Measuring how much of a pool is rented
Reward elasticity is the one quantity that separates deposits that stay from deposits that are being paid to sit still — and it is not symmetric.
“Mercenary liquidity” is a description, not a measurement. To put a number on it you need something observable that moves when the subsidy moves. Emission rates change discretely, publicly and on a schedule — which makes them close to an experiment run for us.
The definition
Let R_P(t)be the incentive-derived reward rate, annualised, in excess of the pool’s organic yield. The reward elasticity of liquidity is the percentage change in deposits per percentage change in that rate:
ε_P = ∂ log L_P / ∂ log R_PEmpirically, incentivised stablecoin pools frequently sit between 1.5 and 4. A pool at ε = 3.1 is not a pool with a risk factor; it is a pool where roughly three percent of deposits leave for every one percent the reward rate falls.
1.5 – 4
typical ε for incentivised stablecoin pools
ε⁻
the only direction that enters the hazard model
Δ log R
estimated off observable rate changes, not surveys
Estimation
For campaign c with rate changes at times τ₁…τk, we regress the change in log liquidity on the change in log reward rate:
Δ log L_P(τⱼ) = α + ε_P · Δ log R_P(τⱼ) + γᵀ Z(τⱼ) + uⱼWithout Z, a pool that shrank because the whole chain shrank would be scored as highly elastic. The controls are what make the coefficient about the pool rather than about the week.
Pools with too few rate-change events do not get a noisy estimate published as though it were a precise one. Elasticity is pooled hierarchically across pools sharing an asset class and venue type, with shrinkage toward the class mean — a partial-pooling estimator that degrades gracefully for a pool with two observations instead of pretending it has twenty.
The asymmetry that matters
Here is the part that looks like a modelling nicety and is actually the difference between a useful system and a permanently alarmist one.
Elasticity estimated from increases in the reward rate is not the same as elasticity on decreases. Capital enters faster than it leaves — partly attention effects, partly exit friction. So the model estimates two coefficients, ε⁺ and ε⁻, and only ε⁻ enters the hazard specification.
Elasticity is not realised in a vacuum
Knowing ε⁻ tells you how responsive deposits are. It does not tell you what happens at a specific cliff, because whether depositors act on that responsiveness depends on where else they can go and how expensive it is to leave. Scheduled shocks are therefore modelled with a dampening term:
E[ Δ log L_P | t_e ] = ε⁻_P · log(1 − m_e) · φ(A_P, F_P)If there is nowhere better to go, or leaving is expensive, elasticity is realised more slowly and less completely. This is also why a thirty-day withdrawal lock changes a pool’s projected horizon far more than the same dollars spent on a longer campaign: the lock raises F_P directly.
The other thirteen features
Elasticity is the core measurable, not the whole vector. Three of the rest are worth naming because they carry weight out of proportion to their simplicity.
| Feature | What it separates |
|---|---|
Tenure Gini G | Two pools with identical median tenure: one a long-standing core plus a churning fringe, one where every depositor arrived the same week. The first retains a floor. The second can empty completely. |
Post-campaign retention κ | The highest-signal feature available, because it is a direct observation of the quantity being predicted under a previous realisation. Unavailable for pools that have never seen a campaign end — which on a new chain is nearly all of them. |
Bridge-in recency ν | Capital that arrived by bridge in the last few weeks has already demonstrated mobility and a willingness to pay bridging costs to chase yield. It is the most likely to leave the same way. |
Every feature derives from public on-chain state or public campaign metadata. None require privileged access — a guarantee that cannot be independently recomputed is not a guarantee.