Beat the Avena Score.
We open-sourced our property scoring engine. Now we're inviting the ML + PropTech community to beat it. Submit your own scoring model against the held-out 2026 backtest. Leaderboard is public, methodology is public, winning model will be cited in the next Avena Score release notes.
Rules of engagement.
- Use the open data. Training set is /api/v1/properties (CC BY 4.0). Town medians at /api/v1/market. Historical closes at /api/v1/indices/avena?history=all.
- Output a 0–100 score per property. Submissions return a JSON mapping
ref → scorefor every ref in the holdout set. Your model can be anything — XGBoost, a transformer, another hedonic decomposition, pure heuristics. - Evaluation. We rank submissions by Spearman rank correlation between your scores and realized 12-month price change on the 2026 holdout set (published 2026-12-31). Ties broken by mean absolute score-error vs. our v1.0 baseline.
- Submit. Open a PR to
github.com/avenaterminal/avena-score/tree/main/submissions/2026with a directory named after your handle, containing a README, your code, and apredictions.json. - Public leaderboard. Published at this URL. Updated on every accepted PR.
- Winner gets cited. Top submission is cited in the Avena Score v1.1 methodology paper and feature-released on /changelog. Additional recognition may include paid collaboration on the v2 model.
Public leaderboard.
| # | Submission | Method | Spearman ρ | MAE |
|---|---|---|---|---|
| — | avena/baseline-v1.0 | hedonic · MIT | 0.——— | — |
| First submissions accepted now · holdout evaluation publishes 2026-12-31 | ||||
Common questions.
Do I need to open-source my submission?
Yes. Every accepted submission is MIT-licensed on merge. Private models can still query /api/v1/score but cannot claim the leaderboard spot.
Can I use the baseline engine as a starting point?
Absolutely. Fork github.com/avenaterminal/avena-score, improve it, submit the diff. Incremental improvements over v1.0 are the most common winning path.
How is the holdout selected?
A random 10% of the 1,881 Spanish working set is held out at the start of the challenge window (2026-04-24 snapshot). Holdout refs are revealed 2026-12-31 along with realized outcomes.
What counts as a "realized outcome"?
12-month change in price/m² of the same property, or of nearest-neighbor properties if the specific unit has no observed resale. Computed by Agent Arbiter with source-of-truth comps from idealista/kyero/fotocasa.
Is there a prize?
Recognition + citation in v1.1 release notes + potential paid collaboration on v2. A cash prize may be added in future cycles.
Baseline engine: github.com/avenaterminal/avena-score
Questions? henrik@avenaterminal.com