Loading…
Methodology
How the forecasts are made, and how far to trust them
Two separate analyses share one dataset. Each was backtested season by season, with every prediction using only data from before the season it predicts.
Dataset coverage
Data fact- 1,212 IPL matches from 2008-04-18 to 2026-05-01 (data cutoff), plus 288,226 ball-by-ball deliveries.
- IPL 2026 is partial: 43 matches. Its later results and champion are not in the data and were not invented.
- Results: 1187 wins, 16 ties (decided by super over), 9 no-results (excluded; their winner field is a placeholder).
- Cleaning: Deccan Chargers separated from Sunrisers Hyderabad (they share an ID in the source). The
team1column, which is the side batting first 75.2% of the time, is never used to orient matches. - Batters were matched by full name, never by surname (e.g.
R Sharmais Rahul, not Rohit). - Not available: squads, auction, injuries, ages, the 2027 schedule and groups.
Photo: Zoshua Colah on Unsplash · Galle International Stadium, Sri Lanka
Champion forecast: models
Baselines (coin flip, career win %), Elo, logistic regression and gradient boosting were compared on pre-match, pre-toss predictions for 2013–2026. Elo was kept: no model beat a coin flip with confidence, and the more complex models were worse.
Match-level backtest (pooled)
Historical result| Model | n | Log loss | Brier | AUC | Accuracy |
|---|---|---|---|---|---|
| coin | 882 | 0.6931 | 0.2500 | — | — |
| winpct | 882 | 0.6966 | 0.2516 | 0.532 | 52.6% |
| elo | 882 | 0.6942 | 0.2505 | 0.527 | 50.8% |
| logreg | 882 | 0.6988 | 0.2527 | 0.515 | 49.8% |
| gbm | 882 | 0.7188 | 0.2622 | 0.482 | 49.1% |
| ens_elo_lr | 882 | 0.6960 | 0.2514 | 0.518 | 50.6% |
| logreg_toss | 882 | 0.6985 | 0.2525 | 0.520 | 51.8% |
Paired bootstrap, log-loss difference
Historical result| Comparison | Mean | Low 95% | High 95% |
|---|---|---|---|
| elo - coin | +0.0010 | -0.0069 | +0.0098 |
| logreg - coin | +0.0057 | -0.0041 | +0.0167 |
| gbm - coin | +0.0257 | +0.0114 | +0.0402 |
| winpct - coin | +0.0034 | -0.0074 | +0.0140 |
| logreg - elo | +0.0046 | +0.0003 | +0.0088 |
| gbm - elo | +0.0247 | +0.0123 | +0.0373 |
| ens_elo_lr - elo | +0.0018 | -0.0003 | +0.0039 |
| logreg_toss - logreg | -0.0003 | -0.0021 | +0.0016 |
Negative means the first model is better. An interval that includes 0 means no detectable difference.
How the title probabilities are calculated
- Elo ratings (K = 4, home advantage 20) are updated through every match to the cutoff. At each season boundary they are pulled toward 1500 (carry-over 0.8).
- 2027 team strength = 1500 + shrink × (start − 1500) + Normal(0, τ²). The shrink and uncertainty τ are not fixed: each of 49 grid points is weighted by the Monte Carlo marginal likelihood of all past seasons' league results.
- Each simulated season draws (shrink, τ) from that posterior, draws team strengths, draws the two groups at random and plays 70 league matches. Summed latent margins act as the net-run-rate tie-break. Then Qualifier 1, Eliminator, Qualifier 2 and the Final.
- The probabilities are frequencies over 20,000 seasons. The Simulation Lab runs the same engine in TypeScript, tested to reproduce the Python results within Monte Carlo error.
Pre-season title forecasts, by season
Historical result| Season | Champion | P(champ) | Its rank | Favourite | P(fav) |
|---|---|---|---|---|---|
| 2013 | MI | 11.1% | 6 | RR | 13.3% |
| 2014 | KKR | 9.1% | 7 | CSK | 20.3% |
| 2015 | MI | 17.3% | 2 | CSK | 25.4% |
| 2016 | SRH | 13.1% | 3 | MI | 23.9% |
| 2017 | MI | 24.5% | 1 | MI | 24.5% |
| 2018 | CSK | 21.3% | 2 | MI | 22.6% |
| 2019 | MI | 18.8% | 2 | CSK | 29.0% |
| 2020 | MI | 22.3% | 2 | CSK | 28.0% |
| 2021 | CSK | 21.8% | 2 | MI | 27.4% |
| 2022 | GT | 7.4% | 5 | MI | 22.9% |
| 2023 | CSK | 12.2% | 1 | CSK | 12.2% |
| 2024 | KKR | 6.4% | 9 | CSK | 16.3% |
| 2025 | RCB | 6.9% | 10 | CSK | 12.3% |
Batter forecast: models
Rows are player-seasons (batted in either of the two previous seasons). A player scores 0 when absent, so availability is part of the target. The chosen model is two ridge regressions, innings per team match × runs per innings, on recency-weighted, shrunk rate features from earlier seasons only.
Season-runs backtest, 2014–2025
Historical result| Subset | Model | MAE | RMSE | Bias |
|---|---|---|---|---|
| all active batters | last_season | 63.8 | 117.1 | 9.1 |
| all active batters | career_avg | 74.8 | 119.2 | 18.5 |
| all active batters | marcel | 64.5 | 107.5 | 6.6 |
| all active batters | ridge | 62.9 | 103.9 | 5.0 |
| all active batters | rf | 62.7 | 104.2 | 2.8 |
| all active batters | gbm | 62.8 | 106.2 | 1.5 |
| all active batters | two_part | 60.4 | 102.6 | -0.2 |
| all active batters | ens_ridge_two_part | 61.4 | 102.8 | 2.4 |
| established (≥250 runs in S-1) | last_season | 165.1 | 211.2 | 91.5 |
| established (≥250 runs in S-1) | career_avg | 149.2 | 188.5 | -6.5 |
| established (≥250 runs in S-1) | marcel | 142.0 | 181.9 | 9.7 |
| established (≥250 runs in S-1) | ridge | 138.7 | 175.7 | -3.5 |
| established (≥250 runs in S-1) | rf | 138.1 | 174.9 | -1.5 |
| established (≥250 runs in S-1) | gbm | 142.6 | 179.6 | 6.4 |
| established (≥250 runs in S-1) | two_part | 136.6 | 175.2 | -2.5 |
| established (≥250 runs in S-1) | ens_ridge_two_part | 137.0 | 174.8 | -3.0 |
Paired bootstrap, MAE difference
Historical result| Comparison | Subset | Diff | Low | High |
|---|---|---|---|---|
| two_part − last_season | all | -3.4 | -5.6 | -1.2 |
| two_part − marcel | all | -4.1 | -5.3 | -2.9 |
| two_part − ridge | all | -2.5 | -3.3 | -1.8 |
| two_part − rf | all | -2.2 | -3.2 | -1.3 |
| two_part − gbm | all | -2.4 | -3.7 | -1.2 |
| two_part − last_season | established | -28.5 | -39.0 | -19.2 |
| two_part − marcel | established | -5.4 | -10.2 | -0.4 |
| two_part − ridge | established | -2.0 | -4.9 | 0.8 |
| two_part − rf | established | -1.4 | -5.7 | 2.7 |
| two_part − gbm | established | -5.9 | -11.5 | -0.2 |
Components vs naive baselines
| Target | Model | MAE |
|---|---|---|
| innings | two-part share model | 2.61 |
| innings | last season innings | 2.71 |
| innings | Marcel weighted share × matches | 2.80 |
| runs/innings (≥5 inn) | two-part rpi model | 5.90 |
| runs/innings (≥5 inn) | last season rpi | 7.57 |
| runs/innings (≥5 inn) | weighted shrunk rpi | 6.31 |
| strike rate (≥100 balls) | weighted shrunk SR | 15.86 |
| strike rate (≥100 balls) | last season SR | 20.09 |
How prediction intervals are calculated
Split-conformal intervals around the point forecast. Out-of-sample errors from 2014–2025 are normalised as (actual − predicted) ÷ (predicted + 50). Their 10th and 90th percentiles are taken separately for predictions below 150, from 150 to 350, and above 350 runs. The interval therefore always contains the forecast, and its width scales with it.
| Prediction band | 10th pct | 90th pct |
|---|---|---|
| 0–150 | -0.489 | 0.610 |
| 150–350 | -0.779 | 0.694 |
| 350–∞ | -0.587 | 0.399 |
Interval reliability (nominal 80%)
Historical result| Method | Subset | Coverage | Below | Above | Width |
|---|---|---|---|---|---|
| conformal around two-part (used) | all | 80.9% | 8.4% | 10.7% | 166 |
| conformal around two-part (used) | established | 80.1% | 8.9% | 11.0% | 469 |
| CQR quantile GBM | all | 87.1% | 0.7% | 12.1% | 179 |
| CQR quantile GBM | established | 80.7% | 3.1% | 16.2% | 441 |
Known limitations and sources of uncertainty
Squads are unknown
Team-level Elo cannot see retentions, the auction or injuries. That is the likely reason title forecasts did worse than uniform in 2022–25.
Partial 2026 season
Only 43 matches of 2026 are in the data, so the freshest evidence is a part-season.
Assumed format
The 2027 format is assumed to match 2022–26. Groups are random in each simulation, NRR is approximated, and rain no-results are not simulated.
Batter availability
Expected innings come from each batter's own playing-time history. Real 2027 selection, injuries and batting order are not known.
Small samples
13 tournament backtests and about 400 established-batter seasons. Calibration estimates are themselves uncertain.
Non-stationarity
Eras differ (expansion, Impact Player rule, mega-auctions), so historical calibration is only an approximate guide to 2027.
Reports and downloads
The two reports are separate analyses and are shown separately.
Photo credits
Each photo was checked on its own Unsplash page: "Free to use under the Unsplash License" (not Unsplash+). Attribution is not required, but credited here. No IPL, BCCI or franchise logos and no player photographs are used.
- Arun Jaitley Stadium, New Delhi by Zoshua Colah · original 4000 × 2252
- D Y Patil Stadium, Navi Mumbai by Harsadh Vikhaas Rajesh Kumar · original 4032 × 2268
- Arun Jaitley Stadium, New Delhi by Zoshua Colah · original 7144 × 4502
- Galle International Stadium, Sri Lanka by Zoshua Colah · original 5941 × 3751