Stock prices screw up validation with back testing. I tried using stock prices to illustrate random walks and the superiority of naive forecasts with random walk data. Only problem: they aren't random walks. They are random sequences of predictable pattern. Frequently, a regression model, or Holt-Winters, or neural net will outperform the naive forecast on backtests, because it was a better model before now. The challenge, is that now it isn't.