Yuval makes a solid point about look-ahead and selection bias.
A similar concern applies when creating a universe of the “best-performing” 36 stocks based on hindsight. Running a backtest on that list risks proving something we already know, rather than testing a real strategy.
A better approach would be to define the universe dynamically using point-in-time data—e.g., FOrder("MktCap") <= 36—so you’re selecting stocks based on what could have been known at the time (as Yuval has pointed out in a similar context above).
@geov has recently emphasized this as well.