InspirationI built CricWP, a ball-by-ball win-probability model and interactive dashboard for the team batting second in T20 cricket. I couldn’t download real match data here, so everything you see in the preview runs on simulated matches, not real games. The dashboard shows a banner saying so.
On your machine, pip install -r requirements.txt then python run.py --fetch pulls real T20I data from Cricsheet and rebuilds dashboard/index.html. I haven’t been able to run it on real data, so expect a possible parsing hiccup on the first fetch. Adding --source ipl uses IPL matches instead.
Problem: the scoreboard shows what has happened in a chase but not how close the game really is. Data: Cricsheet ball-by-ball data, with the state (runs needed, balls left, wickets lost) recorded after every delivery of the chase. Model: gradient boosting that can never say more runs needed helps the chasing team, or that more balls or wickets in hand hurt it. It is tested on the newest 20% of matches, which it never saw in training. Dashboard (works offline): Match replay: scrub through a chase with wickets and biggest swings marked. What-if: set a score situation and see the win probability, a full probability map, and what the next ball would do. Player impact: leaderboards for how much each batter and bowler added to the win probability. Model & data: accuracy against baselines and a calibration plot.
On the simulated data the model’s AUC is 0.88 and it is well calibrated, but a simple logistic baseline scores about the same, since the model only sees three inputs. The README says this plainly. Adding batter and bowler ratings is the obvious way to give the model a real edge.
The four tests (parser consistency, monotonic win-probability surface, curves ending at the actual result) pass, and the dashboard loads in a headless browser with no errors.
The zip has the full project with README and MIT licence, and the dashboard preview is the separate index.html.
Sources: Cricsheet, the data source the pipeline downloads (ODC-By licence, so credit it if you publish results).
Log in or sign up for Devpost to join the conversation.