Curve fitting trading strategy and trading system curve fitting test searches point to the same failure mode: you adjust a model until it hugs historical prices. The backtest looks brilliant because you optimized on the path you judge.

This is not "being smart about data." It is training a camera to photograph yesterday.

How curve fitting sneaks in

Manual tweaks after peeking at equity curves
Parameter searches without a held-out protocol (Parameter optimization)
Many strategy variants on one dataset (Data snooping)
"Just one more filter" after you saw the drawdown in window 3
reusing the same narrative across different symbols without independent evidence

Degrees of freedom are not only parameters

Anything you change after observing results consumes degrees of freedom:

indicator windows
universe selection rules
start dates and warm-up policies
fee assumptions after you did not like the first answer

If you do not budget these choices explicitly, you will always feel honest while overfitting.

What to do instead

Fix a research protocol before touching parameters (write the success metrics first)
Use walk-forward discipline (What is WFA?)
Treat hyperopt winners as suspects (Hyperopt)
Run cost stress tiers and record whether rankings are stable

The emotional trap

Curve fitting feels like progress because the chart improves immediately. Real discovery is slower: it shows up as robustness under stress, not as a prettier backtest line.

A simple test

Ask: "If I freeze parameters and roll time forward, does the edge still exist?" If the answer requires more tuning, you are still fitting (What is WFA?).

A stronger test (preregistered outcomes)

Before you optimize, write down:

which metrics decide pass/fail
what threshold counts as pass
how many variants you are allowed to try

If you change the metric after you see results, reset the protocol.