Loading
Calendar and celestial overlays are narrative/actor-belief context only, not independent predictive signals.
A multi-layer intelligence system that monitors four primary signal layers plus a narrative overlay, scores their convergence, and synthesises high-conviction intelligence briefs. Every assessment traces back to observable data. Every prediction is tracked and scored.
Each primary signal layer operates on fundamentally different data sources with different dynamics. This independence is what makes convergence meaningful. Calendar and celestial data serve as narrative context for understanding actor beliefs, not as independent predictive signals.
NEXUS does not predict events. It identifies conditions under which events become more probable. The distinction matters: prediction implies certainty, while convergence analysis surfaces elevated probability windows.
Every output feeds back into the system. Weights, thresholds, and scoring parameters are living values that evolve with each completed prediction cycle. The system never assumes its current calibration is final.
Every assessment traces back to specific, observable data points. The system does not hallucinate connections or project patterns that aren't supported by the underlying signals.
Each phase builds on the previous. Click to expand.
More primary layers converging means a stronger signal. Two-layer convergences are common. Three or more are significant. Full four-layer convergence is exceptionally rare and always flagged as critical. Narrative overlays (CAL/CEL) add context but do not contribute convergence weight.
Signals that cluster tightly in time score higher than the same signals spread over weeks. The scoring function applies a time-decay weighting that favours narrow convergence windows.
Every convergence event is matched against a library of past convergences. Patterns with strong historical precedent and clear outcome data receive higher confidence scores.
Convergence scoring is non-linear and only counts primary signal layers (GEO, MKT, OSI, and additional data layers). Narrative overlays (CAL/CEL) provide actor-belief context but do not contribute convergence weight. The amplification curve steepens as more primary layers align, reflecting the decreasing probability of coincidental overlap.
The complete system reduces to three equations. Signal detection and Bayesian fusion produce a calibrated posterior. The posterior informs forecast generation. Forecasts are scored against outcomes with a proper scoring rule. Everything below is implemented in production code and runs against live data.
All terms are additive in log-odds space. The logistic function σ guarantees the output is a valid probability in (0, 1) regardless of input magnitude. No dimensional mixing, no unbounded outputs.
Each signal layer contributes a likelihood ratio via the exponential transform. The sensitivity constant k = 0.45 is calibrated so that significance s ∈ [0, 9] spans the full intensity scale against operating priors of 5-10%.
The discount factor dj prevents double-counting correlated evidence. It uses an evidence-weighted harmonic mean of pairwise independence factors, so layers that barely moved the posterior have negligible influence on the discount.
Di,j ∈ [0, 1] is pairwise independence (1 = fully independent). Reduces to direct pairwise independence when one prior layer dominates the evidence.
Applied as ln μ in log-odds. Only primary layers (GEO, MKT, OSI, SYS) count. Narrative overlays contribute zero convergence weight.
Calendar and celestial signals are actor-belief context. They tell you what market participants think matters, not what objectively matters. Capped at βmax = 0.40 log-odds.
Maximum probability shift: ~10% at the midpoint, ~6% at extremes. Deliberately small.
This is the step most systems hide. The posterior probability from Equation 1 is the quantitative anchor, but the final forecast confidence incorporates regime context, game-theoretic structure, and historical precedent. This function is not closed-form because it includes structured analytic tradecraft that resists reduction to algebra. We state this explicitly rather than hiding it behind an arrow.
Output of Equation 1. The quantitative anchor that constrains how far the forecast can deviate.
Tuple of volatility regime (calm/elevated/crisis), risk appetite (risk-on/neutral/risk-off/panic), and commodity regime (normal/tight/supply-shock). Wartime classification invalidates active predictions.
Harsanyi incomplete-information game: actors, type spaces, strategy sets, belief distributions, audience costs (Fearon 1995), and bargaining range. Conflict structurally likely when bargaining range ≤ 0.1.
Vector store of documents with 1024-dim embeddings (Voyage AI). Retrieved via cosine similarity ≥ 0.7. Contains historical precedents, resolved predictions, analyst notes, and ingested intelligence.
Decay-weighted Brier score with a 60-day half-life. Penalises overconfidence, rewards calibration. Recent predictions count more than older ones. This is a proper scoring rule: the only way to optimise it is to state your true beliefs.
Brier decomposes into Bias (systematic calibration error), Noise (confidence scatter), and Information (discrimination power). Positive Information means the signal framework has predictive value.
All constants are implemented in production code and validated against live outcomes. The sensitivity parameter k and layer reliabilities are living values subject to recalibration as the prediction record grows. Signal decay models disconfirmation as reversion to prior rather than negative evidence, a deliberate design choice that bounds the likelihood ratio to [1, ∞) for the current signal space.
Specific data providers, API configurations, and ingestion pipelines are proprietary. The categories above describe the types of data consumed, not the specific sources or methods used to acquire them.
Truly unprecedented events have no historical pattern to match against. The system can detect unusual conditions but cannot anticipate events with no precedent.
Some signal layers operate on delayed data. Geopolitical and OSINT signals may lag minutes to hours behind real-time events. Market data varies by feed tier.
Market regimes change. Correlations that held during one period may break down in the next. The feedback loop mitigates this but cannot eliminate it entirely.
Deep dive into signal detection, intensity scoring, decay functions, and cross-layer amplification.
Read moreNarrative context layers: historical calendar-market patterns as actor-belief overlay, not independent signals.
Read moreLive accuracy tracking, Brier scores, and performance breakdowns by signal layer and time horizon.
Read moreNash equilibria, Schelling focal points, and escalation ladders applied to geopolitical scenario modelling.
Read moreAccess the full NEXUS platform to explore live signal detection, convergence analysis, and AI-driven intelligence briefs.
Request Access