Here are a shop's sales for the last 6 months. Your job is to predict the next 3. Drag the slider to shape your model, from a plain straight line to a curve that bends to hit every month.
Here is the trap. The curve that hit every past month did it by bending to fit the random bumps as well as the real trend. Those bumps are noise, and noise does not repeat, so when the next months arrived your model lurched off chasing a pattern that was never really there. The plain straight line ignored the wiggles and stayed close. Drag the slider back toward simple and watch the future error fall: under real uncertainty, the model that fit the past best is rarely the one that predicts the future best.
The bias-variance trade-off behind less-is-more (Gigerenzer & Goldstein, 1996; Gigerenzer & Brighton, 2009). Sales here are illustrative; the out-of-sample failure is a real consequence of overfitting. This holds when data is scarce and noisy: with plenty of clean data and a genuinely complex pattern, more flexibility can help. The art is matching the model to the world.