — reputational shift, strategic pivot, or emerging deal narrative
• Low drift over time = stable narrative; sudden high drift = context break worth investigating

• Applies Laplacian eigenvalue analysis to the entity co-occurrence graph snapshot
• Tracks changes in the Laplacian spectrum week-over-week to detect structural changes in community topology
• Captures cluster merges, splits, and entity migrations between communities that are invisible to node-level metrics
• Interpretation: a rising spectral gap indicates strengthening community structure; a collapsing gap signals structural fragmentation

• Runs OLS regression of sentiment on momentum across the entity corpus to establish the expected sentiment given a given momentum level
• Computes the residual: actual sentiment minus predicted sentiment
• A large positive residual means sentiment is running ahead of what momentum would predict — potential contrarian signal
• A large negative residual means sentiment is deteriorating despite strong momentum — early-warning divergence often preceding corrections

• Computes the Herfindahl-Hirschman Index across news source domains covering the entity
• High HHI (coverage concentrated in few outlets) indicates potential information asymmetry — institutional or specialist sources breaking stories before broader dissemination
• Low HHI (broad source diversity) indicates consensus coverage — widely known, low information edge
• Contrarian interpretation: high-HHI entities may offer earlier signal; low-HHI entities are already in consensus

• Layer 1 — Raw Signal Computation: Each of the five signals computed independently per entity using signal history and graph state
• Layer 2 — PCA Orthogonalization: Kaiser criterion (eigenvalue > 1) retains principal components; signals with VIF > 5 are rejected to eliminate multicollinearity. Ensures composite score reflects genuinely independent information
• Layer 3 — BOCPD Regime Detection: Bayesian Online Changepoint Detection (Adams & MacKay, 2007) using Normal-Inverse-Gamma conjugate prior. Identifies structural breaks in the signal distribution with a 7-day transition window, allowing the model to adapt to new regimes without overfitting to noise
• Layer 4 — Inverse-Variance Weighted Composite: Signals are combined using inverse-variance weights, with weights updating automatically as signal reliability evolves. More stable signals receive proportionally higher weight
• Layer 5 — Kalman Smoothing: Harvey (1989) local linear trend model with Rauch-Tung-Striebel backward pass. Decomposes the composite into trend, mean-reverting, and noise components. Produces a smoothed score that lags genuine transitions minimally while filtering measurement noise
• Layer 6 — Z-Score Output: Kalman-smoothed composite standardized against the rolling entity-level distribution. Output is a dimensionless z-score with regime label assigned per the classification table

• James-Stein shrinkage: raw signals are shrunk toward BICS sector priors derived from all entities in the same sector. New entities inherit partial sector-level expectations rather than starting from zero
• Linear confidence ramp: PSI confidence scales linearly from 0 to 1 over the first 30 observations (n/30). Score is suppressed entirely below 10 observations to prevent spurious early signals from surfacing in article metadata
• Percentile calibration: during the first 30 days, raw z-scores are converted to sector-relative percentile ranks rather than absolute z-scores, providing interpretable output before full distribution estimation is possible
• Sector inheritance: until an entity has sufficient history, its regime classification defaults to the sector median PSI, clearly labeled as an estimated value

• Momentum: ML momentum score (0–100), the entity’s weighted composite of 9 behavioral features including mention velocity, deal activity, and relationship density
• Sentiment Trajectory: 7-day rolling directional change in targeted entity-level sentiment, extracted via NLP sentiment analysis on per-signal text windows
• Network Position (PPR): Personalized PageRank score measuring structural importance in the co-occurrence graph — entities connected to other well-connected entities score higher
• Narrative Drift: S2 signal value (cosine distance between weekly description centroids) indicating how much the entity’s story is changing
• SMD: Sentiment-Momentum Divergence residual (S4), flagging when sentiment departs from what momentum predicts

• Minimum 5 graph edges (sufficient network embeddedness for PPR to be meaningful)
• At least 3 distinct signals received in the prior 14 days
• Momentum score above 0 (entity must show at least baseline activity)

• Gate 1 — Permutation Test: PSI scores are permuted against entity outcomes 1,000 times. The observed information coefficient must exceed the 99th percentile of the null distribution (p < 0.01) to pass
• Gate 2 — Subsample Stability: The full entity corpus is split into three temporal subsamples. Signal weights must be directionally consistent across all three periods to confirm robustness rather than sample-specific fitting
• Gate 3 — Walk-Forward Cross-Validation: Rolling origin backtests with expanding training windows. PSI must achieve positive information coefficient (IC) on more than 60% of out-of-sample periods to pass
• Gate 4 — Factor Regression: PSI is regressed on news volume and sector momentum to isolate alpha. The model must retain significant explanatory power after controlling for these known confounders
• Gate 5 — Decay Analysis: Signal half-life is estimated by regressing IC on forward horizon. A half-life below 3 days triggers a weights recalibration; below 1 day triggers a full signal review

• Three-model ensemble: graph heuristics (common neighbors, Jaccard similarity, Adamic-Adar index), logistic regression on entity feature pairs, and a Graph Convolutional Network (GCN) trained on the co-occurrence graph topology
• Feature inputs: BICS sector alignment, deal context overlap, momentum score proximity, PPR proximity, geographic co-location, shared signal lane history
• Validation: 20% edge holdout test split from the accumulated graph history. Ensemble precision, recall, and AUC-ROC evaluated monthly
• Output: A daily list of high-probability predicted edges surfaced to the intelligence dashboard as “Emerging Connections” — relationships that do not yet exist in the graph but are structurally likely within a 7–30 day window
• Status: Accumulating 3 months of edge history for initial training; target production deployment Q2 2026

• Signal decomposition: Per-entity breakdown of which of the five signals contributes most to the current PSI score, enabling attribution analysis (e.g., “PSI is ELEVATED primarily due to Narrative Drift and Source Concentration”)
• Sector rotation heatmap: Aggregate PSI by BICS sector, updated daily, showing which sectors are in ELEVATED or HIGH regimes simultaneously — a macro-level intelligence view for portfolio allocation decisions
• Lead-lag detection: Cross-correlation analysis between connected entity PSI series identifies which entities systematically lead or lag their peers in the co-occurrence graph — useful for identifying bellwether entities within a sector
• Peer-relative PSI: Entity PSI expressed as a z-score versus the sector median, normalizing for sector-wide regime shifts and isolating entity-specific signal from broad market conditions
• Event study framework: Tracks PSI behavior in the 30 days before and after confirmed events (deal announcements, regulatory decisions, earnings) to calibrate signal lead times and validate predictive utility
• Clinical trials bridge: Trial status transitions (Phase I → II, enrollment open/closed, primary endpoint readout) are ingested as discrete signals, enabling PSI to reflect pipeline events for healthcare entities not yet captured in general news flow

Individual signals (Hawkes processes, spectral analysis, HHI) are well-documented in academic literature. PSI’s contribution is the specific five-signal combination with PCA orthogonalization, Bayesian changepoint detection, and Kalman smoothing as a unified pipeline. No published system combines these methods for healthcare intelligence.

The VIF rejection step ensures signal independence — adding a sixth signal (e.g., clinical trial transitions) only strengthens the composite if it provides genuinely new information.

The entity co-occurrence graph underlying PSI took months of continuous ingestion from 172 RSS feeds, NewsAPI.ai, and Mediastack to construct. With 400+ active nodes, 800+ edges, and daily temporal history stored in the database, the graph captures relationship dynamics that cannot be replicated from static data.

A competitor starting today would need months of identical signal processing to build a comparable graph.

Every day PSI runs, the model becomes more valuable:

• Kalman smoother improves estimates with each observation
• BOCPD regime detection requires history to identify changepoints
• Lead-lag detection needs 7+ days of paired PSI scores
• Event study backtesting grows more statistically powerful with each deal event
• Link prediction accuracy improves as edge formation patterns accumulate
• Validation gates (day 30 auto-fire) build auditable statistical evidence

A 30-day PSI has materially different capabilities than a day-1 PSI. At 90 days, the link prediction model gains enough edge history for temporal features. At 180 days, event study results carry statistical significance. This temporal moat widens continuously.

The link prediction system ensembles six models spanning three methodological families:

• Classical graph theory: Jaccard, Adamic-Adar
• Feature-based machine learning: logistic regression on PSI features
• Deep graph learning: GCN, Node2Vec, VGAE, TransE knowledge graph embeddings

Each model captures different structural patterns. When 5 or 6 models agree on a prediction, the confidence is qualitatively different from any single model’s output. This ensemble architecture is standard at institutional quantitative firms but unprecedented in healthcare intelligence.

PSI incorporates non-news signals that create information asymmetry:

• SEC EDGAR filing detection: 8-K material events, 13-D activist disclosures, S-1 IPO registrations
• Search velocity acceleration: mention momentum from 24,000+ daily signals
• Clinical trial status transitions: recruiting → completed, suspended, terminated

These signals arrive hours to days before news coverage, providing temporal advantage over news-only intelligence systems.