Methodology

How the predictions are made — live, not marketing.

Pipeline

live · rolling 30 days

Every symbol is scored every cycle. Most attempts never become a published prediction — the gate abstains when conviction is thin, and that attrition is the point, not a shortfall. Validated counts what has since settled; the narrower cohort the published accuracy is computed on, and the exclusions that define it, are stated on the track record.

Universe84ksymbols5 markets
Features392per symbol14 signal groups
Models2champion + challenger7 horizons
Signal gate60%abstainedof 815k attempts
Published328kpredictions40% of attempts
Validated57kscored vs outcome17% matured

Abstained 486,907Published 327,886from 814,793 attempts

Model pipeline

Primary model

DAILY_LIGHT LIGHT_v20260622_1414

Version

LIGHT_v20260622_1414

Source

champion_registry

Challenger

DAILY_COMPLETE v20260917_1749

Last training

Sep 17, 2026

Cadence

weekly on GCP Saturday

Last calibration

—

Calibration error

—

Not carried by the pipeline-state payload. Expected calibration error is published, with its bins, on the Track record page.

Universe

Total symbols

84250

Markets

US, IN, EU, APAC, LATAM

LATAM 770EU 12133US 44458IN 7347APAC 19542

Features

Total features

392

Signal groups

Horizons

1d, 5d, 7d, 10d, 15d, 30d, 90d

trendmomentumvolumevolatilitysectormacrosentimentpatternfundamentalsnews_eventsfibonaccichart_patternsindia_signalsxbrl_fundamentals

Track record · 30d

-11.31%

Net edge

17.5%

Validation rate

57,235

Scored

Net edge is directional accuracy measured against coin_flip_50pct (directional_accuracy_vs_50pct_random), gross_of_costs. A negative figure means the published calls have been landing below a coin flip over this window — shown as measured, not smoothed. This payload does not disclose the confidence floor its cohort was cut at; the floor in force is published on the track record.

Snapshot generated Sep 17, 8:03 PM · data as of Sep 17, 7:07 PM.