Known limitations
WeBall3 min readUpdated Aug 15, 2026Reviewed Aug 15, 2026
This page exists because the alternative is worse. A model presented without its failure modes invites people to trust it in exactly the situations where it should not be trusted.
Thin samples early in a season
The engine builds probabilities from recent results. In the opening weeks of any season there are barely any, so model output is shrunk heavily toward the market and many bets return THIN DATA rather than a grade.
This is the system behaving correctly, but it means the product is least useful precisely when interest is highest. College football in August and the NFL in September are the clearest cases.
Status: inherent. More history does not exist until games are played.
De-vig bias on lopsided markets
Fair prices come from multiplicative de-vig — normalising implied probabilities so they sum to 100%. That assumes books spread margin evenly across outcomes, and they do not: more margin is loaded onto long shots.
The consequence is a systematic bias on heavily lopsided markets, where the method slightly understates the favourite and overstates the underdog. On a −110/−110 spread the effect is negligible; on a −2000 moneyline or a 30-team futures field it is not.
Status: fixable. Power and Shin de-vig methods correct for it.
Garbage time and coaching decisions
The engine models scoring, not intent. It has no variable for how early a coach pulls starters, how aggressively a trailing team plays in the fourth quarter, or whether a game is meaningful for either side.
This matters most on large spreads, especially in college football, where blowout dynamics drive a real share of outcomes. Treat grades on very large spreads as a starting point rather than a finished opinion.
Status: inherent to a score-based model.
Correlation between your own bets
Grades are produced per bet, independently. The engine has no view of your overall exposure, so it cannot tell you that six of today's recommendations are effectively the same bet on the same weather system or the same market condition.
Kelly sizing assumes independence. If you stake several correlated bets at quarter Kelly each, your true exposure is considerably higher than it looks.
Status: requires bet tracking, which does not exist yet. On the roadmap.
Markets the engine cannot ground
Anything whose deciding number cannot be reconstructed from stored history is UNRATED — shown with a real price and no opinion. This includes most exotics, some period and combination markets, and any player whose identity could not be resolved to history with enough certainty to risk a wrong-player grade.
Resolution failures are logged and reviewed rather than silently discarded, so this category shrinks over time. See markets covered for the current list.
Status: ongoing.
No live or in-play grading
Everything is pregame. Prices move once a game starts and the engine does not follow them, so nothing shown during a game should be treated as a live opinion.
Status: deliberate scope decision.
The calibration ledger is not yet populated
The honest test of any probability model is calibration: of the bets graded at 55%, did roughly 55% win? That requires a meaningful number of settled graded bets, and the record is still accumulating.
Until it is published, no track-record claims are made anywhere in the product. A model that will not show you its misses has not earned the confidence it is asking for.
Status: in progress. The computation exists; the sample does not yet.
Common questions
Why publish this at all?
Because a confidence number you cannot audit is decoration. A bettor who knows where a model is weak can use it properly; one who has been told it works everywhere cannot. It is also the difference between analysis and touting.
Are these being fixed?
Some are on the roadmap and some are inherent to the approach. The list marks which is which, and it gets revised as the engine changes rather than left to go stale.