Live Results
Pulled directly from the MCP gateway’s /showcase/ml-prediction/:symbol endpoint. Data refreshes client-side every 60 seconds. No API key required — showcase routes are public but rate-limited.
Reading the widget
Section titled “Reading the widget”p10(red) — 10th-percentile forecast. Lower bound of the envelope.p50(teal) — median forecast. The model’s central estimate.p90(green) — 90th-percentile forecast. Upper bound.▲ / ▼— whetherp50is above or below the last close.last— the most recent close used as reference.
When p90 - p10 is wide relative to recent history, the model is communicating low conviction. When it’s tight, high conviction. On FOMC/CPI/NFP days post–Phase 3, expect spread ≥ 1.5× the non-event baseline.
Prediction ages
Section titled “Prediction ages”Predictions are generated by the Railway cron 55 * * * * (1 h horizon) and 45 5 * * * (1 d horizon). The gateway serves the most recent row from ml_predictions — so between cron ticks, you’re seeing a cached answer, not a fresh one.
A prediction older than 75 minutes (missed cron tick) implies the RTX 4060 service was unreachable. Check cloudflared tunnel and the kronos-batch-1h cron status on Railway.
Historical accuracy
Section titled “Historical accuracy”Historical signal-scoring rollups are available via the gateway at /showcase/signal-performance — those feed the evaluation views in the prediction site. This docs site shows live predictions; the app at prediction.datfxlabs.com is the place to see rolling accuracy and per-instrument deep dives.
Gateway endpoints used
Section titled “Gateway endpoints used”| Endpoint | Auth | Cache TTL |
|---|---|---|
GET /showcase/ml-prediction/:symbol | none | 30 min (gateway LRU) |
GET /api/catalog | none | — |
GET /api/catalog/sample/ml-prediction | none | 30 min |
Full catalogue of showcase endpoints: see GET /api/catalog (27 entries as of 2026-04-23).