Skip to content

SupeRR Selector · Season 2 · Rajasthan Royals × Unstop

Backing players under pressure

Out-of-sample checks · Data as of 18 Nov 2026

Does it work? Modestly

Thirty splits: rate players on one half, rank-check on the half the ratings never saw. The honest answer: modestly. Past clutch is a weak signal of future clutch. We keep the pressure lens for two reasons and no more: the contest question asks for players under pressure, so the metric matches the brief; and the weighting never claims predictive superiority, only relevance. Most entries to this problem will quietly assume clutch is a stable skill. It is not, and we say so.

1

Half-split validation, final configuration

Pressure weighting matches plain stats (rank correlation 0.28 vs 0.26). Plain unweighted stats predict the held-out half better (0.20 vs 0.13). A pure win-probability-added ranking was clearly worst and was dropped.

Figure 3. Held-out rank correlation by group and weighting, from validation_halfsplits_v2.csv. Thirty splits; players rated on one half, ranked against the half the ratings never saw. Hover for exact values.
Table 1. Half-split results, final configuration (v2 export)
GroupMethodRank correlationHeld-out top-10 value
BattersUniform (plain stats)0.200.0362
BattersPS-weighted0.130.0216
BowlersUniform (plain stats)0.260.0462
BowlersPS-weighted0.280.0596

Earlier method comparison

Table 2. The earlier four-method run (validation_halfsplits.csv). The held-out top-10 value in this export is on a different scale from Table 1; read within this table only.
GroupMethodRank correlationHeld-out top-10 value
Battersraw(uniform w=1)0.1160.883
BattersWPA only0.0010.830
BattersPS-weighted (core)0.0600.883
Batterscombined PIS-0.0180.814
Bowlersraw(uniform w=1)0.290-0.621
BowlersWPA only-0.224-0.861
BowlersPS-weighted (core)0.283-0.590
Bowlerscombined PIS0.225-0.619
2

Bootstrap stability

Resampling matches 200 times and rebuilding each top-10 shows only the very top of a list is stable. Drawn only for auction-pool players who appear in the stability tables; names absent from those tables have no share printed, and we do not impute one.

Batters · share of 200 resamples in the top 10

  • Reeza Hendricks44%
  • Pathum Nissanka40%
  • Kamindu Mendis22%
  • David Warner6%

Bowlers · share of 200 resamples in the top 10

  • Sufiyan Muqeem84%
  • Mohammad Nawaz70%
  • Shoriful Islam66%
  • Lockie Ferguson41%
  • Mark Adair39%
  • Abrar Ahmed33%
  • Tim Southee32%
  • Taskin Ahmed31%
  • Sikandar Raza30%
  • Matheesha Pathirana19%
  • Reece Topley19%

Pakistani players Muqeem (84%) and Nawaz (70%) top the bowlers' stability table but are excluded from consideration; Shoriful Islam (66%, rounded from 0.655) is the highest eligible bowler.

3

Limitations

Samples are small where they matter most. Only 6.6% of chase balls are high-pressure, so a full season's finisher might carry 80 to 120 of them; shrinkage handles the arithmetic but cannot manufacture evidence. Three of our six signings rest on fewer than 130 pressure balls.

The pack mixes IPL and T20I cricket. For Pakistan and Bangladesh players the mix is T20I-heavy, and sensitivity tests show those bowlers fall most when IPL-only data is used; for Australians and Indians the lists barely move. The over-level normalisation fixes early- and late-over inflation but imposes a uniform mean across overs, which is an assumption, not a discovery.

Availability is not confirmed for anyone: the registered pool reaches franchises only on 3 December 2026, and the five exclusions for injury, retirement, withdrawal and legal risk rest on public reporting, not the data pack.

Finally, the pack contains no Indian domestic cricket, so Syed Mushtaq Ali form is invisible to the model; the Indian watchlist behind our relaxed domestic threshold is correspondingly thin, and an uncapped domestic spinner who is cheap and in form will not appear in any number we printed.