Avena Terminal

Daily observation of the Spanish coastal new-build market — what was listed, what moved, and what quietly disappeared.

Live: avenaterminal.com · Open data: /open-data · mirror · CC BY 4.0 · DOI 10.5281/zenodo.19520064

---

What problem this solves

Spanish coastal new-build listing data is ephemeral. A unit is listed, repriced two or three times, and then vanishes from the feed — sold, withdrawn, or quietly reshuffled by the developer. No public registry records any of that at listing level. Once a listing changes, its previous state is gone.

The Spanish land registry publishes transactions, but with a long lag, no listing identity, and no asking price — so it cannot answer the question that actually matters to a buyer: was this unit reduced before it sold, and by how much?

So this system writes that history down, once a day, and never stops. The value is not the code — it is the fact that the series exists and cannot be backfilled. A competitor starting today is 70 days behind permanently, and money will not close the gap.

The ledger

Measured 2026-10-13. These are derived live from the database by scripts/capability-stats.ts, which refuses to emit zeros rather than report a failed read as an empty result.

Observation days, unbroken	70 (2026-08-05 → 2026-10-13)
Units observed	2,287
Price moves recorded	657
Delistings (absorption)	276
Relistings	42
Live book	2,039 across 69 towns

Depth is measured as distinct prices per unit, never as row count. An earlier version of this pipeline wrote ~394,000 rows that contained zero price changes — a capped write loop re-inserting the same frozen snapshot nightly. Row counts in this domain lie; the genuine series starts the day the pipeline was repointed at the live feed.

How the capture works

Listing XML feed (~98 MB)
  └─ parse-feed.js          fingerprints the payload, refuses a short read
      └─ public/data.json   the book, committed nightly
          ├─ pricing-history cron   → price_snapshots   (per ref, per day)
          │                         → sold_properties   (delisting ledger)
          └─ build-open-dataset.ts  → public/open-data/*  → public mirror

Two things in that chain are worth more than the rest:

The diff is gated. Delisting and price-move detection only runs against a prior snapshot no older than four days, and only when the two books overlap by at least 50%. A broken feed therefore cannot mass-flag 2,000 units as sold. The gate reports trusted_prior: false and writes nothing rather than writing something plausible.

The scheduler is not trusted. GitHub's cron landed its 05:10 slot between 06:15 and 06:43 for five consecutive nights, and the upstream feed intermittently refuses GitHub Actions egress entirely (a fresh runner 43 minutes later gets straight through). A conditional backstop workflow with 37 tests covers both. It has since carried the capture through a 21-day unattended stretch and an 8-day one with zero missed nights.

The valuation model

scripts/avm-backtest.ts, re-measured on every change, committed to public/model-stats.json:

metric	value
Median absolute error	15.82%
MAPE	21.81%
RMSE	43.28%
Within ±10%	34.2%
Within ±20%	59.5%
n	2,039

Both caveats are published alongside every figure, because without them the numbers are not true:

1. In-sample. Town × type medians are drawn from the same book being scored, so each listing contributes to its own baseline. Out-of-sample error is higher.
2. The target is asking price, not transaction price. Spanish coastal new-build transaction prices are not publicly available at listing level.

This section exists because the site previously claimed "±3% RMSE on backtest" with no backtest in existence. Building one produced a worse number and exposed a real model bug. A defensible 15.82% is worth more than an invented 3%.

The failure mode this codebase is built against

Every serious outage here has had one shape: a missing or failed value silently becoming a zero, so a broken system looks like a working one with nothing to report.

Real instances, all fixed:

• parse-feed downloaded 0 MB, threw, and exited 0 — a green check on a dead feed.
• A citation API returned 401; the catch stored an empty result, overwriting a healthy ~33% measured rate with a fabricated 0.00% for six days.
• The pricing cron did if (!error) count += chunk and took the prior snapshot as the global max date, discarding it whenever that max was today — so any re-run diffed against nothing and reported price_moves: 0 while real moves went unlogged.
• Four routes called createClient(process.env.URL!, ...) at module top level. The ! is a lie when the variable is absent. Production had the vars, so every preview deployment broke — and the signal that a branch was broken was itself broken.
• A cron reported success on every run for 57 consecutive days while transmitting nothing, because the no-op branch returned a bare 200.

So the house rules are: a failed read returns null, never 0; a published count is derived, never a literal; an absent credential throws instead of falling back; and a run that did nothing says skipped, not success.

Several tests assert the bug, not just the fix — test-hf-push-reporting.ts includes the exact pre-fix payload and asserts it still derives to success, so the test demonstrably catches the regression rather than only documenting the repair.

Gates

All four must pass before anything reaches main:

npx tsc --noEmit                   # clean
npx tsx scripts/test-concierge.ts  # all pass
npx tsx scripts/avm-backtest.ts    # median error must not regress
npm run build:preview-sim          # full build with every secret UNSET

The last one exists because production has environment variables that preview deployments do not, so a green production build proves nothing about a branch.

Running it

npm install
export FEED_URL="<listing feed URL>"   # required; no fallback, throws if unset
node parse-feed.js                   # refresh public/data.json
npm run dev

FEED_URL is a credential — the feed has no other authentication — so it lives in the environment and never in the source.

Stack

Next.js (App Router) · TypeScript strict · Tailwind v4 · Vercel · Supabase · 65 scheduled jobs · GitHub Actions for the capture path.

Licence

Code: see repository. Data under public/open-data/: CC BY 4.0 — free to use with attribution to Avena Terminal (avenaterminal.com).