A
AvenaTerminal · Est. 2026
Architectural Commitment 03 · Methodology Evolution

Every weight, every revision, in public.

Static methodologies are a credibility ceiling. Most index providers chose their weights once and never revisit. We publish every version of every methodology Avena has ever shipped — weights, rationale, derivation method, activation date — with the prior version linked. Competitors look frozen. We visibly learn.

Note: weight changes in v1 are gated on human review. Autonomous weekly recalibration is wired but not enabled until we accumulate ≥10k labelled out-of-sample resolutions — chasing noise on a small sample would produce embarrassing oscillations.

Avena Property Cycle Index

1 version
v1.0.0activemanual
2026-05-25

Avena Property Cycle Index — composite of five sub-signals from the BIS residential property cycle literature and the ESRB anti-cyclical buffer framework.

View weights JSON
{
  "policy_risk": 0.1,
  "price_velocity": 0.3,
  "supply_response": 0.2,
  "rate_sensitivity": 0.15,
  "yield_compression": 0.25
}

v1.0 launched 2026-05.

Avena Score

1 version
v1.0.0activemanual
2026-05-25

Initial published weights. Derived manually from comparison against IPD residential index methodology and IMF Working Paper 19/258.

View weights JSON
{
  "risk": 0.05,
  "value": 0.4,
  "yield": 0.25,
  "quality": 0.1,
  "location": 0.2
}

Frozen at launch (2026-05). All subsequent revisions must be documented here.

Automated Valuation Model

1 version
v1.0.0activemanual
2026-05-25

AVM — town × type median €/m² base with multiplicative adjustments. ±3% RMSE on Spanish coastal backtest.

View weights JSON
{
  "energy_band": 0.08,
  "size_adjust": 0.15,
  "town_median": 0.55,
  "view_premium": 0.1,
  "beach_distance": 0.07,
  "amenity_pool_golf": 0.05
}

v1.0 launched 2026-05.

Counterpart Score

1 version
v1.0.0activemanual
2026-05-25

Counterpart Score — developer credit grade. Weights derived from a manual review of factors predicting Spanish promoter insolvency 2008-2015.

View weights JSON
{
  "payment_delay": 0.3,
  "delivery_delay": 0.15,
  "legal_disputes": 0.2,
  "court_judgements": 0.2,
  "financial_stress": 0.15
}

v1.0 launched 2026-05.

Regulatory Radar Classifier

1 version
v1.0.0activemanual
2026-05-25

Regulatory Radar classifier — Claude Sonnet 4.5 + keyword pre-filter + confidence floor 0.30.

View weights JSON
{
  "model": "claude-sonnet-4-5",
  "pre_filter": "keyword",
  "confidence_floor": 0.3
}

v1.0 launched alongside Commitment 8.

Score Confidence (adversarial)

1 version
v1.0.0activemanual
2026-05-25

Adversarial confidence layer — deterministic v1 heuristic. Penalises edge-case properties. v2 will swap in a trained residual model.

View weights JSON
{
  "score_edge": "additive",
  "comp_sparsity": "additive",
  "missing_inputs": "additive",
  "extreme_price_m2": "additive"
}

v1.0 launched alongside Commitment 5.

Recalibration policy

New versions are proposed when prediction_outcomes show systematic bias (mean error > 1σ from zero across ≥1,000 resolutions). Proposed versions A/B test against the active version for 7 days. Activation requires statistically significant improvement (p < 0.05) on out-of-sample data and explicit human sign-off. Methodology fingerprints (SHA-256 of the weights JSON) are committed daily to the verifiable digest at /verify.