Predictions

Candidate models inventory. Nothing trained yet — this is the wishlist + readiness check.

Predictive ML is intentionally last — we want at least 1k accounts and 5k agreements before we train anything, otherwise we'd be fitting noise. Tracked in docs/04-technical-build.md §19.

Churn risk

Logistic regression on PG counters; zero net-new infra.

inputs ready

Inputs

last_seen_at, agreements/30d trend, channel-config completeness

Target

P(churn within 30 days)

Deal-success probability

Useful at Step 1 of seal flow to flag risky proposals.

inputs ready

Inputs

creator CVI, counterparty CVI, country-pair history, amount distribution

Target

P(seal → reflect within 7 days)

Country LTV

FX vol needs OpenExchangeRates historical pull.

partial

Inputs

sealed-deal volume, average reflection velocity, FX volatility

Target

expected lifetime XBN per signup

Sybil cluster confidence

Needs the abuse-detection worker on Memgraph (Q3 2026).

pending

Inputs

shared IPs, device fingerprints, reflection-rating uniformity

Target

P(account is part of a wash cluster)

Real-time signals to feed any of these live in Metrics, cohorts in Cohort, and the abuse signal stream in Fraud.