← Open Watchtower

Forecast Lab: models, evidence and resolution

Forecast Lab brings together statistical market projections, Watchtower’s existing cross-domain scenarios, source weather forecasts and upcoming economic catalysts. Each retains its own meaning, source clocks and evaluation status.

Research and shadow mode. Live calibration and a persistent predictive advantage have not been established. Historical reconstruction and authored scenario probabilities are not a proven live track record. Outcomes, uncertainty and baseline comparisons remain visible, including targets where the model does worse than its baseline.

Explicit market targets

The initial pilot covers nine US-listed ETF reference-price series at one, five and twenty NYSE sessions. The event is a strictly positive price return from the frozen base close to the target session close. A tie is not positive. Prices are in USD.

Cash distributions, execution prices, fees, slippage and trading costs are excluded. These are reference-price forecasts, not total returns or a simulated investment strategy.

Historical inputs and availability

The pilot uses up to five years of daily Yahoo Finance public-chart observations through Watchtower’s bounded provider transport. The nine series are aligned by exchange session. Missing, stale, duplicate, wrong-currency or incompatible records are withheld. Historical dates are source session labels: the timestamp of a daily bar does not mean its close, high or low was already known at the opening bell.

A conservative 20:00 America/New_York cutoff admits a session as finalized. Live issuance runs only before 09:00 or after 20:00 New York time, against the expected latest finalized session. This avoids intraday and early-close ambiguity. Research projections may be displayed before issuance; they are never backdated into the live ledger.

Session calculations include regular NYSE holidays and the January 9, 2025 closure. Exchange calendars govern the reference convention. Unforeseen closures or missing observations do not silently shorten the target horizon; an unavailable required session remains awaiting source evidence.

Three empirical model components

1. Historical distribution: non-overlapping historical price-return blocks, ending no later than the forecast origin, within a maximum 756-session training lookback.
2. Volatility-scaled history: historical returns are centered, scaled by current twenty-session daily volatility relative to training-origin volatility (ratio bounded to 0.5–2), and given half the historical mean drift. These are fixed modeling choices, not an estimated causal law.
3. Cross-asset regime analogs: historical origins close to the current standardized price-feature vector. Up to forty greedily selected non-overlapping analog windows are used; at least twelve are required.

Features include the target’s one-, five- and twenty-session log returns, twenty- and sixty-session daily volatility, and twenty-session equity, Treasury, credit-minus-Treasury, dollar, gold, oil and emerging-minus-developed equity trends. Feature scaling is fitted only to the training window. Driver values describe the observed price regime; they are not causal attributions or independently verified reasons for a future move.

Mixture weights use only previously matured chronological evaluations: equal weights until twelve evaluations, then a bounded weighting of the last sixty component Brier scores. For component k, weight is 0.1 + 0.7 × exp(−8 × loss_k) / Σ exp(−8 × loss). The three components are correlated. Their disagreement is displayed in percentage points.

The median, central 80% interval and directional estimate come from the weighted empirical return distribution. Extreme probabilities are bounded to 0.001–0.999. An 80% model interval is nominal; measured historical and prospective coverage determine how well it behaves. No neural-network, causal-discovery or universal-predictability claim is made.

Chronological reconstruction

Each historical origin uses only features and fully matured labels at or before that origin. Evaluation windows do not overlap within each symbol/horizon partition. Model weights use preceding evaluations, not their future outcomes. Reported metrics include Brier score, log loss, median-return error, nominal-interval coverage and probability bins. Comparisons include the historical-distribution baseline, a 50/50 direction forecast and an unchanged-price forecast.

This uses the current published historical vintage. Providers may revise observations or adjust historical prices for corporate actions. It is not an original-vintage point-in-time backtest. The selected instrument universe also introduces selection and survivorship limitations. Positive sample-relative improvement is not proof of statistical significance, robustness or profitability. Results across assets and horizons are dependent and are not pooled into a fictitious independent sample.

Prospective ledger and outcome revisions

Every live issuance freezes its actual issue time, target and base session, model and definition versions, source cutoff, probability, range, baseline and exact normalized input bundle. Stable content identifiers let the record be inspected and exported. Repeated worker runs do not duplicate the same target/base-session issuance. The archive begins with real collection and issuance, not reconstructed earlier records.

After the target session is finalized and a compatible source close is available, the deterministic resolver records the actual price return and binary direction. Missing target sessions remain awaiting source evidence. A source change to the frozen base price beyond one part per million—such as a split adjustment or correction—places the result in source-revised state and withholds scoring. Target-price corrections append a new outcome revision. Original issuances and previous outcome revisions remain retained; current scores use the latest compatible outcome.

The primary live performance table uses a non-overlapping subset within each symbol/horizon. All-issuance metrics are also available in the read-only data view, with their dependence retained. Open, awaiting-source and source-revised records are excluded from proper scores. The existing adjudicated cross-domain Forecast Calibration ledger is separate and unchanged.

Cross-domain evidence

Forecast Lab reuses Outcome Studio’s eight source-linked scenario families: geopolitical escalation, strategic supply corridors, banking/liquidity, cross-asset markets, grid/digital infrastructure, mobility/official controls, health/biosecurity and food/energy prices. Authored priors, dependency-capped model contributions, confirm/falsify conditions, model clocks and source references remain explicit. Their sensitivity ranges are not confidence intervals and the scenarios are not mutually exclusive. Expired estimates are withheld.

MET Norway coordinate weather forecasts retain source issue and validity clocks, units and CC BY 4.0 attribution. The bounded point sample is not a continuous global forecast surface. Rainfall is not flood probability, wind is not a tornado prediction, and a reference-coordinate forecast is not a country-wide outcome. NOAA official outlooks preserve their geographical/event definitions and expiry where available.

The economic catalyst view uses existing official/provider Calendar records over a bounded upcoming window. Scheduled release dates, market consensus, reference periods and past values stay separate; a passed date does not prove an actual release. Date-only clocks are not invented midnight timestamps. Maritime hazard estimates remain subject to their existing history gate. Missing model probabilities are not zero.

Current cross-domain evidence is contextual; it is not inserted into historical market training as if it had been known years earlier. A longer original-vintage feature history is required before such cross-domain predictive relationships can be evaluated honestly.

Conditional scenarios and watches

The What-if Lab selects non-overlapping historical windows where one ETF’s return falls within a user-specified band, then reports the other ETF’s same-window return distribution. At least twelve matching windows are required. This is historical co-movement conditional on the specified fund-price return, not a causal shock-response estimate or calibrated future conditional probability.

Watchlists and saved baselines are scoped to the verified account in this browser. Material estimate changes of at least five percentage points are shown while the workspace is visible. This does not configure closed-browser monitoring or external messages. Owner changes, access loss and replay clear the active workspace; local corrupt preferences are preserved for recovery.

Collection, source rights and exports

Existing public providers are used without a new paid subscription. History collection is bounded, ordinarily six-hourly with an additional post-close refresh; context refreshes approximately every five minutes. Permission/rate pauses and last-good clocks are retained. Read requests do not force upstream collection. Source availability and provider terms continue to apply.

JSON exports preserve evidence, model versions and distinctions. Target CSV exports contain the selected public market view. The reviewed read-only dataset supports overview, detail, sources, performance, ledger, immutable forecast/evidence and conditional-scenario views. Personal watches and account records are excluded.

Open Forecast Lab in Watchtower · Outcome Studio methods · Calendar methods