Signal inputs
On-chain protocol health
Utilisation curves, supply and borrow caps, oracle staleness,
reserve factors, kink slopes, interest-accrual fragility.
TVL flow + cohort behaviour
Institutional vs. retail cohort flow, multi-window TVL deltas,
address-level rebalance event clustering.
Prediction-market + sentiment
Prediction-market-derived risk premia, sentiment on
protocol-specific tickers, governance-vote dispersion.
Macro indicators
Stablecoin supply trajectories, central-bank rate prints, funding
rates, category-level capital flow.
Per-protocol specialists
Mechanism-specific signals: AMM-yield discount evolution,
vault-level curator action, parameter-drift detection, governance
activity.
Cross-protocol correlation
Pair-wise APY co-movement matrix, regime-flip detection, leader
and laggard graph from a rolling correlation window.
Capability surface
Signal Mesh emits predictions on four horizons. Each horizon is a gradient-boosted classifier trained on the data shape that horizon answers best.Sub-day direction
Sub-day direction
Short-horizon directional signal for tight-cadence rebalance
triggers. The right horizon when utilisation can shift inside a
single trading day and rebalance cost is low.
1-day direction
1-day direction
The primary directional model. Calibrated for confidence-score
interpretability and used as the headline directional signal on
product surfaces. Output works as a binary trigger or a
confidence-weighted scaler.
7-day APY estimate
7-day APY estimate
Continuous APY forecast for the 7-day forward window, decomposed
into base yield, reward yield, and demand components per protocol
family. The decomposition lets the estimator handle protocols
with structurally different rate mechanics (lending,
AMM-style fixed yield, governance-set rates) inside one stack.
30-day stability
30-day stability
Three-class signal: trending up, trending down, stable. The
long-horizon allocation context. Surfaces pools where the
short-horizon signal is noise around a stable mean rather than a
real directional move.
- Per-mechanism specialists. AMM-style yield runs through its own specialist because the underlying pricing is mechanically different from utilisation-curve lending. Specialists outperform the generalist on their own protocol family.
- Regime detection. A separate model surfaces correlation regime flips, so the directional heads can be re-weighted when cross-protocol relationships break.
- Hybrid composite score. Combines the directional output, the continuous APY estimator, native risk signals, and liquidity-impact context into the single score that drives the allocation ranking.
Confidence calibration
Every directional output ships with a calibrated confidence score on a 0 to 100 scale. The calibration step is run as a separate stage after model training, so confidence values map to empirical hit-rates and can be used as a gating threshold by downstream consumers. A live calibration verifier monitors drift between the score and observed forward outcomes; calibration is the class of silent bug a model pipeline is most exposed to, and the verifier keeps that loop closed.Explainability
Every prediction carries a per-signal attribution vector showing which inputs contributed most to the directional and confidence output. Allocators see the top contributing signals alongside the recommendation, so the model output is auditable at the level of a single decision rather than as a black-box score.How predictions surface
Every prediction is published with:- The horizon (sub-day, 1 day, 7 days, 30 days)
- A direction class (up, down, stable)
- A calibrated confidence score (0 to 100)
- A continuous APY estimate
- A per-signal attribution vector
- A liquidity-impact context that answers “what would my deposit do to this pool’s APY”