Rohan Desai

Senior Data Scientist at Boundary Arc Analytics | Cricket win-probability models | Better calibrated than my biryani portions

I build cricket win-probability models at Boundary Arc Analytics in Mumbai, mostly trying to explain what changed between “comfortable chase” and “why is everyone standing up?” My work spans ball-by-ball data, feature engineering and calibration, with a particular soft spot for T20 match situations. Outside work, I’ve been learning tabla since 2018, get to Wankhede whenever I can, and cook biryani for groups that mysteriously grow after I buy the ingredients. I’m usually up before six; it’s the only time Marine Drive and my group chats are both reasonably quiet.

Build and maintain ball-by-ball win-probability models for T20 cricket. Lead work on calibration, venue effects and live inference, and collaborate with analysts to turn model movements into explanations useful to broadcast and team-analysis clients.

Developed innings-progression features and evaluation pipelines for limited-overs cricket models. Built match-level holdout tests, investigated data-quality issues and created internal dashboards for reviewing predictions over complete innings.

Worked on demand forecasting and operational reporting for retail clients. Automated recurring SQL reports and moved from spreadsheet-heavy analysis into Python-based modelling.

Python, SQL, Machine Learning, Probability Calibration, Cricket Analytics, Feature Engineering, Time-Series Analysis, Model Validation, Data Visualisation, Live Model Monitoring

English, Hindi, Gujarati, Marathi

A win-probability chart should not need its own commentary team. This month we’ve been working on clearer explanations for the bigger movements: wicket lost, required rate changed, fewer recognised batters left. The interesting bit is separating what happened on the field from what the model inferred. “It went from 62% to 44%” is a number. Explaining why is the actual job.

Eight years into learning tabla, I still occasionally speed up a phrase because I’m pleased I remembered it. My teacher is considerably less pleased. There’s a useful parallel with modelling here: getting one difficult thing right doesn’t excuse losing the underlying structure. Today’s practice was slow, repetitive and extremely unshareable. Probably the most productive half hour of my morning.

Watching a T20 at Wankhede is a good reminder that the crowd experiences probability very differently from a model. Two boundaries and the entire stand has mentally finished the chase. One dot ball and we’re all reconsidering our life choices. Back at work, I’m looking at calibration by match phase—not just one reassuring metric averaged over everything. The last four overs deserve their own uncomfortable questions.

Small cricket-data lesson from this week: split your evaluation by match, not by delivery. Neighbouring balls share a lot of context, and a careless split can make a model look cleverer than it is. We caught this while reviewing an experimental pipeline, before it went anywhere near production. No dramatic breakthrough. Just a useful check, a corrected notebook and slightly less confidence in suspiciously beautiful results.