Plocamium Holdings

What PSI Measures

The Plocamium Signal Index (PSI) is a daily composite score measuring narrative disruption across the entity graph. It is not a sentiment score. It is not a volume count. PSI measures how much the graph’s behavior deviates from its own learned baseline.

How to Read a Score

The PSI score is mapped to five labeled regimes, providing immediate interpretive context for non-quant readers. Regime labels appear on entity dossiers, cards and reports.

PSI regime classification thresholds
Regime PSI range Interpretation
QUIET below -1 Below-normal signal activity across all five dimensions
NORMAL -1 to +1 Baseline market conditions; no anomalous signal detected
ELEVATED +1 to +2 Heightened activity across signal dimensions; monitor closely
HIGH +2 to +3 Significant regime shift detected; multiple signals converging
EXTREME above +3 Rare event; maximum signal intensity across all dimensions

Entities with fewer than 10 daily observations have no PSI score and no regime. They are shown as Insufficient data, never as NORMAL with a score of 0, and are left out of the market PSI.

What the Evidence Supports

Each study has explicit limits. The notes behind these results are under Studies in Full.

Claims we support

  • GBM leads forward link holdout (Phase-0 diagnostic, 2026-06-22, not externally anchored: mean AUPRC gbm_psi 0.725 / GBM 0.690 vs Node2Vec 0.594)
  • Production link ensemble is predictively Node2Vec-only (leave-one-out contribution 100%, last run 2026-10-29)

Claims we cannot support

  • PSI predicts deal announcements (not shown: the one external test failed its rule, and a later review found it could not settle the question)
  • Attention Diagnostic Suite “passes” (3 of 5 as of 2026-11-15)
  • GNN attention beats GBM on forward link holdout
  • GPU val AUC alone justifies GNN promotion (the 2026-06-21 GPU run reached val AUC 0.7217, but its holdout was 0.4286 against prod 0.4293)

Latest diagnostic run (2026-11-15): 3 of 5 probes passed

  • Permutation test: PASS (p=0.003; r=+0.20 vs forward mention growth, n=195)
  • Subsample stability: PASS
  • Walk-forward CV: FAIL (39% positive-IC windows; threshold >60%)
  • Factor regression: FAIL (PSI β −0.33 after news volume and sector momentum controls)
  • Decay analysis: PASS (half-life 17 observations)

What this suite cannot claim

A 4/5 or 5/5 pass would not authorize deal-announcement forecasting. LF-1 attempt #2 (rule fixed 2026-06-08) failed its rule: PSI hit rate 10.8% (CI 4.8-18.1%) vs a comparison rate of 49.4% (CI 39.8-60.2%), n=83 evaluable events. A November 2026 review found the comparison was not a coverage spike: it read a running total of mentions, and most of its hits came from events with no earlier data to compare against. Two companies also supplied most of the events. Deal forecasting remains unproven, not disproven. Do not conflate internal event-study coverage with external LF-1.

The Five Signals

Note: This section describes the index as designed. The method running today differs in places: sentiment-momentum divergence does not yet feed the score, scores are not yet standardized against each entity’s own history, and they are not yet published with confidence intervals. A corrected method is running alongside it, and the changelog will record the switch.

Each signal is designed to capture a distinct, non-redundant dimension of entity behavior. Orthogonality is enforced at the processing layer via PCA, but the five signals are conceptually independent by construction: attention volume, narrative direction, structural network position, sentiment-momentum divergence, and source breadth.

In brief:

  • 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 can turn quietly negative while staying prominent.
  • 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.

S1: Attention Cascade Intensity

  • 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

S2: Narrative Embedding Drift

  • 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

S3: Graph Spectral Shift

  • 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

S4: Sentiment-Momentum Divergence (SMD)

  • 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
  • A large negative residual means sentiment is deteriorating despite strong momentum. No study on this page tests whether either precedes market moves

S5: Source Concentration (HHI)

  • Computes the Herfindahl-Hirschman Index across news source domains covering the entity
  • High HHI means a few outlets account for most of the coverage
  • Low HHI means coverage is spread across many outlets

How a Score Is Computed

The note under The Five Signals applies here too.

PSI combines five signals into one dimensionless score through six processing layers. It is published daily on entity dossiers and in the subscriber dashboard.

The Six Layers

Raw signals pass through six processing layers before producing a PSI score. Each layer serves a specific statistical purpose: removing redundancy, detecting regime changes, weighting by reliability, and smoothing noise without obscuring genuine transitions.

  • 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; components with VIF > 5 are dropped
  • 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: Score Output. The Kalman-smoothed composite is pulled toward its sector’s average (James-Stein shrinkage) and scaled down for entities with fewer than 30 observations. Output is a dimensionless score with regime label assigned per the classification table

Output and Calibration

  • Calibration: The bands are fixed score thresholds. Separately, the Attention Diagnostic Suite runs weekly to characterize signal structure. Suite results are diagnostic only. They do not block production PSI publication (latest results).
  • 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.

Cold Start

Newly tracked entities have insufficient signal history to produce reliable PSI scores. The cold start protocol applies Bayesian shrinkage and graduated confidence ramps to avoid publishing noisy scores for entities with sparse data.

  • James-Stein shrinkage: an entity’s score is shrunk toward the average of its BICS sector, less as observations accumulate
  • Confidence ramps: Below 30 observations, signal weight scales linearly (n/30). Below 10 observations, an entity has no PSI score or regime and is shown as Insufficient data.

Entity Cards

A data card for each eligible entity, in the subscriber dashboard and PDF reports, shown alongside its PSI score.

  • Momentum Score: 0-100 weighted composite.
  • Network Position: Personalized PageRank. Requires 5+ edges.
  • Narrative Drift: how far this week’s coverage of the entity has moved from last week’s.
  • Sentiment Trajectory: recent daily sentiment. Needs at least 5 days of data.
  • Sentiment-Momentum Divergence: flags when sentiment and momentum move apart. Reads 0 where an entity has too little data.

Until the index moves to its new method, narrative drift, sentiment trajectory and divergence come from that method’s daily shadow run.

Eligibility

  • 5+ co-occurrence edges, 3+ signal appearances in 14 days, momentum > 0

Signal Collection

Plocamium continuously ingests from curated RSS feeds and news APIs spanning content lanes. Each run, every 2 hours, samples a shuffled set of feeds in each lane and takes at most 5 items per feed, so no single feed dominates.

  • Curated RSS feeds: shuffled randomly each run for source diversity
  • News APIs, including NewsAPI.ai (Event Registry): short keyword queries per lane, returning global wire/agency articles (Reuters, AP, AFP, etc.)
  • 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)

Topic Discovery & Trend Acceleration

Twice daily, the Trend Acceleration Engine scans the signal archive to identify topics with accelerating mention velocity. Acceleration is calculated as the ratio of 24-hour mentions to the 7-day daily average. Topics at 3× acceleration or more with no existing coverage automatically trigger the content pipeline.

  • 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 4 pipeline triggers per scan for high-acceleration uncovered topics

Entity Extraction

Entities are now extracted from every individual signal, not just published articles, using natural-language entity recognition. Each signal receives a unique content-hash-based article key (e.g., HEALTHCARE-3a5e65588bb5), ensuring co-occurrence reflects real editorial relationships rather than artificial batching. Entities are extracted with a confidence threshold of >0.85 and classified as ORGANIZATION, PERSON, or LOCATION. Entity extraction runs on every signal ingestion (every 2 hours) and again on published articles.

  • 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

Entity Enrichment (BICS Classification)

Entities on the enrichment target list are enriched via AI classification using the Bloomberg Industry Classification System (BICS), a 4-level institutional taxonomy: Sector → Industry Group → Industry → Sub-Industry. Enrichment runs when articles are published.

  • 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: a brief description of any M&A, investment or deal activity
  • Geography: headquarters city/country
  • Market cap tier: mega-cap through micro-cap, or unknown
  • Relationships: related entities

Machine Learning Models

Two ML models run on the entity corpus, retrained weekly. These models transform the static entity catalog into a dynamic intelligence signal, surfacing which entities are gaining momentum and revealing non-obvious groupings that BICS classification alone cannot detect.

  • Entity Momentum Scorer: Weighted composite of 9 features: mention count, 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).
  • Rising Entities panel: Top 10 entities by momentum score surfaced on the intelligence page.

Momentum scores also influence which topics Plocamium covers and how its articles and social posts are scored and framed.

Network Construction

The entity network is a co-occurrence graph: two entities are connected when they appear in ≥2 shared signals or articles with unique per-signal keys. Edge weight reflects the number of shared items. The graph is kept as daily snapshots. Each news signal gets a unique hash-based key (e.g., HEALTHCARE-3a5e65588bb5) so entities only co-occur when they appear in the same news story.

  • 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: 400 entities and 22,844 connections on 2026-11-17; the map shows the top 150

Variance & Historical Analysis

Network state is archived daily. This enables temporal analysis: which connections are strengthening, which are new, and which entities are gaining or losing centrality over time. The daily insight narrative explicitly highlights new edges and momentum shifts, so readers see what changed, as well as the current state.

  • Daily snapshots archived 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

AI-Generated Daily Insight

Each morning, an AI model writes a short briefing from the day’s data in four parts: Signal (the entities that stand out), Surprise (one new connection, chosen from the day’s real changes), So What (what the PSI reading means) and Action Item (what to research next). It is research, not investment advice.

Visualization

The network is rendered as an interactive D3.js v7 force-directed graph with custom SVG rendering. Nodes are sized by ML momentum score (square root scale) and colored by ML cluster assignment, with BICS sector color as fallback for unclustered entities. The visualization is pre-rendered for fast page loads.

  • 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
  • Interactions: hover highlighting (connected nodes stay, others fade), drag, zoom/pan
  • Sector filter pills: remove/add nodes from simulation dynamically
  • 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 5 alert triggers (momentum threshold, momentum delta, cluster change, new connection, attention spike)
  • 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

Advanced Analytics

Beyond the core PSI score, the intelligence platform supports a suite of advanced analytical layers for clients requiring deeper quantitative context. These modules operate on the same underlying signal infrastructure.

  • Signal decomposition: Per-entity view of which signal reads highest
  • Sector rotation heatmap: Aggregate PSI by BICS sector, updated daily, showing which sectors are in ELEVATED or HIGH regimes simultaneously, showing which sectors are drawing unusual coverage
  • 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, computed daily but not yet shown to clients
  • 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 (internal): Tracks PSI around confirmed events for coverage calibration. Not interchangeable with the LF-1 deal test. Latest results are under Event Study

Studies in Full

Continuous studies measure component contribution, link-prediction generalization, and attention-signal structure. Current results and claims are under What the Evidence Supports.

Ablation Study

Schedule: retired October 2026; figures are from the last run, 2026-10-29. Can claim: marginal AUC contribution per model (Node2Vec 100% in that run). Cannot claim: all eight models contribute equally in production.

Event Study (internal)

Latest run (2026-11-18): 23.7% of 1,566 entity-event pairs had elevated PSI before the event. Average PSI was 0.39 before and 0.30 after.

Can claim: PSI level distribution around known events in corpus. Cannot claim: anticipatory deal-forecast power. LF-1 is that test.

Schedule: daily with enrichment (scheduled). Production link predictions are ranked by Node2Vec similarity. The eight-model blend selects the candidate pairs; its daily weight search was retired in October 2026.

LF-1 Anchored Test (external)

Attempt #2 (rule fixed 2026-06-08) failed its rule. PSI regime vs SEC-filing-dated deals, n=83. A November 2026 review found its comparison rate unsound, so the result shows no evidence of deal forecasting rather than evidence against it. Attempt #3+ requires a new rule, fixed in advance.

Attention Diagnostic Suite

The Attention Diagnostic Suite (formerly “validation gates”) runs five statistical probes on PSI history each week to characterize signal structure, not as a deployment go/no-go and not as authorization for deal-forecast claims. Clean-corpus floor: dates ≥ 2026-04-27. It does not gate production, enrichment, or link prediction. Current results are under Latest diagnostic run.

The Five Probes

  • Permutation test (p < 0.01, 1,000 shuffles): Is PSI-outcome correlation distinguishable from a shuffled null?
  • Subsample stability: Variation in the mean across 3 non-overlapping time periods is small relative to the spread within them.
  • Walk-forward cross-validation: Positive information coefficient in >60% of out-of-sample windows.
  • Factor regression: Residual alpha after controlling for news volume and sector momentum. A pass needs PSI β > 0 and R² > 0.
  • Decay analysis: Characterized half-life (signal mean-reversion profile).

Fit Health Study

Paused from November 2026 until a corrected fit study replaces it.

Last run: 2026-10-04, on 175 days of history. Ensemble fit-gap 0.0591, status OVERFIT.

The study ran monthly until it was paused. It compared train and holdout ROC-AUC across graph-history lookback settings (7, 14, 30, 45 days, and production “full”) for seven link-prediction models (VGAE was not run) and their ensemble. Purpose: an overfit check, not model promotion.

The headline metric is fit-gap: train ROC-AUC minus holdout ROC-AUC. Fit-gap above 0.10 flags overfit; gap below 0.05 with reasonable absolute AUC is healthy. The study also surfaces the underfit signal: both curves low and flat across training sizes indicates insufficient model capacity.

Status Thresholds

OVERFIT: ensemble fit-gap > 0.10, or rising across the last three runs
WATCH: ensemble fit-gap between 0.05 and 0.10, or any single model above 0.10
HEALTHY: otherwise

Secondary Credibility Panels (top-50)

Two panels are reported alongside the learning curves but do not drive the fit-gap diagnosis. EDGAR co-filing hit rate measures whether the model surfaces real-world ties confirmed by SEC filings. PSI co-movement hit rate measures whether high-confidence predicted pairs move in lockstep on PSI regime change. Both serve as credibility checks on model output quality.

Request Full Study Results

Diagnostic reports, ablation tables, Phase-0 panel, and LF-1 artifacts available to qualified parties.

Request study results by email

Frequently Asked Questions

What is the Plocamium Intelligence Network?
An interactive visualization mapping entity relationships across healthcare, industrials, and geopolitics. It shows how companies, regulators, and decision-makers connect through shared news signals and published analysis.
How often is the network updated?
The network visualization is re-rendered every morning. Signal ingestion runs every 2 hours from curated RSS feeds plus news APIs. Entity extraction runs on every signal ingestion and again on published articles. Entity enrichment runs when articles are published.
What is BICS and why is it used?
BICS (Bloomberg Industry Classification System) is the taxonomy used by global asset managers for sector classification. It provides four levels of granularity: Sector → Industry Group → Industry → Sub-Industry. Using BICS ensures compatibility with how institutional investors already categorize companies.
How are entities connected in the graph?
Two entities are connected when they co-occur in two or more shared signals or articles. Each signal gets a unique content-hash key (e.g., HEALTHCARE-3a5e65588bb5) so co-occurrence reflects real editorial relationships. Additionally, relationships inferred by AI during BICS enrichment create edges even without co-occurrence. Edge weight reflects the number of shared items. This surfaces non-obvious relationships, such as entities appearing together in deal coverage and regulatory filings.
What does the daily insight tell me?
A short AI-generated briefing in four parts: the entities that stand out, one new connection chosen from the day’s real changes, what the PSI reading means, and what to research next. It is research, not investment advice.
Can I access detailed entity data?
Detailed signal timelines, entity relationship graphs, BICS profiles, and coverage analysis are available to Plocamium clients and partners. Click any entity node on the network to request access, or contact us directly.
How does variance analysis work?
Network state is archived daily. New edges are detected by comparing today’s graph against yesterday’s. Centrality drift tracks which entities are gaining or losing connections over time. The daily insight highlights these changes.
What are the ML models and how do they work?
Two models retrain weekly: an Entity Momentum Scorer computes a 0-100 score from 9 features (mention count, deal context, recency, relationship density, market cap, enrichment completeness, and more). A Sector Clustering model (KMeans) groups entities by behavioral similarity rather than just BICS label, revealing clusters like “deal-active heavyweights” or “resource geographies” that span traditional sector boundaries. Momentum drives node sizing; clusters drive node coloring.
What is the Rising Entities panel?
The Rising Entities panel surfaces the top 10 entities by ML momentum score. Momentum combines mention count, deal context, relationship density and enrichment, among other features.
How do ML models affect content production?
Momentum scores also influence which topics Plocamium covers and how its articles and social posts are scored and framed.
What is a cluster-crossing signal?
When an article covers entities from different ML clusters, data-driven groupings that often span traditional BICS sector boundaries, the system flags this as a cluster-crossing signal. For example, an article connecting a “deal-active heavyweights” cluster entity (UPMC) with a “resource geography” cluster entity (Permian Basin) would trigger this flag. They show coverage linking sectors that rarely appear together.
What is an entity watchlist?
Users can subscribe to watch specific entities and receive daily email alerts when tracked entities cross momentum thresholds, change ML clusters, gain new graph connections, or see significant momentum shifts. Watchlists use double opt-in confirmation via email and support five distinct alert trigger types. Alerts are delivered as a daily digest via email.
How does source diversity work?
RSS feeds are shuffled randomly each run and capped at 5 items per feed, so no single feed dominates the signal pool. Using both curated RSS feeds and news APIs broadens geographic and editorial coverage. This produces more representative entity co-occurrence data and reduces bias toward any particular news source.
What are centrality movers?
The Rising/Falling Centrality panel tracks which entities are gaining or losing graph connections compared to the previous day. Rising centrality means more co-occurrence edges formed, the entity appeared alongside more peers in recent signals. Falling centrality means connections dropped. Tooltips show the exact degree change, giving investors a quick read on which entities are becoming more or less central to the network.
What is the Plocamium Signal Index (PSI)?
A daily composite score measuring how much the entity graph is deviating from its own learned baseline. It combines five orthogonal signals, attention cascades (Hawkes process), narrative drift (vector embeddings), graph structure (spectral shift), sentiment divergence, and source concentration (HHI), into a single regime-aware indicator. PSI uses Bayesian Online Changepoint Detection to distinguish permanent shifts from temporary anomalies, and a Kalman filter to separate signal from noise.
How is PSI different from a sentiment score?
Sentiment measures tone (positive/negative). PSI measures structural disruption in the entity network, new relationships forming, communities reorganizing, narratives shifting. A market can be highly negative (bearish sentiment) but structurally quiet (low PSI), or neutral in tone but structurally chaotic (high PSI). They measure fundamentally different things.
What are Entity Intelligence Cards?
Data cards in the subscriber dashboard and PDF reports, showing momentum, network position, narrative drift, sentiment trajectory and sentiment-momentum divergence alongside the PSI score. Until the index moves to its new method, the last three come from that method's daily shadow run.
How does PSI handle regime changes?
BOCPD (Bayesian Online Changepoint Detection) maintains a posterior distribution over how long since the last structural break. When a changepoint is detected, PSI enters a 7-day transition window with widened confidence bands and a visible “regime shift detected” flag. After transition, baseline parameters are re-estimated from post-break data only. No arbitrary rolling windows, the model adapts to the data.
What validation does PSI undergo?
The Attention Diagnostic Suite runs five statistical probes weekly to characterize PSI signal structure, not as a deployment gate. External deal forecasting is tested via LF-1 (attempt #2, rule fixed 2026-06-08, failed its rule). The latest results are under What the Evidence Supports. Production link predictions are ranked by Node2Vec similarity.

Changelog

Changes to this methodology, newest first.

  • 2026-11-03 · Bare titles such as Foreign Minister and Chief Executive Officer are no longer scored or shown as entities. Daily briefings from November 4, 2026 leave out sentences that forecast events or say what investors or markets are doing.
  • 2026-11-03 · Dossiers count connections across the full network for every entity; those on the map used to read a count against the 150 entities on it. The live score is no longer described as a z-score: it is a smoothed composite pulled toward its sector’s average and scaled down for entities with little history.
  • 2026-11-03 · Attention spike alerts now compare an entity’s articles in the latest day with its own recent daily median, and fire at four times that median and at least 10 articles. They used to read a running total of mentions, and they no longer cite the 49% deal figure. The rule describes coverage and predicts nothing.
  • 2026-11-03 · New connections now count only pairs where both entities were tracked the day before, and movement in connections is measured across the full network rather than the 150-entity map. Daily briefings no longer call PSI a z-score, quote running mention totals as coverage or forecast announcements, and Plocamium’s own articles no longer count toward a Surprise. The November 3 briefing was corrected in place.
  • 2026-11-03 · Reviewed the LF-1 deal test. Its rule still failed, but its comparison rate of 49.4% read a running total of mentions rather than a coverage spike, so this page no longer treats the test as evidence against deal forecasting or for a mention spike. Deal forecasting remains unproven.
  • 2026-11-03 · Corrected descriptions that did not match the running code: topic trigger thresholds, feed sampling, enrichment timing and fields, watchlist triggers, and where entity cards appear. Lead-lag analysis now says it is not yet shown to clients. Entity cards now carry narrative drift, sentiment trajectory and sentiment-momentum divergence, read from the new method’s daily shadow run. A note marks the sections that describe the index as designed rather than as it runs today.
  • 2026-11-03 · This page now opens with what PSI measures, how to read a score and what the evidence supports, and reports each study from its latest stored run. Fit Health had been showing a June run as the latest; its last run was October 4. The claim that PSI is a contra-indicator for mention growth was removed, because the latest diagnostic run no longer supports it (permutation p=0.061). Language suggesting PSI anticipates events or markets was removed, along with the section on competitive defensibility.
  • 2026-11-02 · Titles, name fragments and adjectives (such as President, CEO, of State, American) are no longer scored as entities; their slots go to the next most-mentioned real entities, which show as Insufficient data until they have 10 daily observations.
  • 2026-11-02 · The client portal, its PDF reports and the daily insight image now use the published regime bands and colours, read each regime from its score, and show an entity without a score as No score yet or Insufficient data instead of 0.00 NORMAL.
  • 2026-11-01 · The public entity glossary now states the published regime bands, no longer lists a CRITICAL regime (there never was one), and leaves out entities with fewer than 10 daily observations instead of listing them at 0.00 NORMAL.
  • 2026-11-01 · The dispatch email shows momentum on its 0 to 100 scale with its band in words, and its bar now fills in proportion.
  • 2026-11-01 · The Pro API now serves the current day's scores; it had been serving a snapshot frozen on 2026-03-23.
  • 2026-11-01 · The weekly client digest now compares each score with the one a week earlier, so its week-over-week and regime changes are real, and a watched entity without a score is listed as No score yet.
  • 2026-11-01 · Dossier pages now lead with the PSI score on its five-band ladder with the current band marked, and the subscriber view's labels meet the 4.5:1 contrast standard.
  • 2026-11-01 · Entities with fewer than 10 daily observations are now shown as Insufficient data, with no PSI score or regime, instead of NORMAL with a score of 0, and the market PSI now averages scored entities only. Regime thresholds are unchanged.
  • 2026-11-01 · Dossier regime tooltips now state the published z-score bands, regime colours match the rest of the site, and dossier momentum shows its band in words (15 and up high, 5 and up mid).
  • 2026-11-01 · Dossier confidence now shows how much of a full 30-day history a score rests on; at full confidence it wrongly read 1%.
  • 2026-11-01 · Removed two cold-start claims the scoring never implemented: percentile calibration for new entities and a sector-median estimated regime.
  • 2026-10-31 · Link prediction: the daily ensemble weight search and ablation were retired because they always settled on Node2Vec alone; published predictions are unchanged.
  • 2026-10-31 · GNN attention refinement retired: it had been switched off in production since June 2026, and its code was removed; PSI scores are unchanged.
  • 2026-10-31 · The monthly link-prediction fit study is paused until a corrected study replaces it.
  • 2026-10-31 · Removed the statement that clinical-trial status changes are ingested as signals; they never were.
  • 2026-10-31 · This page is now published automatically with each daily intelligence render.

Explore the Network

See the live entity intelligence map with momentum scoring, cluster analysis, and daily AI-generated insights.

View Intelligence Network →
↑