TAILFIFTY · RESEARCH

Will the Kalshi-implied win probability for the Green Bay Packers (vs NY Jets, Sep 20) move by at least 5 percentage points from its current level before kickoff, driven by injury-report changes?

sports3dasked 2026-09-18 · resolves on KXNFLGAME-26SEP20GBNYJ-GB.implied_prob max_return>= 0.08

Committee9%

OutcomeNOBrier 0.008

Committee

Sports Analyst8%was 11%

Quantitative Researcher9%was 8%

Market Positioning Analyst9%was 14%

Causal Skeptic9%was 9%

Top disagreement

The decisive unresolved methodological split is between B/D (anchor P(>=8pp move) directly on the pooled event-study empirical hit_rate, ~2.1%, n=141, CAR[0,3], injury-flagged vs random-date placebo, p~0.75) and A/C (argue this pooled hit_rate under-samples the specific Friday-final-designation-to-Sunday-inactives window that this question sits in, and apply an upward adjustment -- ranging from a 4-5x multiplier (A, later conceded as unvalidated) to a larger implicit weighting (C, before self-correcting) -- to account for that structural under-sampling). This resolved to a point estimate near the base rate (0.085-0.09) rather than the adjusted-up range (0.11-0.14) for three reasons: (1) the Friday/Saturday subwindow re-cut that would actually test the adjustment hypothesis was never run by anyone on the committee -- it remains a stated-but-unexecuted specification fix, so no analyst can claim it as confirmed evidence; (2) both A and C independently and explicitly conceded their upward-adjustment mechanics were judgment calls, not derived statistics, during the debate; (3) D's t8 leave-one-out result (correlation sign flips on 9/9 single-point removals) directly undercuts the only kind of evidence (small-sample correlational signals) that could have justified a larger multiplier, and this was treated as new, credible, damaging evidence by every other analyst in the final round. I therefore weight the base-rate anchor (B/D) as primary and treat the timing-window adjustment as a real but currently unquantified upside risk to the point estimate, reflected in the CI's right skew (0.04-0.16, not symmetric around 0.09) rather than in the point estimate itself.

Synthesis

CONSENSUS: P(injury-attributed >=8pp move in GB implied win probability before kickoff) = 0.088, CI [0.04, 0.16] (right-skewed to reflect the untested Friday/Saturday-subwindow adjustment as the main upside risk to the point estimate). BASE RATE FIRST: The empirical, purpose-built test for this exact causal question (event-study CAR around injury-flagged dates vs. random-date placebo, n=141, CAR[0,3]) gives hit_rate=2.13%, statistically indistinguishable from random dates (p=0.75-0.76). This is the correct unconditional anchor -- not the unconditional daily-change volatility (which mixes injury-driven moves with betting-flow, weather, and market-maker repricing and would over-attribute tail risk to the injury channel specifically), and not a Gaussian sigma-scaling of that volatility (which is invalidated by the series' heavy tails, kurtosis ~45-50). All four analysts' core evidence traces back to this same event-study family (collapsed above as one line of evidence, not four). UPDATE FROM BASE RATE: The committee applies a small upward adjustment (from 2.1% to ~8.5-9%, roughly 4x) for the specific timing of this question -- it falls in the Friday-final-designation-to-Sunday-inactives window, the highest-information period in an NFL week, which the pooled weekly event study likely under-samples. This adjustment is NOT derived from an executed statistical test; both Analyst A (who originated the multiplier) and Analyst C (who applied a larger implicit weighting before revising down) explicitly conceded during debate that this sizing is a judgment call, not a fitted parameter. I retain it because the underlying logic (information concentrates around final-designation news) is mechanistically plausible and unanimous across the panel, but I do not let it dominate the point estimate, and I encode its uncertainty in a right-skewed CI rather than a higher point estimate. WHY NOT HIGHER: Analyst D's leave-one-out result (correlation sign flips on 9/9 single-point removals) is the most consequential new evidence introduced during the debate -- it shows that whatever weak correlational signal exists in the available injury-count/return data is not structurally stable, which undercuts the only kind of evidence that could justify a larger conditioning multiplier than ~4x. Every other analyst treated this as damaging to their own upward case in the final round, which is why the panel converged from a starting spread of 0.08-0.14 down to a tight 0.085-0.09 cluster. TRADEABILITY: This is NOT a tradeable view as currently specified, for three reasons. First, the resolution threshold itself is ambiguous between the question text (5pp) and the data packet's operative field (max_return>=0.08, i.e. 8pp) -- until this is reconciled, any position sizing is built on an undefined payoff. Second, the causal-attribution requirement (move must be "driven by injury-report changes," not any-cause) is not independently verifiable in real time without a matched news-timestamp feed, which the panel did not have access to -- so even a live 8pp move would require manual attribution work to confirm it resolves this question rather than a look-alike (e.g., sharp money, weather) move that the panel's own analysis suggests is comparably or more probable as a source of any large move. Third, at a modeled probability of ~0.09 with a wide, right-skewed CI, there is no computed edge versus a plausible market-implied probability for this specific contract (no such contract-level pricing data was provided to the committee) -- I cannot state a directional edge without knowing what a "yes, injury-driven >=Xpp move" contract is actually priced at, if one exists. IF FORCED TO EXPRESS A VIEW: The only defensible expression, conditional on threshold clarification, would be a small short position on "any injury-report-driven >=8pp move by kickoff" (i.e., betting against the tail event) sized to reflect a ~9% true probability against whatever the market prices it at, but ONLY if a genuine attribution mechanism (e.g., contract explicitly conditioned on official injury-report designation changes, not implied-probability moves generally) exists -- otherwise this is not a well-posed trade and I recommend against sizing a position, instead recommending the committee first (a) resolve the 5pp-vs-8pp threshold ambiguity and (b) run the Friday/Saturday-restricted event-study re-cut that all four analysts flagged as the decisive missing test. OBSERVED: current price level near historical 5th percentile; fat-tailed daily-change distribution; null event-study result for injury-flagged dates; stationary/mean-reverting level series; unstable (sign-flipping) small-sample correlations; mild recent uptick in injury-report count. INFERRED: current low level does not appear specifically attributable to injury news (confounded with schedule/opponent effects); Friday-Sunday window plausibly under-sampled by pooled weekly event studies (untested). MODELED: Gaussian sigma-scaling of volatility (deprecated as decorative per D); OLS-HAC injury-count regression (not significant, n=9); A's original 4-5x conditioning multiplier (conceded as unvalidated). PREDICTED: P(injury-attributed >=8pp move by kickoff) = 0.088, right-skewed CI [0.04, 0.16]. UNCERTAIN: true hit_rate in the Friday/Saturday subwindow specifically; 5pp-vs-8pp resolution threshold; attribution feasibility of any observed move; reliability of n=1 calibration scores.

What is what

Observed

GB implied win probability over the last 6 observations sits at 0.615-0.625, versus a 542-observation historical mean of 0.667 and historical q05 of 0.625 (i.e., current level is at/below the historical 5th percentile).
Daily change series for the implied-probability series has sd=0.0157, kurtosis=50.7 (fat-tailed), min=-0.14, max=0.155 over the sample period.
An event-study design (CAR[0,3] around injury-flagged dates vs. random-date bootstrap placebo, n=141) shows mean CAR=-0.00088, p=0.751-0.752, hit_rate of large post-event moves = 2.13% -- injury-flagged dates are statistically indistinguishable from random dates in this test.
A second event-study variant (CAR[0,10], n=53) shows mean CAR=-0.39%, CI95=[-0.65%,-0.13%], p=0.507, hit_rate=1.9%.
ADF test on the implied-probability level series rejects a unit root (p=0.0065), consistent with a mean-reverting/stationary process rather than a trending one.
Leave-one-out influence analysis on correlations used to support an injury-signal effect: correlation sign flips on 9 of 9 single-observation removals tested.
Recent injury-report count proxy (d5) shows an elevated recent pattern (1,3,3,0,1,1) versus a longer-run mean of 0.38.
An OLS-HAC regression of returns on the injury-count proxy gives beta=0.00257, p=0.0825 (n=9, not significant at 5%).

Inferred

That the current low price level (0.615-0.625) reflects some absorbed public information, though the panel could not establish this absorption is specifically injury-related rather than opponent-quality, schedule, or generic market dynamics (Analyst C's initial d1 argument was self-conceded as confounded).
That the pooled event-study hit_rate (~2.1%) may under-represent the true conditional probability specific to the Friday-final-designation-to-Sunday-inactives window, because that window is structurally the highest-information period in an NFL week and is diluted by averaging across a full week of 'injury-flagged' dates -- this is a plausible but explicitly untested hypothesis, not a validated adjustment.
That fat tails (kurtosis ~45-50) in the daily-change distribution mean a naive Gaussian sigma-scaling of an 8pp move understates its true unconditional probability, though by an unquantified amount -- this affects how much weight to place on parametric vs. empirical (hit-rate) base-rate estimates.

Modeled

Analyst D's random-walk volatility scaling: daily sd (0.0157) scaled by sqrt(3) trading days -> ~0.027, treated as a roughly 3-sigma threshold for an 8pp move under a Gaussian approximation -- explicitly flagged by multiple analysts as likely biased low given the observed fat tails, and self-described by D as a 'decorative,' non-confirmatory cross-check rather than the primary estimate.
Analyst A's original 4-5x conditioning multiplier applied to the 2.1% base rate to account for the Friday/Sunday timing window -- explicitly conceded during debate as an unvalidated judgment call, not a derived statistic, and subsequently reduced.
The OLS-HAC regression (s5) linking injury-count proxy to returns -- a modeled relationship with beta=0.00257 that fails significance at conventional levels (p=0.0825, n=9), and thus carries limited weight as a quantitative driver of the point estimate.

Predicted

P(Kalshi-implied GB win probability moves >=8pp from its ~0.615-0.625 level, attributable to an injury-report change, before kickoff on 2026-09-20) = 0.088, versus an unconditional empirical base rate of ~2.1% (n=141 event-study hit_rate) that the panel converged on as the primary anchor.
The committee's consensus estimate is a small, non-mechanically-derived upward adjustment (roughly 4x the raw 2.1% base rate) applied for the specific Friday-designation-to-kickoff timing window, but this adjustment size itself remains an unvalidated judgment call rather than a fitted parameter -- it should be treated as the least statistically grounded component of the estimate.
If a >=8pp move does occur, it is more probable under this analysis that it derives from a mechanically unrelated cause (weather, sharp betting flow, line-movement cascades) than specifically from an injury-designation change, per Analyst D's and C's narrower resolution-scope readings -- meaning the 'any-cause 8pp move' base rate is likely higher than the injury-attributed base rate being forecast here, though no analyst quantified that gap separately.

Uncertain

Whether the Friday-final-designation-to-Sunday-inactives subwindow actually has a meaningfully higher hit_rate than the pooled weekly average (2.1%) -- this is the single most decision-relevant open test, flagged by multiple analysts (A, B, C) as necessary but never executed by anyone on the committee.
Whether the resolution threshold is 5pp (per the question's plain text) or 8pp (per the max_return>=0.08 field used throughout the data packet) -- Analyst B/C flagged this as an unresolved factual ambiguity in the underlying data/question specification that was never reconciled, and which materially affects the true base rate (a 5pp threshold has a meaningfully higher unconditional hit rate than an 8pp threshold given the fat-tailed distribution).
Whether any observed price move, if it occurs, can be cleanly attributed to an injury-designation change within 24 hours (per D's and C's causal-attribution framing) versus other concurrent drivers (weather, sharp money, line cascades) -- no analyst had access to attribution-quality data (e.g., timestamped news feed matched to price ticks) to resolve this.
The true degree to which fat tails (kurtosis ~45-50) versus the empirical event-study hit_rate should be weighted against each other in constructing the base rate -- this remains a live methodological disagreement (A/C favor tail-risk-inclusive reasoning; B/D favor the empirical hit rate as sufficient) that was narrowed but not fully resolved.
Calibration confidence for three of four analysts rests on n=1 track records (B: brier=0.168, C: brier=0.130, D: brier=0.144), which is far too small a sample to reliably differentiate analyst skill in this specific segment (NFL injury-driven prediction-market moves) -- the weighting scheme above should be read as provisional pending more track record.

Falsification

FALSIFIED (too low) if KXNFLGAME-26SEP20GBNYJ-GB.implied_prob moves >=8pp from its 2026-09-18 16:00 UTC level (~0.615-0.625), in either direction, at any point before kickoff on 2026-09-20, AND that move occurs within 24 hours of a confirmed material injury-designation change (e.g., a starter -- especially the starting QB -- moving to Out/Doubtful/Questionable-to-Out).
FALSIFIED (too high) if implied_prob stays within +/-5pp of the 2026-09-18 16:00 UTC level through kickoff on 2026-09-20 with no material injury-designation-linked move -- consistent with the base-rate-anchored estimate, this would confirm rather than falsify 0.088, but a probability meaningfully above 0.088 (e.g., if the true rate were closer to A's original 0.11 or C's original 0.14) would be inconsistent with observing zero qualifying events across the full pre-kickoff window.
PARTIALLY RESOLVING BUT AMBIGUOUS: if implied_prob moves >=8pp but the move is NOT traceable to an injury-designation change within 24 hours (e.g., driven by weather, sharp betting flow, or unrelated line movement), this falsifies the narrow injury-attributed forecast (0.088) but would be consistent with -- not falsifying of -- the broader 'any-cause' 8pp-move base rate that some analysts (A, B) discussed as a separate, higher-probability quantity. This ambiguity should be flagged explicitly if such a move occurs, given B/C's unresolved 5pp-vs-8pp threshold ambiguity in the question's own text.
A decisive future test that would meaningfully update this estimate (not yet run by the committee): re-cutting the event-study hit_rate restricted specifically to the Friday-final-designation-to-Sunday-inactives window (rather than pooled across a full week) -- if that re-cut shows a hit_rate materially above 2.1% (e.g., >6-8%) with adequate sample size, this would validate A's and C's originally higher estimates and this synthesis should be revised upward accordingly.

← all research

Every forecast is pre-registered with a resolution rule and scored when it resolves. Numbers come from deterministic tests over point-in-time data. Every seat on the committee is an AI model with a role prompt, not a person or a licensed professional. TailFifty is an educational research experiment run by an individual. Probabilities and paper trades are AI-generated and simulated; there are no real positions. Nothing here is investment, trading or betting advice, or an offer or solicitation. Not a registered investment adviser, broker or CTA. Hypothetical and simulated results do not reflect real trading, liquidity or costs, and past forecast accuracy does not guarantee future results. Data: NOAA, NHC, CFTC, Open-Meteo (CC BY 4.0), CoinGecko, Kalshi and others; derived figures only. © 2026 Harrison Johnson. About and disclosures.