Data source

eloratings.net publishes a “starting” Elo snapshot for every year, e.g. https://eloratings.net/2026_start. The page itself renders client-side (a JS grid) and has no public API, but the grid is populated from two plain TSV files that can be read directly:

• https://eloratings.net/{year}_start.tsv — one row per team, ranked by rating, with 31 tab-separated columns: rank-change arrow, rank, team code, rating, highest/average/lowest rank and rating, rank/rating changes over the trailing 3 months, 6 months, 1, 2, 5, and 10 years, total matches played (home/away/neutral), wins/losses/draws, and goals for/against.
• https://eloratings.net/en.teams.tsv — maps each 2-letter team code to its full country name.

Building the panel

Starting Elo ratings are pulled for every calendar year from 1930 through 2026 — not just World Cup years — and stacked into a single panel, so downstream posts can look up any team’s rating in any year without re-scraping. Years already present in the output file from a previous run are skipped, so re-running this post only hits eloratings.net for new years.

Normalizing country names

Output

The combined panel — sourced from eloratings.net — is written out for downstream use by the data-cleaning and hosting-premium posts:

Visual check

As a sanity check that the scrape actually worked, here’s Elo rating over time for two teams read straight back from the parquet file:

Data sources

Output

The cleaned dataset is written out for downstream use (and eventual publication to Kaggle):

Model definition

Methodology

Dataset

Difference-in-differences (DiD) measures the effect of a treatment by comparing the change over time in a treatment group against the change over time in a control group.

The raw dataset is cleaned in separates data-cleaning posts for the matches and for the ELO ratings, which documents every transformation — column renaming, deduplication, dtype coercion, and stage-label normalization across World Cup rule changes — and produces the cleaned datasets used below.

Fitting the model

One possibility in fitting the model was using statsmodels and C(team_id). However, this forces to explicitly construct a matrix column for every nation in the dataset — the Least Squares Dummy Variable (LSDV) approach. As the number of teams grows, inverting the matrix becomes slow, memory-heavy, and clutters the summary table with hundreds of unwanted coefficient rows.

We choose instead an alternative approach where we use linearmodels, which de-means and absorbs fixed effects without ever constructing dummy variables.

Conclusions

The TWFE estimate suggests hosting the World Cup is associated with a 1.02 goal-differential-per-game increase (95% CI: 0.55–1.49, p < 0.001) relative to a team’s own non-hosting baseline, after absorbing team- and tournament-level fixed effects and controlling for pre-tournament Elo. This is a large effect in practical terms — for context, a team that typically draws (0.0 avg goal diff) would be expected to outscore opponents by roughly a goal a game while hosting.

The pre-tournament Elo control is also significant (p = 0.047) but its coefficient is small (0.002 per Elo point), consistent with Elo capturing gradual team quality rather than being a strong predictor of a single tournament’s goal differential.

The model’s within-R² (0.079) is modest: fixed effects and Elo explain only a small share of within-team variation in tournament performance, which is expected given the high game-to-game variance inherent in short single-elimination-heavy tournaments (avg. ~5 games/team). The F-test for poolability (p = 0.0004) rejects the null that entity/time effects are jointly zero, supporting the TWFE specification over pooled OLS.

Taken together, the results are consistent with a genuine home-hosting advantage in the World Cup, though the estimate should be interpreted as an average effect across only 21 hosting instances — a small effective treatment count that warrants caution before treating the point estimate as precise.