Scenario Methodology

Each scenario starts with a calibrated base rate probabilityderived from a specific historical reference class (e.g., “quarters with CPI >4% AND unemployment >5% since 1970”). This is not a guess — it’s the empirical frequency of similar conditions in the historical record, adjusted for current structural differences.

Probabilities are updated using Bayesian log-odds updating. New evidence shifts the log-odds of the probability proportional to a calibration constant. This ensures updates are:

• Bounded between 5% and 95% (no false certainty)
• Symmetric in log-space (rising from 10% to 20% requires the same evidence strength as rising from 80% to 90%)
• Time-decaying toward the base rate when evidence is absent (half-life parameterised)

Time decay: Between updates, the probability decays toward the calibrated base rate. The decay follows an exponential half-life — after one half-life period with no new evidence, the probability moves halfway back toward the base rate. This prevents stale probabilities from persisting.

Coherence enforcement: After individual scenario probabilities are computed, a simultaneous iterative pass enforces structural constraints between related scenarios. For example, mutually exclusive scenarios (Stagflation + Fed Pivot) cannot exceed their joint probability ceiling. This is order-independent — the same result regardless of which scenario is processed first.

Each scenario monitors 5 key indicators. Evidence strength is computed via z-scores on first-differences (changes, not levels) — this captures how unusual the recent movement is relative to the rolling 252-day volatility of changes. First-differences prevent spurious correlation from trending series.

Z-scores are weighted by per-metric importance weights and adjusted for cross-metric correlation using an effective degrees-of-freedom correction. If all 5 metrics are highly correlated, they collectively carry less information than 5 independent signals.

Regime-conditional weights: Some metrics have defined threshold levels where their economic interpretation changes (e.g., unemployment below 4% vs. above 6%). When the metric is in a specific regime, its weight is adjusted accordingly.

The evidence model requires at least 60 data points per metric for z-score computation. Metrics with fewer observations fall back to a threshold-proximity proxy.

The platform’s full macro data estate — 150+ indicators across rates, credit, commodities, currencies, and equities — feeds a regime classification that modulates scenario sensitivity.

Each scenario has a polarity(escalatory, benign, or neutral). During risk-off macro regimes, escalatory scenarios receive amplified evidence sensitivity while benign scenarios are dampened. This prevents the model from assigning high probability to “calm” scenarios during obvious stress periods.

Update frequency: Scenario articles are generated every 6 hours via automated pipeline, with cadence adjusted by heat classification. CRITICAL scenarios may update daily; COLD scenarios may go weeks between updates. Manual editorial triggers are available when events warrant immediate analysis.

Scenarios have a full lifecycle managed by four independent automated pipelines:

• Scenario Radar (every 6 hours): Scans 6 independent data source categories (FRED surprises, price moves, news clusters, CFTC positioning, composite indices, cross-source divergences) for emerging macro configurations. Regime-aware thresholds adjust for VIX levels.
• Article Generation (every 6 hours, offset +1h from radar): Runs the full 22-gate pipeline for active-tier scenarios. Internal heat-based cadence gates prevent over-publication.
• Lifecycle Evaluation (daily): Checks demotion criteria (probability below 5% for 3 consecutive cycles, or COLD heat for 30+ days), evaluates structured resolution conditions against live data, and cleans up stale watchlist entries and emergence signals.
• Coherence Audit (every 2 hours): Verifies probability coherence across mutually exclusive pairs, sum constraints, and conditional relationship consistency. Flags violations without auto-correcting.

Tiered tracking: Scenarios exist in either active tier (full computation, article generation) or watchlist tier (reduced monitoring every 48 hours, no article generation). The system maintains a maximum of 8 active + 5 watchlist scenarios to prevent analysis dilution.

Resolution conditions are structured and machine-checkable: each condition specifies a metric, comparison operator, threshold, and optional sustained-days requirement. The system checks these deterministically against live data — no LLM in the loop for resolution checking. When all conditions are met, the scenario is flagged for human review rather than auto-retired.

Full technical documentation of the radar, cascade templates, emergence gates, and lifecycle transitions is available on the Scenario Engine methodology page.