Probability is one output, not the full decision
A model can estimate that a market is likely while other evidence says the estimate should not be promoted. Recommendation logic exists to combine probability with input quality, market conditions and risk patterns.
This is why non-recommended matches remain visible in BetData's general table. Their percentages are useful analytical information, but their status remains separate and explicit.
Reason 1: the available price may offer too little room
A high probability often comes with low odds. If the market price already reflects the estimated chance, the difference between model and market may be too small. A likely event and a well-priced selection are not the same thing.
Price floors and edge checks reduce selections where a small estimation error could remove the apparent advantage.
Reason 2: data coverage may be weak
A precise percentage can hide an imprecise foundation. Missing matches, inconsistent team mappings, a new season or too little relevant home and away history can weaken the estimate. Good filtering should respond to the quality of the sample, not just the number produced by the model.
Reason 3: match patterns can conflict
Different indicators may point in opposite directions. A broad scoring average can look positive while recent home or away behaviour, low-event patterns or another market signal raises caution. Cross-market consistency checks help prevent one strong feature from dominating the full decision.
Reason 4: recommendation thresholds are intentionally narrower
A recommendation is designed to be a filtered subset, so it should be normal to have days with few or no green cards. Filling a page with recommendations for the sake of activity would change the meaning of the label and weaken transparency.
BetData can also visually highlight an analytical pattern without changing its status to recommended. A label such as “SHADOW · NOT A RECOMMENDATION” is descriptive; it does not enter the recommendation count or become a betting instruction.
How to judge the filter
Evaluate both sides: the results of selected recommendations and the matches the filter rejected. Over a sufficiently large, time-ordered sample, this shows whether the extra rules improve selection quality rather than merely reducing volume.
No filter removes normal football variance. A recommended prediction can lose and a rejected forecast can win. The purpose of filtering is consistent selection under predefined rules, not perfect foresight.