In what way does the presence of a 'leaker' compromise the prediction quality?

The presence of a 'leaker' can undermine the integrity of a predictive model by introducing information from the target variable into the feature set in an inappropriate manner. This situation can lead to artificially inflated prediction quality, as the model appears to perform better than it truly would in practice. This is often because the model inadvertently learns this information that it shouldn't have access to during its training phase, resulting in misleading performance metrics when evaluated.

In this context, other answers may not accurately capture the overarching issue related to prediction quality. Adding unnecessary complexity could refer to alterations in the feature set that do not truly reflect the underlying data patterns, but it doesn’t focus on the dilution of model trustworthiness caused specifically by leakers. While limiting training data could hinder the model, the primary concern is the quality of the insights it generates, not the quantity. Additionally, a decrease in speed of model training may happen if the model is more complicated, but this isn't a direct reflection of prediction quality compromised by leakers. Thus, the correct focus is on the detrimental impact on prediction accuracy due to artificially inflated performance metrics caused by the leaker.