For longer training, prefer medium volatility, and cut large volatility to have brief, capped samples