WATCHTOWER · SYSTEMIC RISK / FORECAST LAB
Conditions, forecasts and outcomes
Systemic Risk combines typed current source context, explicit research targets, original evidence and a prospective outcome record. A source-model index is distinct from an event probability. A historical fit is distinct from validated live performance.
Current conditions
Only explicitly identified bounded source-model indices can populate an index. Alert counts, confidence values, numbers of monitored countries and unrelated scalar fields are not added into risk. Relevant, dated, unique source records are shown as context with their original URLs and clocks. Unknown or stale indices remain unavailable.
Source relevance uses disclosed topic/lifecycle/time rules, not a universal truth classifier. Excluded evidence and reasons can be inspected. Structural vulnerability, current shocks, source availability and forecast skill remain separate. Shared-source links explain dependence; they do not constitute independent confirmation or measured causal propagation.
Financial target definitions
The registered fund references are SPY, QQQ, EFA, EEM, TLT, HYG, GLD, USO, UUP, KBE and KRE. Horizons are 1, 5 and 20 NYSE sessions. Source-published USD fund prices exclude distributions, transaction costs and execution effects; fund proxies are not the underlying spot commodity, credit spread, yield or bank-failure outcome.
| Target | Definition |
|---|---|
| Terminal loss | Price return at the declared horizon is at or below the registered negative threshold from the frozen base close. |
| Within-window loss from base | At least one subsequent session close breaches that threshold relative to the starting close. This is not an intraday or moving-peak-to-trough drawdown definition. |
| Joint fund decline | At least two of SPY, HYG and TLT cross their declared terminal thresholds in the same horizon. The joint event is observed directly; marginal probabilities are not multiplied. It is not a definition of a worldwide systemic crisis. |
Thresholds are fixed in the versioned target registry before model comparisons. Equity and bank-fund thresholds are 2%, 3% and 5% by horizon; HYG and UUP 0.75%, 1.5% and 3%; TLT 1.5%, 3% and 5%; GLD 2%, 3% and 5%; USO 4%, 7% and 10%. Within-window targets use the 5- and 20-session horizons. There are 58 registered financial targets.
History, time and source finalization
Longer histories are read through the existing bounded Yahoo public-chart transport using explicit UTC periods and verified daily granularity, with up to 6,000 bars per reference. Available dates, gaps and original retrieval clocks are shown. Historical reconstruction uses the current published vintage; corporate actions, revisions and selection limitations remain relevant. It is not an original-release-vintage backtest.
Regular NYSE holidays and reviewed exceptional closures are respected. The conservative finalization boundary is 20:00 America/New_York. A retained bar is considered finalized only if the source was captured after the corresponding finalization boundary. New market forecasts issue only before 09:00 or after 20:00 New York, using the expected latest finalized base. Missing sessions do not shorten a horizon. A frozen base-price revision beyond one part per million withholds scoring for review. Source and outcome revisions remain auditable.
Baselines and challengers
- Historical base rate: non-overlapping historical target windows with conservative probability smoothing.
- Regime model: volatility/trend-matched empirical information, shrunk toward the base rate. Empirical reference ranges and smoothed event estimates are evaluated separately.
- Regularized logistic model: price-derived features, training-only scaling and minimum positive/negative example requirements.
- Ensemble: prior matured comparisons determine weights; weak components may receive zero weight and the baseline remains an anchor.
- Calibration: held-out predictions and sufficient observations are required. The label “calibrated model” describes its estimation step, not proof of live calibration.
Current authored cross-domain scenario probabilities are contextual. Because they already consume risk inputs, they are excluded as independent systemic predictors. Current weather and schedules are never inserted into old market training as if they were known historically.
Evaluation
Chronological origins use only preceding matured target windows. Normalization and model fitting use prior data. Ensemble weights are selected before the calibration segment; calibration is frozen before the later test segment. One-session historical tests are sampled every five sessions; five- and twenty-session tests use their respective non-overlapping steps. Earlier crisis-period replay diagnostics are labelled separately from the later test period.
Metrics include Brier and log loss, baseline-relative skill, reliability bins with uncertainty, calibration slope/intercept, precision–recall, warning lead time and false alarms under a declared alert budget. Continuous reference ranges are evaluated using coverage, weighted interval scores, tail quantile losses and a frozen 101-quantile CRPS approximation. A wide interval is not automatically a useful forecast.
Prospective promotion requires at least 60 non-overlapping windows, 10 positive and 10 negative outcomes, sufficient independent blocks, positive paired baseline improvement, acceptable reliability and better warning utility. These counts are minimum eligibility floors, not proof on their own. Time-block uncertainty and simultaneous-target comparisons are retained. Assets and horizons share information and are not treated as independent replications. Unsupported targets remain research or unavailable.
Regional reporting and physical impacts
The regional research target is a newly issued severe/extreme Actual/Public CAP message first observed by the scheduled sampler for a country within 24 UTC hours. The sampler normally runs about every 15 minutes plus processing time. Short-lived messages between samples may be missed. This target concerns sampled reporting, not all hazards or a probability of an outage, crop loss or human impact. JMA native municipality/hazard states remain separately labelled current context.
A positive observation retains the original dated message and its first-seen capture. A negative requires a healthy complete sampled observation window with no gap exceeding 30 minutes. Incomplete windows remain inconclusive; partial positives are excluded from calibration/training to avoid selectively learning only observed adverse events. At least 60 complete days and both classes are required for base-rate research; seasonal estimation additionally requires 730 days of span and matching seasonal examples.
Source weather forecasts retain their original coordinates, issuances, validity and units. Natural Earth country joins are approximate context. Numerical exposure sensitivity is available only for explicitly typed, dated measured dependency shares; qualitative links never acquire invented weights. Specialist bank-failure, crop, outbreak, cyber and geopolitical event probabilities need appropriate labels and coverage before qualification.
Immutable evidence and outcomes
Each forecast binds its target, model, policy and exact input evidence. Baselines and challengers coexist in a new versioned ledger. Original forecasts are immutable; outcomes and compatible source revisions append records. Issued evidence is retained. Unissued feature snapshots have a 90-day retention window; source and ledger capacity are bounded and reviewed rather than silently deleting forecasts.
Existing authored scenarios retain their separate resolution rules and evidence-backed adjudication workflow. Unresolved judgments are not invented as binary labels. Source permission pauses and rate backoff remain binding; a successful download with an old session triggers bounded catch-up rather than a false “current” state.
Sources and navigation
Yahoo Finance public reference histories · Severe Weather Information Centre and named issuing authorities · Natural Earth. Other source-model context retains its original publisher/adapter references. No publisher endorses Watchtower's derived models.
Read the exact machine-readable registry and policy. Open Systemic Risk → Current, Outlook, Impact & dependencies, Evidence, Track record or Methods. Forecast Lab also provides a Systemic risk view with the same snapshot and calculations.