Optimization without overfitting in trend-system evaluation

Overfitting in system optimization

Overfitting occurs when parameters are tuned so tightly to historical patterns that the same rules fail on later, different data.

Scanning every possible value and then keeping only the profitable ones is described as a first step toward overfitting.

Use more history, not less

Using more history is preferred because it adds bull and bear regimes, price shocks, and a larger trade count for evaluation.

Commit the range before the scan

Parameter ranges should be chosen in advance from the expected holding style, such as 40 to 120 days for macrotrend work and 5 to 40 days for short-term work.

Trend rules used for the tests

A one-parameter trend test can use the slope of a moving average to stay long when the line is rising and short when it is declining, avoiding many false signals from price-versus-line crossings.

A Moving-average crossover is long when the faster average is above the slower one and short when the faster average is below.

Retest the same range on later windows

Testing the same range over three successive windows, each about half as long as the previous, is used to check whether the edge still appears in later history.

A robust result is defined as a large share of tests remaining profitable, not as a single peak profit, with a trend example of as much as 70% successful tests.

When a new rule is treated as generalized

A new rule is treated as generalized only if the average of all tests improves and most individual tests improve. Improvement confined to one pocket while others worsen is treated as overfitting.

A volatility filter on new entries

A volatility filter that blocks new entries when annualized price volatility exceeds 50% and waits until it falls back below that level is presented as typically cutting return a little while cutting risk much more.