The symptom is a large gap between in-sample and out-of-sample performance.
Why finance is especially vulnerable. The signal-to-noise ratio in returns is very low, so a flexible model has abundant noise to fit. This is why linear models remain competitive in quant research long after being abandoned elsewhere.
The subtle form. Even with a holdout set, repeatedly checking it and adjusting turns it into training data. Overfitting through researcher decisions is harder to see and more common than overfitting through parameters.
Fixes: regularisation, fewer features, walk-forward validation, and honesty about how many variants were tried.