The survival model
How a feature vector becomes a horizon: hazard specification, scheduled shocks, and the three estimators that vote.
Argued in full in the whitepaper at §4.
Hazard and horizon#
Retention failure is modelled as a survival process — Cox proportional hazards with time-varying covariates:
λ_P(t | X) = λ₀(t) · exp( βᵀ X_P(t) )
S_P(t) = exp( − ∫ λ_P(s | X) ds )The horizon is then the longest lookahead over which survival stays at least σ-likely, with σ fixed at 0.5 as the reference:
H_P(σ) = sup { h ≥ 0 : S_P(t₀ + h) ≥ σ }The baseline hazard is estimated non-parametrically and stratified by venue class — AMM pool, lending market, vault, staking contract. A lending market with utilisation-dependent withdrawal availability behaves nothing like a constant-product AMM.
Scheduled shocks are not hazard#
So hazard ratios are assumed constant within an incentive regime and not across regime boundaries. Campaign expiry, subsidy termination and step changes in emission rate all re-stratify the baseline. Known future events enter as deterministic jumps rather than estimated hazard; the horizon computation integrates the smooth hazard between events and applies the discrete drop at each one, taking whichever binds first — cumulative expected retention falling below θ, or survival falling below σ.
Three estimators#
| Estimator | What it contributes |
|---|---|
| Cox PH | Interpretable, well-calibrated on sufficient data, degrades gracefully when features are missing |
| Gradient-boosted survival trees | Interactions the additive model misses — particularly concentration × expiry, where high HHI shortens horizons far more sharply near a campaign boundary |
| Structural cohort simulation | Each depositor cohort gets a reservation yield drawn from the fitted tenure distribution; the simulation steps through scheduled rate changes and records first breach time across paths |
Outputs are combined by inverse-variance weighting on a rolling validation window. Divergence between them is itself a signal — when the structural simulation and the statistical models disagree materially, the conformal interval widens rather than the disagreement being averaged away.
Where language models are used#
Two bounded roles, neither of them estimation:
Extraction. Campaign terms are published inconsistently — forum posts, governance proposals, dashboard metadata, tweets. A small model normalises them into strictly typed JSON validated against a schema. Anything failing validation is dropped to a human review queue rather than guessed.
Anomaly triage. Flagging conditions the feature pipeline does not anticipate — an unannounced parameter change, a governance vote in progress, a pool migration — and routing them for handling.