How the replacement-date model is computed.

Signal taxonomy

The PLX-Forecast model fuses 47 weighted signals across five families: capability, commercial, workforce, capital, and discourse. Each signal arrives at its own cadence and is normalized into a unitless posterior contribution.

• Capability: 23 frontier-model benchmarks (MMLU, GPQA, SWE-bench Verified, ARC-AGI, GAIA, plus 18 internal proxies).
• Commercial: 38 quarterly earnings transcripts across hyperscaler, vertical SaaS, and semis cohorts.
• Workforce: postings across 14 job boards, normalized into per-100 indexes (developer down, AI engineer up).
• Capital: private and public AI-related capital flows reconciled against a quarterly anchor.
• Discourse: 12,847 sentiment- weighted expert posts a day across X, Bluesky, and LinkedIn.

Update cadence

The full posterior is recomputed every minute. Capability and discourse signals stream in continuously; capital and commercial signals arrive on quarterly anchors and are smoothed by a Kalman-style filter.

Calibration

Calibration error has tightened from ±18 months to ±9 days over the last four quarters, primarily due to compounding evidence accumulation. Interval coverage is reported at the 94.2% level and re-evaluated against holdout signals each quarter.