← all projects
Case study · simulation under uncertainty

IPL 2026 Simulator
modeling futures, not picking winners.

Can you predict a champion in a league famous for chaos? Not with a single answer — but you can model the distribution of possible seasons. An ensemble trained on 18 seasons of ball-by-ball data, run through 2,000 Monte Carlo simulations, gave RCB the top title probability at 25.4%. RCB won.

Live dashboard ↗ Repository ↗
18
SEASONS OF DATA
279K
DELIVERIES PARSED
49
FEATURES · 7 GROUPS
2,000
SIMULATED SEASONS
54.9%
WALK-FWD ACC. '24
IPL 2026 prediction dashboard showing championship win probability and playoff qualification probability for all ten teams
The live dashboard — championship and playoff probabilities across all 10 teams, re-simulated as matches complete.

01 · The problem

T20 cricket is one of the most stochastic team sports there is. Match outcomes swing on match-day pitch behavior, dew, dropped catches, and individual brilliance — none of it modelable from history. Published research tops out around 55% match accuracy; even professional betting markets only reach 55–58%.

So the goal was reframed: don't predict a single outcome — model the distribution of possible futures. "RCB wins in 25% of simulated seasons" is a fundamentally more honest and more useful output than "RCB will win."

02 · Architecture

Cricsheet ball-by-ball1,175 match JSONs · 279K deliveries · 2008–2026
Kaggle pitch data344 matches · pitch type · dew risk
2026 live resultsupdated as matches complete
FEATURE ENGINEERING · 49 FEATURES / 7 CATEGORIES
Elo (20% season decay) · form · H2H · venue · powerplay/death phase stats · pitch type · run margins
ENSEMBLE · GRIDSEARCH CV
XGBoost 50% weight
Logistic Regression 25%
Random Forest 25%
MONTE CARLO · 2,000 FULL SEASONS · PLAYOFF BRACKETS · INJURY ADJUSTMENTS · PROBABILITY SMOOTHING
Championship & playoff %per team, P10–P90 points ranges
Orange & Purple Capplayer projections with uncertainty
Streamlit dashboardlive on Streamlit Cloud

03 · Validation — walk-forward, no leakage

The model was backtested the only honest way for time-series sports data: train strictly on the past, predict a full unseen future season. Then compared against three baselines — because 54.9% means nothing until you know what naive strategies score.

SEASONOUR MODELELO ONLYHOME FAVOREDRANDOMLOG LOSS
2023 (train 2009–22)52.1%46.6%42.5%50.0%0.683
2024 (train 2009–23)54.9%47.9%56.3%50.0%0.695

Worth reading closely: in 2024 the "home favored" baseline actually edged the model on raw accuracy — that's on the page because hiding it would be cherry-picking. The model's value isn't single-match accuracy; it's calibrated probabilities across all matches and teams, which naive baselines can't produce. In a domain this stochastic, a consistent 2–5 point edge over random with good log loss is a real signal.

Six-chart suite: championship and playoff probabilities, walk-forward backtest accuracy, Orange and Purple Cap projections with error bars, current standings
The visualization suite — probabilities, walk-forward backtest vs the 50% random line, and player projections with uncertainty bars.

04 · The call — and the receipts

As of match 13, the simulation gave RCB the highest championship probability at 25.4% (DC 21.9%, RR 18.8%), with CSK collapsed to 1.6% — their early-season struggles already priced into the Elo and form features. RCB went on to win IPL 2026.

A 25.4% call being "right" deserves honest framing: the model didn't say RCB would win — it said RCB won in a quarter of possible futures, more than anyone else. That's exactly what a calibrated model of an unpredictable league should sound like. Alongside the title race, the simulator projected individual awards with P10–P90 uncertainty ranges:

Orange Cap and Purple Cap projection tables showing projected runs and wickets with low and high ranges per player
Orange & Purple Cap projections — point estimates with honest ranges, e.g. Rizvi ~805 projected runs (709–907).

05 · Design decisions

Elo with 20% seasonal decay.

IPL squads reshuffle every auction. Standard Elo assumes stable teams, so ratings revert 20% toward base each season — teams carry 80% of earned strength forward.

Monte Carlo over point predictions.

2,000 simulated seasons — including full playoff brackets and per-team injury adjustments — produce a probability distribution instead of a guess.

Probability smoothing against overconfidence.

Raw model outputs are too extreme. Temperature scaling (75% model + 25% prior) keeps the championship distribution realistic across all 10 teams.

Feature importance told cricket truths.

Home ground is the single strongest feature (~56% home win rate); powerplay and death-over strike rates outrank overall win rate; GT wins 0% on spin pitches vs 84% on balanced ones.

Roadmap: Bayesian probabilistic modeling, live score ingestion, transformer-based sequence models, and a player-level simulation engine.