A favourable bet selection is ideal, but in my opinion there is no such animal when you're talking about random outcomes. This is because, by definition, random outcomes are unpredictable, so there can be no factors which act as "predictors" for the next bet or series of bets.

As you learn in statistics 101 courses, "correlation is not causation". Correlation means that one factor or variable "goes with" another. There are many correlations in roulette; for example the number of hits on the third column tend to be correlated with the number of hits on red (more so than the other columns, because there are more reds in it), but these kinds of correlations don't give you a favourable bet selection. There are different kinds of correlations though, and if it were possible for there to be some kind of serial correlation, "serial" meaning that some past outcomes were correlated with future outcomes, then there would be some hope. However, to discover such correlations you would have to look at physical variables (dealer or ball changes, etc).

Why? because in that case you're not relying merely on correlation, but also causation.

I don't think that progressions necessarily mean chasing losses; you may as well say that they mean "chasing wins". Nor do they have to rely on the gambler's fallacy. You may not be able to get more wins than losses (relative to the payout on a bet), but this is what a well designed progression can make up for. You may not be able to predict what's coming next with any more certainty than random, but the statistical behaviour of a bet can nevertheless be an aid in designing your wagering plan. For example, for an even chance bet 87.5% of outcomes consist of streaks no longer than 3.