Signal Evaluation & Replay

What it does

Connects live events to decision-support candidates and validates them with replay and historical calibration.

Why it exists

To keep live signal interpretation and historical validation in the same operator loop without making replay the product itself.

Inputs

• events, themes, and transmission outputs
• market time series
• source and mapping priors
• historical replay frames

Outputs

• decision workflow objects
• signal candidate cards
• false-positive and confidence guardrails
• replay and walk-forward run summaries
• expanding-window walk-forward folds with 6-month validation and 6-month test segments
• backtest lab visuals and decision comparisons
• coverage-aware universe summary, review queue, and gap tracking
• approved dynamic candidates that join the next refresh and subsequent backtests
• universe policy modes for manual, guarded auto-approval, and full auto-approval
• scheduler-driven replay and nightly walk-forward when the automation worker is enabled
• theme discovery queue items that can become reusable backtest themes after Codex proposal and guarded promotion
• source registry candidates that can now auto-approve and auto-activate under guarded score and diversity policy
• coverage-gap candidate expansion that can now be requested by the scheduler instead of only by button clicks
• candidate auto-approval that now uses composite scoring plus sector and asset-kind caps
• repeated uncovered theme pressure that can now propose and guard-register missing historical datasets
• weak theme motifs and low-signal autonomous keywords that can now be auto-retired or auto-rejected
• weak idea cards that can now be auto-suppressed before the operator view
• cross-corroboration scoring that penalizes rumor-heavy or contradictory source clusters
• calibrated confidence and constrained autonomy actions: deploy, shadow, watch, abstain
• time-decay and recent-evidence floors so stale mapping priors cannot dominate current recommendations
• reality-aware execution checks for spread, slippage, liquidity, and session state
• shadow-book rollback signals that can force the engine back into shadow mode after weak recent performance
• macro kill-switch and hedge overlay that can override attractive micro themes
• hidden graph-propagated candidates that can surface second-order transmission plays
• explainable attribution that splits corroboration, graph, beta, macro, and reality penalties
• self-tuning experiment history and active weight profile summaries
• cost-adjusted replay summaries in Backtest Lab, not only raw signed returns
• default-on adaptive modules for ATR stops, Kalman noise tuning, execution-cost modeling, and theme sensitivity

Key UI surfaces

• Investment Workflow
• Auto Investment Ideas
• Replay Validation
• Transmission Sankey / Network
• Coverage-aware universe review queue

Algorithms involved

• event-to-market transmission
• regime weighting
• Kalman-style adaptive weighting
• Hawkes intensity, transfer entropy, bandits
• historical replay and warm-up handling
• coverage-aware candidate retrieval
• deterministic ranking over core plus approved expansion assets
• Codex-assisted candidate expansion as a reviewed queue, not an execution path
• guarded auto-approval with probation and auto-demotion
• composite auto-approval scoring with source, role, supporting signals, and crowding penalties
• dataset registry plus scheduler worker for unattended replay cadence
• theme discovery queue built from repeated unmapped motifs in replay frames
• Codex theme proposer automation with guarded auto-promotion, novelty overlap limits, and promotion scores
• source automation sweep for discovered feed and API registry acceptance using score, health, and diversity caps
• scheduler-driven candidate expansion and replay refresh after accepted universe changes, with cooldown and per-region balancing
• autonomous keyword lifecycle review and theme-queue hygiene
• pre-render idea-card triage and suppression
• cross-corroboration and contradiction penalties over clustered sources
• confidence calibration, no-trade gating, and shadow-only fallback
• execution-reality penalties for session state, spread, slippage, and liquidity
• recency weighting and stale-prior decay inside deterministic ranking
• shadow-book rollback and constrained autonomy state carried into the operator workflow
• guarded dataset discovery and auto-registration for replay-safe historical coverage expansion
• experiment registry and self-tuning weight promotion / rollback
• graph-driven hidden candidate propagation beyond direct trigger keywords
• macro risk overlay with kill-switch, hedge bias, and exposure caps
• explainable attribution over corroboration, graph support, beta, macro pressure, and penalties
• default-enabled adaptive ATR stop logic, Kalman auto-tuning, execution-cost curves, and theme sensitivity

Limits

The public site documents the system behavior but not private operational data or sensitive market configurations.

The engine remains:

• a decision-support and paper-trade research surface first
• a cost-aware replay engine second
• a human-reviewed or policy-gated execution candidate generator, not an unconstrained execution bot

What changed in practice

The replay stack no longer behaves like a static "theme scorecard" only.

It can now:

• widen its historical dataset registry when the current research surface is too narrow
• tune its own weight profile through guarded experiments
• discover hidden candidates through graph propagation
• apply top-down macro kill-switch logic before attractive micro trades survive
• explain whether an idea was driven by corroborated event evidence, graph support, generic beta, or penalty-heavy noise

This means the system is still constrained, but it is noticeably less dependent on manual queue curation than earlier versions.

What must exist before replay really starts

The repository now ships with a pilot registry that is already enabled, but replay still depends on live provider access.

Minimum unattended pilot inputs:

• coingecko-btc-core
• fred-core-cpi
• gdelt-middle-east
• acled-middle-east

Key requirements:

• coingecko and gdelt-doc: no key required
• fred: FRED_API_KEY
• acled: ACLED_ACCESS_TOKEN

Once those keys exist and the scheduler task is installed, the replay stack can start operating without repeated manual console runs.

Variant coverage

Primary: finance. Extended and shared support also exists in tech.