Overfitting
How can a trading rule that looked perfect on past data fail as soon as the market writes new data?
In plain words
Overfitting happens when a rule is tuned so precisely to the quirks of past data that it captures noise instead of a real pattern, and then fails on new data it has never seen.
Why it matters
A rule with many settings can always be adjusted to look good on a sample it was fitted to. When the next period arrives, those quirks vanish, and risk that seemed measured turns out to have been assumed rather than earned.
In a real week: ETH, week of 20 September 2021
ETH's history in the data covered 188 weeks. In the week of 20 September 2021 it fell 8.09% with a 26.24% range, an 11.21% best day on 2021-09-22 and a -10.55% worst day on 2021-09-20, closing in the middle third of that range. A rule tuned to smoother weeks would have been tested by swings this size.
A common mistake
Assuming performance on the fitting sample says anything about future performance. The fit quality rises automatically as settings are added, even when the added settings follow no repeatable market behaviour.