pith. sign in

arxiv: 2604.13861 · v2 · submitted 2026-04-15 · 💻 cs.LG · stat.AP

Simulation-Based Optimisation of Batting Order and Bowling Plans in T20 Cricket

Tinniam V Ganesh This is my paper

Pith reviewed 2026-05-10 13:42 UTC · model grok-4.3

classification 💻 cs.LG stat.AP
keywords T20 cricketbatting order optimizationbowling plan optimizationMonte Carlo simulationMarkov Decision Processwin probabilityIPLsimulation-based optimization
0
0 comments X

The pith

Optimal batting orders and bowling plans raise T20 win and defend probabilities by 4.1 and 5.2 percentage points in two IPL test cases.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a unified Markov Decision Process framework that selects batting orders and bowling plans by directly maximizing simulated win or defend probabilities instead of expected runs. It builds three-phase player profiles from historical IPL data using James-Stein shrinkage and scores every candidate decision through vectorised Monte Carlo simulation of 50,000 innings trajectories. When applied to two 2026 IPL matches, the optimised batting order lifts Mumbai Indians' win probability from 52.4% to 56.5% and the optimised bowling plan lifts Gujarat Titans' defend probability from 39.1% to 44.3%. The gains arise because many standard choices ignore phase-specific performance differences that the model captures. If the simulations track real outcomes, teams gain a repeatable method for turning data into higher-probability tactics.

Core claim

A three-phase player profile engine with James-Stein shrinkage, estimated from 1,161 IPL ball-by-ball records, feeds into Monte Carlo evaluation of all feasible batting-order permutations and guided search over bowling-over assignments; this process identifies arrangements that improve win probability by 4.1 points and defend probability by 5.2 points over the plans actually used in two 2026 IPL matches, showing that phase-agnostic deployment is measurably costly.

What carries the argument

Vectorised Monte Carlo simulation over 50,000 innings trajectories inside a Markov Decision Process that scores every feasible batting order and bowling plan by its direct effect on win or defend probability.

If this is right

  • Batting orders are chosen by exhaustive comparison of all remaining-player arrangements to maximise the simulated win probability.
  • Bowling plans are refined through guided search that respects constraints such as no consecutive overs by the same bowler.
  • Decisions based on aggregate statistics prove sub-optimal once phase-specific profiles are applied.
  • Both batting and bowling optimisations are expressed uniformly in terms of match outcome probability rather than expected runs.
  • The same simulation engine can be reused for any T20 match given sufficient historical ball-by-ball data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Teams could run the model in real time to adjust plans as wickets fall or the pitch changes.
  • The approach could be extended to ODIs or other formats where phase performance matters.
  • Systematic comparison across many matches might reveal consistent biases in how coaches currently allocate overs and batting positions.
  • Incorporating live pitch or weather data into the profiles would test whether the current gains hold under more dynamic conditions.

Load-bearing premise

Historical phase-specific profiles with shrinkage, when fed into Monte Carlo trajectories, accurately predict real-match win probabilities without opponent adaptation, pitch changes, or other unmodeled in-game factors.

What would settle it

Running the optimised batting order and bowling plan in a live match and checking whether the realised win or defend rate matches the simulated 4-5 percentage point lift.

Figures

Figures reproduced from arXiv: 2604.13861 by Tinniam V Ganesh.

Figure 1
Figure 1. Figure 1: MI batsman phase-specific profiles (KKR vs MI, over 12 intervention). The [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Best achievable win probability by next-in choice at Position 3. SA Yadav’s [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: GT bowler landscape at the over 10 intervention point. The optimal Death [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Best achievable defend probability by over 10 opener choice. Siraj’s two-over [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Top 10 GT bowling plans heatmap. The vertical divider separates Middle [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Actual vs optimal GT bowling plan (overs 10–19). The total gain of +5.2 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
read the original abstract

This paper develops a unified Markov Decision Process (MDP) framework for optimising two recurring in-match decisions in T20 cricket, namely batting order selection and bowling plan assignment, directly in terms of win and defend probability rather than expected runs. A three-phase player profile engine (Powerplay, Middle, Death) with James-Stein shrinkage (a technique that blends a player's individual statistics toward the league average when their phase-specific data is sparse) is estimated from 1,161 IPL ball-by-ball records (2008-2025). Win/defend probabilities are evaluated using vectorised Monte Carlo simulation over N = 50,000 innings trajectories. Batting orders are evaluated by comparing all feasible arrangements of the remaining players and selecting the one that maximises win probability. Bowling plans are optimised through a guided search over possible over assignments, progressively improving the allocation while respecting constraints such as the prohibition on consecutive overs by the same bowler. Applied to two 2026 IPL matches, the optimal batting order improves Mumbai Indians' win probability by 4.1 percentage points (52.4% to 56.5%), and the optimal Gujarat Titans bowling plan improves defend probability by 5.2 percentage points (39.1% to 44.3%). In both cases, the observed sub-optimality is consistent with phase-agnostic deployment: decisions that appear reasonable under aggregate metrics are shown to be costly when phase-specific profiles are applied.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript develops a simulation-based optimization framework framed as a Markov Decision Process for T20 cricket, using a three-phase (Powerplay, Middle, Death) player profile engine estimated with James-Stein shrinkage on 1,161 IPL ball-by-ball records (2008-2025). Win and defend probabilities are computed via vectorized Monte Carlo simulation (N=50,000 trajectories), with batting orders optimized by exhaustive enumeration of feasible arrangements and bowling plans via guided search respecting constraints such as no consecutive overs. Applied to two 2026 IPL matches, it reports that the optimized batting order raises Mumbai Indians' win probability from 52.4% to 56.5% and the optimized bowling plan raises Gujarat Titans' defend probability from 39.1% to 44.3%, attributing sub-optimality to phase-agnostic decisions.

Significance. If the Monte Carlo trajectories prove well-calibrated to real outcomes, the work would usefully extend sports analytics by showing how phase-specific profiles can quantify the cost of aggregate-metric decisions in cricket. The vectorized simulation and shrinkage for sparse data are established strengths that enable efficient evaluation of discrete plans. The specific numerical claims, however, rest on an unvalidated simulator, limiting immediate applicability.

major comments (3)
  1. [§5] §5 (application to the two 2026 IPL matches): The headline improvements of 4.1pp and 5.2pp are obtained by comparing the actual plan against the MDP-optimized plan entirely inside the same Monte Carlo engine; no hold-out matches, comparison to realized match outcomes, or calibration metrics (e.g., Brier score on past games) are reported. This directly undermines the claim that the differences represent real sub-optimality rather than simulator mismatch.
  2. [§3.1] §3.1 (three-phase player profile engine): The model assumes conditional independence of balls within each phase and fixed opponent profiles with no adaptation, pitch degradation, or unmodeled state variables. Because the optimization and evaluation occur inside this simulator, any systematic bias in these assumptions scales the reported gains; a concrete test (e.g., sensitivity to phase-transition probabilities or opponent shrinkage) is needed to bound the robustness of the 4-5pp figures.
  3. [§3.2] §3.2 (James-Stein shrinkage): The shrinkage intensity is a free parameter whose value is not stated, nor is any sensitivity analysis shown on how different intensities alter the phase profiles or the resulting optimal orders/plans. This parameter directly affects the input distributions used for all Monte Carlo trajectories and therefore the claimed improvements.
minor comments (2)
  1. The abstract refers to a 'unified Markov Decision Process framework' yet solves the problems via enumeration and guided search rather than dynamic programming or policy iteration; a brief clarification of the approximation would improve precision.
  2. No mention is made of Monte Carlo standard errors or convergence diagnostics for the N=50,000 trajectories; adding these would strengthen the reliability of the probability estimates.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below and describe the revisions we will make to improve the manuscript.

read point-by-point responses

  1. Referee: [§5] §5 (application to the two 2026 IPL matches): The headline improvements of 4.1pp and 5.2pp are obtained by comparing the actual plan against the MDP-optimized plan entirely inside the same Monte Carlo engine; no hold-out matches, comparison to realized match outcomes, or calibration metrics (e.g., Brier score on past games) are reported. This directly undermines the claim that the differences represent real sub-optimality rather than simulator mismatch.

    Authors: We agree that the reported gains are internal to the simulator and that external validation would strengthen the claims. The two 2026 matches function as prospective case studies using publicly available team line-ups and conditions. Because these matches have not yet occurred, direct comparison to realized outcomes is not feasible. In the revision we will add a dedicated validation subsection that evaluates the simulator on a hold-out set of historical IPL matches (2023–2025), reporting Brier scores and reliability diagrams for simulated win/defend probabilities. This will provide evidence of calibration before presenting the case-study results. revision: yes

  2. Referee: [§3.1] §3.1 (three-phase player profile engine): The model assumes conditional independence of balls within each phase and fixed opponent profiles with no adaptation, pitch degradation, or unmodeled state variables. Because the optimization and evaluation occur inside this simulator, any systematic bias in these assumptions scales the reported gains; a concrete test (e.g., sensitivity to phase-transition probabilities or opponent shrinkage) is needed to bound the robustness of the 4-5pp figures.

    Authors: The assumptions of conditional independence and static opponent profiles are deliberate simplifications required for tractable Monte Carlo evaluation. We acknowledge that systematic bias in these assumptions could affect the magnitude of the reported improvements. In the revised manuscript we will add sensitivity experiments that vary phase-transition probabilities by ±15 % and opponent shrinkage factors across a plausible range, then report the resulting spread in optimal plans and win/defend probability gains. These bounds will quantify the robustness of the 4–5 pp figures under the stated modeling choices. revision: yes

  3. Referee: [§3.2] §3.2 (James-Stein shrinkage): The shrinkage intensity is a free parameter whose value is not stated, nor is any sensitivity analysis shown on how different intensities alter the phase profiles or the resulting optimal orders/plans. This parameter directly affects the input distributions used for all Monte Carlo trajectories and therefore the claimed improvements.

    Authors: We thank the referee for highlighting this omission. The shrinkage intensity was selected by cross-validation on the training data and fixed at 0.45. In the revision we will state this value explicitly in §3.2 and include a sensitivity table showing how intensities ranging from 0.2 to 0.8 affect the estimated phase profiles and, consequently, the optimized batting orders and bowling plans for the two case studies. This will clarify the dependence of the results on the shrinkage parameter. revision: yes

Circularity Check

0 steps flagged

No circularity: MDP optimization and MC simulation derive improvements from data-estimated profiles without self-definition or reduction to inputs

full rationale

The paper estimates three-phase player profiles (Powerplay, Middle, Death) with James-Stein shrinkage from 1,161 IPL ball-by-ball records (2008-2025), then uses vectorised Monte Carlo simulation (N=50,000 trajectories) to compute win/defend probabilities for candidate batting orders and bowling plans. Optimization proceeds by exhaustive or guided search over discrete feasible arrangements (respecting constraints such as no consecutive overs), selecting the argmax of the simulated probability. The reported 4.1pp and 5.2pp gains are simply the differences between simulated probabilities of the actual versus optimized configurations inside this forward model. No equation defines a quantity in terms of itself, no fitted parameter is relabeled as a prediction, and no load-bearing step relies on self-citation or an imported uniqueness theorem; the derivation chain remains self-contained as a simulation-based search procedure.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The central claim rests on estimated player performance parameters from historical data and standard assumptions about Markovian game states; no new physical entities are postulated.

free parameters (1)
  • James-Stein shrinkage intensity
    Controls blending of individual phase stats toward league average; value is implicit in the estimation but not numerically reported.
axioms (2)
  • domain assumption Cricket innings can be modeled as a Markov Decision Process where future state depends only on current overs, wickets, and target.
    Invoked to justify the MDP framework for sequential decisions.
  • domain assumption Monte Carlo sampling of 50,000 trajectories sufficiently approximates true win probabilities.
    Used to evaluate orders and plans.
invented entities (1)
  • Three-phase player profile engine no independent evidence
    purpose: Captures distinct performance in Powerplay, Middle, and Death overs.
    New structuring of player data to enable phase-aware optimization.

pith-pipeline@v0.9.0 · 5558 in / 1471 out tokens · 34710 ms · 2026-05-10T13:42:05.005278+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

  1. [1]

    Bellman, R. (1957). Dynamic Programming. Princeton University Press, Princeton, NJ. Duckworth, F. C. and Lewis, A. J. (1998). A fair method for resetting the target in interrupted one-day cricket matches. Journal of the Operational Research Society , 49(3), 220–227. Ganesh, T. V. (2016). yorkr: An R package for analytics of cricket. R package version 0.0....

    work page 1957

  2. [2]

    and Stein, C

    James, W. and Stein, C. (1961). Estimation with quadratic loss. In Proceedings of the 4th Berkeley Symposium on Mathematical Statistics and Probability , Vol. 1, pp. 361–379. University of California Press. Kirkpatrick, S., Gelatt, C. D., and Vecchi, M. P. (1983). Optimization by simulated annealing. Science, 220(4598), 671–680. Puterman, M. L. (1994). Ma...

    work page 1961