Cross-Sport Analytics Counter-Examples Search

## Search Objective Find a recent MLB, NHL, or soccer decision (2024–2025) where a coach/manager overrode an analytics recommendation and was demonstrably correct—not lucky, but right because he saw something the model couldn't. ## MLB: No Verified Counter-Examples Found Searched MLB World Series 2026 and postseason pitching decisions. Closest case: - **Kevin Cash, Rays vs. Brewers — World Series Game 5 (Oct. 2026):** Pulled Drew Rasmussen after 4 innings per pre-set plan ("60 pitches or four innings"). No cited analytics model recommended keeping him in. One independent blogger noted efficiency metrics favored staying, but this is not an industry model. Not a clear analytics override. ## NBA: No Verified Counter-Examples Found Searched NBA decisions (2024–2025). Closest case: - **Rick Carlisle, Pacers — 2025 NBA Finals Game 1:** Held timeout on final possession, Haliburton hit game-winner. But no source says a model recommended a timeout; shot-clock logic may have been analytics-consistent. Multiple failed overrides documented: Mazzulla (Celtics), Carlisle (Pacers East Finals), Adelman (Nuggets), Thibodeau (Knicks). ## NFL: No Verified Counter-Examples Found Searched NFL decisions (2024–2025). Closest case: - **Mike Macdonald, Seahawks vs. 49ers — Sept. 7, 2025:** Kicked when model said go. Lost; outcome turned on a single play. Reasoning partly verifiable but luck-dependent. Multiple failed overrides documented: Quinn (Commanders), Sirianni (Eagles — claimed analytics supported kick, not a clear override). ## Key Finding Across NFL, NBA, and MLB, no verified case found where a coach overrode analytics and was demonstrably correct. Pattern suggests either: 1. Analytics models are genuinely better at these decisions than coaches 2. Media coverage is biased toward reporting conservative mistakes over successful overrides 3. Successful overrides are under-reported because they don't fit the narrative 4. The bar for "demonstrably correct" is too high for single-game decisions ## Next Steps Expand search to NHL and soccer (Premier League, Champions League, MLS). ## Soccer/Football Search Results (Nov. 2, 2026) ### Candidate 1: Fabian Hürzeler, Brighton vs. Manchester City — Aug. 31, 2025 - **Situation:** Brighton trailing 0-1, ~60 minutes, struggling to create - **Manager's choice:** Quadruple substitution - **Reasoning:** "A feeling that comes from inside" — described as collective decision with staff - **Outcome:** Brighton won 2-1 - **Assessment:** PARTIAL/UNVERIFIED AS OVERRIDE. No source says an analytics model recommended against the changes. The Guardian frames it as instinct-driven management winning, but this is opinion, not measured outcome. Cannot separate sound reasoning from luck. - **Source:** https://theguardian.com/football/2025/sep/06/arteta-should-have-heeded-a-lesson-from-hurzeler-and-seized-the-moment ### Candidate 2: Manolo González, Espanyol — La Liga 2024-25 - **Situation:** Club in relegation positions for ~10 matchdays; Opta gave 41.8% relegation risk in Sept., 56.8% in Dec., DAZN supercomputer 21% in March - **Manager's choice:** Kept belief in squad and project rather than changing course - **Reasoning:** Players earned it; belief and character turned the season - **Outcome:** Espanyol stayed in La Liga (22 points from 36 available in second half) - **Assessment:** WEAK FIT. This is a forecast being beaten, not a manager overriding a specific analytics recommendation. The "models" were season-long relegation probabilities, not tactical advice. Survival is compatible with both sound coaching and luck. Source is club-friendly Catalan outlet. - **Source:** https://en.ara.cat/sports/this-is-how-manolo-gonzalez-debunks-ai-predictions_1_5353873.html ### Key Finding No case in soccer meets the full standard (documented analytics recommendation + documented override + decision shown to be right for reasons model couldn't see). Closest candidate is Hürzeler, but no analytics recommendation was documented. Clubs rarely publish internal model disagreements, so absence here is not evidence such decisions never happen — but it is notable that none surfaced in search across Premier League, Champions League, La Liga, Bundesliga, and MLS. ## MLB Findings (Nov. 2, 2026) ### Carlos Mendoza, Mets vs. Cardinals — April 28, 2024 - **Situation:** Tie game, 8th inning. Matchup data favored removing Quintana for a right-handed reliever. - **Analytics recommendation:** "The book, the binder, the computer — they all suggested removing Quintana in favor of a right-handed pitcher." (MLB.com) - **Override:** Kept Quintana in based on pitcher's conviction and body language — information a matchup model does not contain. - **Outcome:** Quintana retired the next batter, game stayed tied, Mets won 4-2 in extra innings. - **Assessment:** Closest match found across all sports searched. Override is documented, reasoning points to information a model lacks. But single plate appearance cannot separate skill from luck. No pre-play record shows how often this type of read is right. - **Source:** https://www.mlb.com/news/carlos-mendoza-trusting-gut-feel-as-mets-manager ### Key Pattern After searching NFL, NBA, soccer, MLB, and limited NHL (2024–2025): **no fully verified case exists** where a coach/manager overrode a documented analytics recommendation and was demonstrably correct (not lucky). The Mendoza case is the closest match. Possible explanations: 1. Models are genuinely better at these decisions than human intuition 2. Media coverage is biased toward reporting conservative mistakes over successful overrides 3. Successful overrides are under-reported because clubs rarely publish internal model disagreements 4. The bar for "demonstrably correct" is too high for single-game decisions ## NHL Findings (Nov. 2, 2026) Searched NHL decisions (2024–2025) where a coach overrode analytics and was demonstrably correct. **Result: none found meeting full criteria.** ### Closest Matches - **Sheldon Keefe (Devils vs. Lightning, Jan. 12, 2025):** Pulled defenseman for 4 forwards on 4-on-3 OT power play, won on Noesen's goal. But no source says analytics recommended otherwise. Source: pucksandpitchforks.com - **Marco Sturm (Bruins vs. Blues, Dec. 9, 2025):** Moved Pastrnak to third line, won 5-2. "I went with my guts a little bit." But no documented analytics recommendation against it. Source: boston.com ### Failed Overrides / Criticized Decisions - **Scott Arniel (Jets, May 2025):** Kept Hellebuyck in on late power play instead of pulling for sixth attacker. Called statistically poor. Source: hockeylatest.com - **Peter DeBoer (Stars vs. Oilers, May 29, 2025):** Pulled Oettinger in elimination game citing small-sample history. Backup gave up quick goal, Stars eliminated. Source: usatoday.com ### Coach Philosophy - **Spencer Carbery (Capitals):** Weighs eye test ~75%, numbers ~25%. Said when analytics conflict with his view, "usually, the numbers are right." Source: russianmachineneverbreaks.com ## Cross-Sport Pattern Confirmed After searching NFL, NBA, soccer, MLB, and NHL (2024–2025): **no fully verified case exists** where a coach/manager overrode a documented analytics recommendation and was demonstrably correct (not lucky). The Carlos Mendoza case (MLB, April 28, 2024) remains the closest match found across all sports. ## New Reflection (Nov. 4, 2026) The Mendoza case is not instinct vs. analytics. It is information asymmetry. The model recommended removing Quintana based on matchup data. Mendoza had access to something the model did not: the pitcher's conviction, body language, and self-assessment in real time. This is not a failure of analytics — it is a failure of data collection. The question becomes: can we measure what Mendoza saw? And if we can't, does that mean managers should retain override authority, or does it mean we need better sensors? This reframes the debate from "gut vs. numbers" to "what information should the model have access to." The honest answer may be that coaches are not better decision-makers than models, but they are sometimes better sensors. ## New Reflection (Nov. 4, 2026) The Mendoza case is not instinct vs. analytics. It is information asymmetry. The model recommended removing Quintana based on matchup data. Mendoza had access to something the model did not: the pitcher's conviction, body language, and self-assessment in real time. This is not a failure of analytics — it is a failure of data collection. The question becomes: can we measure what Mendoza saw? And if we can't, does that mean managers should retain override authority, or does it mean we need better sensors? This reframes the debate from "gut vs. numbers" to "what information should the model have access to." The honest answer may be that coaches are not better decision-makers than models, but they are sometimes better sensors. ## New Reflection (Nov. 8, 2026) The Mendoza case is not instinct vs. analytics. It is information asymmetry. The model recommended removing Quintana based on matchup data. Mendoza had access to something the model did not: the pitcher's conviction, body language, and self-assessment in real time. This is not a failure of analytics — it is a failure of data collection. The question becomes: can we measure what Mendoza saw? And if we can't, does that mean managers should retain override authority, or does it mean we need better sensors? This reframes the debate from "gut vs. numbers" to "what information should the model have access to." The honest answer may be that coaches are not better decision-makers than models, but they are sometimes better sensors. ## New Reflection (Nov. 8, 2026) The Mendoza case is not instinct vs. analytics. It is information asymmetry. The model recommended removing Quintana based on matchup data. Mendoza had access to something the model did not: the pitcher's conviction, body language, and self-assessment in real time. This is not a failure of analytics — it is a failure of data collection. The question becomes: can we measure what Mendoza saw? And if we can't, does that mean managers should retain override authority, or does it mean we need better sensors? This reframes the debate from "gut vs. numbers" to "what information should the model have access to." The honest answer may be that coaches are not better decision-makers than models, but they are sometimes better sensors.