Squadro Match Lab

Predict: win probability with the seven factors behind it, plus a live update after game 1
Draw odds: exact title chances for a 16-entry knockout and how much the draw moves them
Scout: power rankings, rating history since 2021 and wins above expectation
Insights: six findings from 10,852 unseen matches, each with a 95% interval and a coaching action
Evidence: frozen-model audit (ROC-AUC 0.808 in 2025, 0.798 in 2026) and live reproducibility check

Inspiration

I play and follow badminton, and for my final-year project I scraped five seasons of BWF World Tour results and built a match-prediction model. It worked, but it lived behind a Python server and a long report. A coach preparing for next week's tournament cannot use either. I wanted to turn that research into something a coach, analyst or player could open on a phone and act on.

What it does

Squadro Match Lab answers the questions a coach asks before and during a tournament.

Predict: pick any two players or pairs in any of the five disciplines and get a win probability, plus the seven factors behind it in plain English (rating gap, recent form, career record, experience, head-to-head, rest days, rating certainty). Each factor shows how far the probability would move without it.
Live after game 1: enter the game-1 score and the forecast updates. On all 3,822 matches of 2026, which this model never saw, ROC-AUC rises from 0.798 before the match to 0.901 after game 1.
Match context: optional adjustments for home advantage and three-game fatigue, two effects the base model cannot see.
Draw odds: exact title odds for a 16-entry knockout and a "draw luck" view that shows how much the draw alone moves each contender.
Scout: power rankings, rating history since 2021, form, and 2026 wins above expectation, which flags players whose rating has not caught up with their results.
Insights and Evidence: six findings from 10,852 out-of-sample matches, each with a 95% interval and a coaching action, plus the full audit and a live reproducibility check.

Six findings for coaches

Home players win 3.3 points more often than the model expects (95% CI +0.4 to +6.1).
Coming off a three-game match costs 3.1 points of win probability next round (CI −5.0 to −1.1).
82.5% of game-1 winners take the match; an underdog who wins game 1 wins 57.6% of the time.
Men's singles is the hardest event to call (ROC-AUC 0.73; one favourite in three loses). Women's doubles is the easiest (0.86).
Upsets peak in finals: favourites win 74% in the round of 32 but 65% in finals.
Under a 50-point Glicko-2 gap the stronger side wins 51%; above 450 points, 93%.

How I built it

Data: 29,792 completed World Tour matches (Jan 2021 to Aug 2026), collected with my open-source scraper BWFWebScrap and cleaned in my FYP pipeline.
Model: an order-symmetric logistic regression over Elo, Glicko-2, form, experience, rest and head-to-head features. It scores each fixture both ways round and averages in log-odds, so P(A beats B) = 1 − P(B beats A) exactly. Trained on 2022–2024 and frozen before 2025.
Making it portable: because of that symmetry, the intercept and every side-invariant input cancel algebraically, leaving a short weighted sum of differences. I extract those weights from the scikit-learn bundle and assert that the portable model reproduces the served model on all 10,852 audit matches, to a largest difference of 3.3 × 10⁻¹⁶. The app re-runs the check in the browser.
New analyses: residual analyses against the frozen model's out-of-sample predictions, with tournament-cluster bootstrap intervals; an in-match model fitted on 2025 and tested on 2026; an exact dynamic-programming bracket calculator.
App: a single static page (HTML, CSS, vanilla JavaScript, hand-drawn SVG charts) served from GitHub Pages. No server and no keys.

Challenges I ran into

A fatigue effect that was too good. My first "second match of the same day" analysis showed a +29-point effect. It was leakage: match IDs within a day are not in round order, so a "previous" match was sometimes a later one. Ordering by round fixed it, and the effect became inconclusive (126 matches), so I dropped it. The three-game fatigue effect survived the fix.
Proving the browser model is the real model. Rounding the exported ratings produced differences of 8 × 10⁻⁵. Exporting full precision brought the live check down to 2.2 × 10⁻¹⁶.
Separating a real effect from recalibration. Adding home and fatigue terms improved 2026 log loss, but most of that gain came from rescaling the model's margin. I report the effects with their intervals and keep the adjustment optional instead of claiming a better model.

Accomplishments that I'm proud of

Every probability in the app can be traced to its factors and reproduced exactly from the frozen model.
The findings are measured on seasons the model never saw, with honest intervals, and each one ends in an action a coach can take.
It runs instantly on a phone with no backend.

What I learned

Statistical honesty is a feature. The most useful thing to show a coach is often not a single number but how sure that number is: the men's singles caveat, the draw-luck range, the effect with its interval.

What's next for Squadro Match Lab

Refresh ratings weekly from the scraper so forecasts track the live tour.
Load real tournament draws as soon as BWF publishes them and compute everyone's path to the title.
A point-by-point live model for broadcasters and bench coaches.
Calibrate probabilities per discipline before using them for operational decisions.