placa

placa forecasts the winner of Gran Hermano Argentina’s thirteenth season (Generación Dorada, 2026) from the eviction vote shares Telefe announces after each gala. It exists because those percentages are not an opinion poll: they are a census of the audience that actually pays to vote, and because eviction voting and final voting are different scales that barely correlate across the last two seasons, the model tracks two separate parameters per contestant, rejection (mu) and positive support (psi), rather than one.

What it does not do

• It is not affiliated with Gran Hermano, Telefe, Kuarzo or Banijay, and nobody from the production reviewed or endorsed it. The percentages are public facts published by Telefe; no rights over them are claimed. The page’s mark is a forecast-shaped eye, not the show’s logo.
• It does not simulate departures or expulsions, thirteen of which had already happened in this season by the time the limits were written. That is named as the most likely way the forecast breaks.
• It does not treat the final’s date or format as settled. Telefe had not confirmed 6 September against 31 August, and the model simulates by number of eliminations rather than by calendar; the finalist count and final mechanics are carried over from 2024-25 as inference, not fact.
• Two contestants have no gala with a complete vote breakdown, so their rejection parameter is almost entirely prior, and the page marks them with a dotted outline; two others each have a rejection parameter based on a single gala.
• It publishes no photographs of contestants. Each is shown by national flag instead, chosen because the cast this season spans five countries.

Install and run

python3 model/final_model.py && python3 model/ramas.py && \
  python3 model/evolucion.py && python3 model/bootstrap.py && \
  python3 model/camino.py && python3 gui/build.py && python3 gui/tarjeta.py

Before publishing, the same check that runs in CI:

python3 gui/verificar.py

The page recalculates only after each gala, and a deploy does not go out if the page cannot be rebuilt from the data committed to the repository.

How it works

Rejection (mu) is estimated with a conditional logit over eight galas of negative voting; positive support (psi) with a Plackett-Luce model over twelve positive-voting instances. Eight galas qualify as complete observations because Telefe’s published shares satisfy an arithmetic identity, published sum minus 100 equals the sum of the shares outside the final pair, in 8 of 11 galas, which is also the check the automatic updater runs against each new gala. Rejection is not a fixed trait: one contestant’s estimate moved from 38.45% to 0.10% across three weeks, and the variance decomposition splits roughly 52% stable trait and 48% weekly noise. An external poll is folded in weighted by its own measured accuracy, which the repository states is good for first place only and much worse from second place down.

The data behind the page is a public file, web/datos.json, readable without touching the HTML.

Live page: https://nerln.github.io/placa/

placa on GitHub, Apache-2.0 licensed.