Who's going to score?

It's the question every supporter asks before kick-off. Will Wainiqolo bag another brace? Will Bielle-Biarrey find the corner? Rather than going on gut feeling, I built a model that estimates the try-scoring probability for every starting player — before the game begins.

In this article I walk through the methodology, the data, the 2025-2026 season results, and some surprising insights from the backtest.

Live predictions are available on the Analytics page 👈

1. The data

The model draws on three data sources collected for every TOP 14 and PRO D2 match:

Match sheets: for each game I gather the official line-up (15 starters), individual player statistics, and the full match timeline (tries, cards, substitutions).

Historical results: all matches across multiple seasons, including scores and team statistics.

Official standings: points scored and conceded per round, used as a proxy for each team's offensive and defensive level.

2. Model features

The model uses 18 features to estimate the probability that a given player scores in a given match.

Player-level

• player_try_rate: rolling average of tries scored over the last 10 games (each prediction uses only past matches)
• games_played_season: games played this season — a proxy for the coach's trust and the player's consistency

Team-level

• team_attack_rate: average tries scored per game over the last 10 games
• opp_defense_rate: average tries conceded by the opponent per game
• expected_match_tries: sum of the two above — a measure of how open the game is likely to be. A tight Toulouse-Racing clash won't create the same opportunities as two free-scoring sides going at each other
• is_home: home advantage, measurable even for individual try scorers

Relative strength

• elo_team and elo_diff: team ELO rating and differential against the opponent
• rank_pts_scored_pg, rank_opp_conceded_pg: points scored/conceded per game in the standings

Position: 8 binary variables — wing, fullback, centre, halfback, flanker, back row, lock, front row.

3. The model

Architecture

I use logistic regression with balanced class weights to handle the natural imbalance: a player only scores in roughly 20% of their matches. The model is trained on all seasons prior to 2025-2026, then evaluated on the current season.

Normalisation

Raw probabilities are normalised per team so that their sum equals the expected number of tries in the match. If a team is expected to score 3 tries and one winger captures most of the raw probability mass, his normalised score can reach 50-60%.

Platt Scaling

Platt scaling is a calibration technique introduced by John Platt (1999). The idea: after training the main model, a second sigmoid function is fitted that maps the model's raw outputs to better-calibrated probabilities. Concretely, if the model tends to be overconfident at 70%, the calibration pulls it toward a more realistic value.

In my implementation, I use 5-fold cross-validation (CalibratedClassifierCV, method='sigmoid', cv=5): five independent models are trained on different data subsets and their predictions averaged. This ensemble effect improves player ranking (AUC +9%) beyond simple calibration alone.

⚠️ The scores displayed are relative rankings, not absolute probabilities. A player at 60% won't score in 60% of their games — it means they are the most likely scorer for their team in that specific match-up.

Feature Importance

Logistic regression is fully interpretable: after standardisation (StandardScaler), every model coefficient is directly comparable. Each value represents the effect of one standard deviation of that feature on the log-odds of scoring a try, holding all other features constant.

Five findings stand out clearly:

1. Individual history dominates. player_try_rate — the 10-game rolling try average — is the strongest individual signal (+0.30 in TOP 14, +0.28 in Pro D2). Players who score consistently keep scoring: regularity is structural, not random.

2. Position is a blueprint. Being a wing is almost as predictive as individual history (+0.29 in TOP 14). Conversely, lock (-0.10 / -0.14) and front row (-0.13 / -0.07) carry negative coefficients. These capture each position's intrinsic role in the game — wingers finish phases, props create them.

3. Standings beat rolling averages. Among collective context features, the standings-based indicators (rank_pts_scored_pg, rank_opp_conceded_pg) have two to four times the impact of raw rolling averages (team_attack_rate, opp_defense_rate). Season-long rankings are a more stable quality proxy than the last 10 games.

4. Home advantage is real and sizeable. The is_home coefficient is +0.20 in TOP 14 and +0.21 in Pro D2 — one of the most consistent and strongest effects among non-positional features. Playing at home increases try-scoring probability even at the individual player level.

5. ELO adds little at the player level. The ELO differential between the two teams has a coefficient of only +0.018 — the weakest of all continuous features. At match scale, relative team strength is already captured by the standings variables; at the player level, it becomes marginal.

These results are remarkably stable across TOP 14 and Pro D2, which gives the findings added robustness.

4. Results — 2025-2026 season

How to read AUC and top-k hit rate

AUC (Area Under the ROC Curve) is the standard metric for ranking models. Here's the intuitive interpretation: imagine randomly picking one player who scored and one who didn't in the same match. AUC is the probability that the model gave the scorer a higher score. An AUC of 0.5 means the model performs no better than chance; 1.0 means perfect. Our 0.596 in TOP 14 and 0.612 in PRO D2 mean the model correctly identifies the scorer in nearly 60% of such comparisons — on an inherently difficult problem where rugby keeps all its unpredictability.

The top-k hit rate is more practical: in what percentage of matches did at least one of the model's top-k picks actually score a try?

Hit 5 performance — By selecting the 5 highest-ranked players across both teams, the model covers a real try scorer in 85% of TOP 14 matches and 78% of PRO D2 matches. That means only 1 in 6 matches (TOP 14) produces a try scorer outside the top 5 picks.

Scoring by position

Wings and fullbacks dominate — expected. More interesting: the model systematically underestimates back rows and flankers relative to their actual scoring rate. These positions have a less predictable but real try-scoring profile — contact offensives, mauls that burst through, forwards appearing at the tail of a lineout. The model still has room to improve on these atypical profiles.

Most predictable teams

LOU Rugby is the most predictable team (52%): the concentration of tries around Wainiqolo makes prediction relatively reliable. Racing 92 and Stade Toulousain are tied at 48% — but for opposite reasons. Racing has a few regularly-identified scorers, whereas Toulouse is the championship's most prolific team (72 tries across the season, 27 different scorers), which actually dilutes individual predictions. When literally anyone in the squad can score, pinpointing exactly who is harder.

At the other end, USA Perpignan and Stade Rochelais are the hardest to call (24%). Perpignan scores rarely (18 tries total); La Rochelle, despite 55 tries, spreads them across 17 different players.

Players who surprise the model

Gaël DREAN (Toulon) is the model's biggest miss: +40 percentage points above prediction. An actual rate of 71% against a predicted 31%. The model structurally underestimates him, likely because he plays in a high-density attacking team where his individual profile is diluted in the overall picture.

Conversely, some highly-rated players underdelivered this season — a reminder that even the strongest predictions are probabilistic estimates, and that rugby always keeps its element of surprise.

Most consistent scorers

Jiuta Naqoli WAINIQOLO (18 tries in 25 rounds, 13 different matchdays) leads the rankings by a clear margin. Louis BIELLE-BIARREY (13 tries) is a strong second in TOP 14. Both have in common that the model ranks them #1 in the vast majority of their matches — confirming that their consistency is structural, not accidental.

Weather effect

The analysis reveals a clear weather effect on which positions score tries. Centres are hit hardest, with a drop of -12.7 percentage points in the rain (23.8% dry → 11.1% wet). Fullbacks and wings also lose ground (-7.4pp and -6.7pp respectively). The logic is clear: rain reduces wide play, offloads and line breaks that benefit backline positions.

Conversely, halfbacks see their scoring rate rise significantly (+9.1pp, from 15.9% to 25.0%), alongside back rows (+7.2pp) and, to a lesser extent, front rows (+2.7pp). In wet conditions, the game becomes more direct, more physical, closer to scrums and rucks — you'd naturally expect front row forwards to benefit through picks-and-gos and dominant mauls. The data does show a positive trend for front rows in the rain, though the effect is more moderate than for halfbacks or back rows.

Worth noting: these figures are based on a limited number of rainy matches this season. The trend is consistent with the logic of the wet-weather game, but should be treated as directional rather than conclusive until confirmed across multiple seasons.

5. What the model doesn't know yet

Several interesting improvements remain to explore:

• Minutes played: a full 80-minute starter has different opportunities to an impact substitute
• Individual ELO: two wingers on the same team don't have the same skill level
• Weather as a training feature: the data is available and the results above show a real signal
• Current streak: does a player who has scored in the last 3 matches have a higher chance next week?
• Head-to-head records: some players perform very differently depending on the opposition

Conclusion

Predicting try scorers remains a challenge — rugby is unpredictable, which is what makes it great. But the data shows there is real, exploitable signal: players who score consistently tend to keep doing so, and attacking teams facing leaky defences structurally create more opportunities.

Predictions are updated each round on the Analytics page. If you want to see who the model picks for the next match, that's where to look 👈

If you have ideas for features to add or questions about the methodology, feel free to reach out!