A hazard curve fitted per pool.
A hazard curve fitted per pool, with campaign expiries applied as discrete jumps rather than smoothed into a background rate.
The proportional
hazards trap.
The dominant failure of a naive survival model here is treating a scheduled cliff as if it were background hazard.
A campaign expiry is a known date, not a random arrival.
Smoothing it understates hazard after expiry and overstates it before.
So hazard ratios are held constant only within an incentive regime.
How this works
Three estimators run in parallel and are combined by inverse-variance weighting. Where they disagree, the published interval widens rather than picking a winner.
01
Cox proportional hazards
Interpretable, well-calibrated on sufficient data, and degrades gracefully when features are missing.
02
Gradient-boosted survival trees
Captures interactions the additive model misses, particularly concentration against a near campaign boundary.
03
Structural simulation
Cohorts drawn from the fitted tenure distribution, stepped forward through scheduled rate changes.
What it reads, derives,
and does with it.
What it reads
The state each estimator is fitted on.
What it derives
What the ensemble produces.
What it does with it
The action each signal triggers.
“The scoring model is publishable and copyable, but the accuracy record is not. A competitor who reimplements the ensemble from this paper starts with an identical model and a calibration history of zero.”
Thirty-six surfaces across contract, API, agent, and automation
Liquidity Intelligence Infrastructure. Read the horizon before the capital moves, not after.