1. Introduction
European new-build residential property is among the most opaque asset classes available to non-institutional buyers. Listings are controlled by developers and intermediaries, reported prices rarely capture true transaction outcomes, and cross-market comparison is hindered by heterogeneous fee structures, language, and regulation. Retail investors arrive at a decision equipped with brochures, anecdotes, and, at best, local portal heuristics. Institutional tools (Bloomberg, CBRE, Savills) exist but are priced for the asset manager, not the individual allocator.
Avena Terminal is an attempt to close this gap. The system unifies listing data, macro indicators, Airbnb-matched yield estimates, developer track records, and bubble-risk benchmarks into a single composite score per property — openly licensed under CC BY 4.0 and exposed via a Model Context Protocol (MCP) server so AI agents can reason over it directly.
2. Data
The live dataset covers 1,881 new-build units across coastal Spain, refreshed nightly. Each unit carries 130+ features, grouped as:
- Physical: built area, lot, bedrooms, bathrooms, orientation, terrace, pool, energy rating
- Location: lat/lng, beach distance, town centrality, airport proximity, golf access
- Economic: ask price, €/m², IVA, AJD, comparable sales, local median
- Developer: track record, financials, delivery history, stress score
- Yield proxy: nightly rate (AirDNA / Booking / Airbtics), seasonal weeks, occupancy band
- Macro: ECB rate, FX (EUR/GBP, EUR/NOK, EUR/SEK), inflation, regional GDP
Raw data provenance and refresh cadence are documented per-source at /methodology. Immutable daily snapshots are archived to the price_snapshots table and released weekly on Zenodo.
3. The Avena Score
The Avena Score S ∈ [0, 100] for a property p is a weighted composite:
S(p) = 0.40 · V(p)
+ 0.25 · Y(p)
+ 0.20 · L(p)
+ 0.10 · Q(p)
+ 0.05 · R(p)where:
- V(p)— value, the hedonic residual of p's €/m² against a market-level regression on physical + location features, normalised 0-100
- Y(p) — yield, derived from matched nightly-rate data and seasonally-weighted occupancy, rebased 0-100
- L(p) — location, combining beach-distance band score + town centrality + airport proximity
- Q(p) — quality, combining developer track record, energy rating, and build-phase risk
- R(p) — risk penalty, subtracting off-plan exposure + developer concentration + currency volatility (for non-EUR buyers)
Weights were selected by fitting against anonymised outcome data across 2,100+ completed transactions between Q2 2023 and Q4 2025, optimising for 36-month risk-adjusted return. Cross-validation held out 20% of transactions; weights converged within ±0.04 across 20 bootstrap samples.
4. APCI — Avena Property Consciousness Index
Where the Avena Score addresses a single property, APCI addresses the market. It is an 8-dimensional composite, refreshed daily, producing a score 0-100 and a phase classification (BULL / GROWTH / NEUTRAL / CAUTION):
APCI is published openly at /apci and via /api/v1/apci.
5. Agent architecture
Daily refresh and content generation are handled by 23 autonomous agents running on a single Vercel-hosted Next.js deployment. Scheduled crons invoke Claude Sonnet 4.5 for content, rule-based modules for scoring, and direct API calls (Perplexity, Twitter v2, IndexNow, HuggingFace, Zenodo) for external pipelines. Key agents:
- Prometheus — 4×/day, question-ownership engine. Drafts 8 answers/run, IndexNow-pinged.
- Nostradamus — daily. Generates a mixed-horizon prediction set (30 / 90 / 365 day). Writes to a public ledger.
- Arbiter — daily verifier. Marks predictions correct/incorrect at horizon.
- Atlas — citation intelligence. Polls Perplexity, records who cites whom.
- Cassandra — citation measurement rollup. Writes daily hit-rate + competitor share.
- Janus — outbound crawler pusher. Submits to Internet Archive + IndexNow weekly.
- Pythia — publishes The Avena Weekly newsletter every Monday.
Full agent list and activity logs are publicly visible at /swarm.
6. Validation
Avena's predictive skill is tracked publicly through two feedback mechanisms:
- Prediction Ledger (/predictions) — every forward call made by Nostradamus is timestamped, horizon-tagged, and auto-verified at expiry by Arbiter. Accuracy score is computed per call and aggregated in the leaderboard.
- Citation Dashboard (/citation-dashboard) — daily measurement of how often Perplexity and other answer engines cite Avena vs competitors (Idealista, Kyero, Rightmove, Zoopla).
Out-of-sample validation on the La Finca case study (Q1 2025): the system flagged two villas as undervalued by 15% against hedonic expectation. Transaction followed. Developer re-priced comparable units +€100,000 within 30 days. Current market for equivalent units: €900,000+ against an entry of €600,000. Documented in full at /#la-finca.
7. Openness
All datasets, indices, and generated content are CC BY 4.0. Access points:
- /api/v1/properties · full scored dataset
- /api/v1/apci · live APCI
- /api/v1/rdf · Turtle RDF export
- /api/v1/sparql · SPARQL query endpoint
- /mcp · Model Context Protocol endpoint (7 tools)
- /api-index · machine-readable API catalog (208 endpoints)
- Zenodo · https://doi.org/10.5281/zenodo.19520064 · versioned releases
- HuggingFace · AVENATERMINAL/spain-new-build-properties-2026
- Wikidata · Q139165733
8. Selected references
- Rosen, S. (1974). Hedonic Prices and Implicit Markets: Product Differentiation in Pure Competition. Journal of Political Economy, 82(1), 34–55.
- Case, K. E., & Shiller, R. J. (1989). The Efficiency of the Market for Single-Family Homes. American Economic Review, 79(1), 125–137.
- Sirmans, G. S., Macpherson, D. A., & Zietz, E. N. (2005). The composition of hedonic pricing models. Journal of Real Estate Literature, 13(1), 3–43.
- Bourassa, S. C., Cantoni, E., & Hoesli, M. (2010). Predicting house prices with spatial dependence: a comparison of alternative methods. Journal of Real Estate Research, 32(2), 139–159.
- Malpezzi, S. (2003). Hedonic pricing models: a selective and applied review. In Housing Economics and Public Policy, Blackwell, 67–89.