Crypto Radar Pro · Multi-Market

The scheduled monitor is a separate Cloudflare Worker. The Pages app can read D1 history even when the monitor Worker is not deployed.

SERVER-LITE is intentionally separate from the full browser Opportunity Engine. It uses a smaller technical model suitable for scheduled cloud monitoring.

Volume-at-price and TPO are approximations reconstructed from OHLCV candle ranges. They are not exchange tick-by-tick Market Profile data.

Crypto windows use UTC; US Stocks use America/New_York.

Sweep confirmation remains advisory in v38 and is saved for later validation.

These structure features remain advisory until enough resolved cost-aware outcomes exist.

Slippage is a research proxy, not a guaranteed executable spread.

Context is advisory and is not injected into base signal weights in v39.

A high-impact event within 30 minutes creates a research blackout gate in Master Verdict. Event times come from the provider and are treated as UTC.

Headline classification uses transparent keyword heuristics. It is event-risk context, not a trained sentiment model.

The model is chronological train/test, runs locally, and remains shadow-only. It never changes the base engine automatically.

The empirical and ML values are research estimates from prior resolved samples. They are not guaranteed probabilities of profit.

A regime model is allowed into the ensemble only after its own chronological OOS gates pass. Base engine weights remain frozen.

PSI and performance shifts are diagnostics, not proof that the market regime has permanently changed.

This waterfall explains the research lens from explicit components. It is not SHAP and does not claim causal attribution.

Expanding-window folds train only on observations strictly before the purge gap. Test windows are separated by an embargo.

Experiments are versioned locally. Re-testing the same parameters on the same history is flagged instead of being treated as fresh evidence.

Meta-labeling does not create direction. It decides whether the base LONG/SHORT setup should be TRADE or SKIP based on resolved historical setups.

Three independent model forms are trained on the same chronological train window. The ensemble receives a usable gate only when all three component models and the fixed-weight ensemble pass OOS checks. Test data is not used to optimize ensemble weights.

IndexedDB is an archive layer, not a destructive migration: the current small working sets stay in localStorage for compatibility while larger historical records are retained in IndexedDB.

Paper-only execution simulator. Entry ladders, partial fills, OCO-style exits, configurable scale-outs, break-even, ATR trailing, volatility stop, time stop, fees, slippage and gap-through logic are simulated from OHLCV. Same-candle sequencing remains model-dependent and is resolved conservatively by default.

Tests nearby thresholds, ATR stops, target multiples and higher modeled costs on an embargoed historical OOS segment. Positive variants are evidence of parameter stability, not proof of future profitability.

Cone percentiles are descriptive and use overlapping rolling windows; v57 also reports a non-overlapping sensitivity percentile.

Cross-timeframe comparison uses each timeframe's native ATR%, RV percentile and TTM squeeze state; absolute non-annualized window values are not compared directly.

Historical VaR/CVaR uses recent daily returns and current paper exposures. It is a research risk estimate, not a guarantee of maximum loss.

Marginal and Component VaR use a covariance/normal approximation. Historical VaR/CVaR above remain distribution-free empirical estimates.

Risk budgets are paper-only hard limits based on initial-risk USD and paper equity. They do not place, modify or block real exchange orders.

Run an active LONG/SHORT setup, then assess its marginal portfolio effect before adding it to Paper Trading.

Dynamic correlation compares recent and longer windows. Common-factor beta is an internal portfolio factor diagnostic, not CAPM beta to a published index.

Sizing is a paper/research diagnostic. It does not resize existing orders or positions.

Stress scenarios are deterministic what-if tests. They are not forecasts and can understate losses if market liquidity or correlations move beyond the configured shocks.

Families use feature-profile similarity between strategy segments. Similar segments should not be treated as independent evidence merely because their labels differ.

Block bootstrap preserves short runs of wins/losses better than IID resampling. It still cannot prove future edge.

The overlay is advisory and does not rewrite the base signal engine. Small samples are shrunk strongly toward neutral.

Historical rank is computed from prefix-only technical features. It deliberately excludes today's journal reliability, today's microstructure and today's external context. Outcome is evaluated only after ranking. Nasdaq membership is reconstructed for anchors within the v65 inherited coverage window (from 22 Dec 2025). Corporate-action reconciliations are tagged and residual membership/provider bias is never hidden; crypto still uses current Pionex membership.

Ranking combines scanner strength, confidence, ADX, structure, volume, relative liquidity, regime quality and shrunk journal reliability. It is a research ranking, not a profit guarantee.

A challenger can become the research champion only after it leads on the locked holdout and on enough post-start forward points. Promotion does not alter the base directional engine.

Threshold selection occurs only inside each training window; the following fold remains untouched OOS.

Block-bootstrap simulation of R outcomes. This is a stress model, not a forecast.

Eligibility combines strict Profit Readiness, Shadow evidence, execution-gap evidence, Provider/Data Health, drift/governor state and a valid cost model. It is not a profit prediction or trade instruction.

The review is deterministic attribution from stored evidence, not an external generative-model opinion.