Hedonic Regression for European New-Build Property Valuation: An Open Methodology

1. Introduction

Property valuation in Europe is fragmented. Closed proprietary models (Zillow's Zestimate, Zoopla's AVM, CoStar's index) dominate the US and UK, but no European equivalent is open, reproducible, and institution-grade. The Avena Score fills that gap: a public composite that scores individual new-build properties against their local market, with every component exposing its reasoning.

We deliberately optimize for auditability over opaque accuracy gains. A 2% improvement in back-tested error is not worth a loss of reproducibility; analysts, regulators, and academics must be able to inspect every weight.

2. The Composite

The Avena Score for a property i at time t is:

S_i,t = 100 × clamp(0, 1, Σ w_k · s_k(i, t))

where   k ∈ {V, Y, L, Q, R}
        w_V = 0.40    Valuation
        w_Y = 0.25    Yield
        w_L = 0.20    Location
        w_Q = 0.10    Quality
        w_R = 0.05    Risk

Sub-scores s_k each map feature inputs to [0, 1]. Weights are fixed in v1.0 based on the back-tested contribution of each dimension to capital appreciation + yield over the 2023–2025 Spanish new-build cohort.

3. Sub-Score Construction

3.1 V — Valuation

gap_i = (pm2_comp - pm2_i) / pm2_comp
s_V   = clamp(0, 1, 0.5 + 2 · gap_i)

Where pm2_comp is the town-median €/m² when available, otherwise the regional median computed from at least 5 local scored comparables. Properties priced at the comp median score 0.5; a 25% discount scores 1.0; a 25% premium scores 0.0.

3.2 Y — Yield

monthly_rent_i = max(400, pm2_i · built_m2_i · 0.0004)
yield_gross_i  = 12 · monthly_rent_i / price_i
s_Y            = clamp(0, 1, (yield_gross_i - 0.02) / 0.06)

The 0.04% per €/m² monthly-rent ratio is the baseline fitted from Costa Blanca + Costa del Sol seasonal long-lets 2023–2025. Override via Avena Terminal's live _yield.gross field when available.

3.3 L — Location

Region tier lookup (costa blanca 0.78, costa del sol 0.82, algarve 0.82, paris 0.92, ...) with beach-distance adjustment: ≤ 0.5 km → +0.12, ≤ 1.5 km → +0.06, > 10 km → −0.08. Capped at 0.95 top, 0.30 floor.

3.4 Q — Quality

Property-type base (villa 0.75, penthouse 0.78, apartment 0.62) + size-band bonus (+0.05 for 120–250 m²) + bedroom bonus (+0.03 for 3+). Future v1.1 adds finish-grade via developer embeddings.

3.5 R — Risk

The public engine uses a static default (0.6) so reproduction doesn't require licensed macro-feed access. Internal Avena Terminal substitutes a 20-feature regime model (ECB path, 10Y spreads, HICP, unemployment, liquidity) with the same weight.

4. Reproducibility

The reference implementation is published at github.com/avenaterminal/avena-score under the MIT license. Training data is accessible via:

• · GET /api/v1/properties — scored inventory
• · GET /api/v1/market — town medians
• · GET /api/v1/bubble-scanner — regional regime signal

Every call returns X-Data-License: CC BY 4.0 headers. The full daily-close history of the flagship AVENA Index is available at /api/v1/indices/avena?history=all&format=csv.

5. Limitations

v1.0 is deliberately narrow: it scores new-builds, uses hedonic sub-scores rather than end-to-end neural regression, and folds risk into a static default in the public engine. These are defensible for auditability but concede 1–3 percentage-points of back-tested error vs. black-box models. Community contributions that improve accuracy without sacrificing auditability are welcome at the 2026 Scoring Challenge.

6. Citation

@article{kolstad2026avena,
  title   = {Hedonic Regression for European New-Build Property Valuation:
             An Open Methodology},
  author  = {Kolstad, Henrik},
  journal = {Avena Terminal},
  year    = {2026},
  doi     = {10.5281/zenodo.19520064},
  url     = {https://avenaterminal.com/research/avena-score}
}