Feed a model a 10-day, 12-day and 14-day moving average and it cannot tell which one is doing the work. Coefficients swing wildly between samples, sometimes flipping sign, while overall fit barely changes.
This matters most when you interpret coefficients or when you use them to size positions. It matters less for pure prediction, though it still inflates variance and makes the model fragile out of sample.
Remedies: drop redundant features, combine them into one composite, or use regularisation, which shrinks correlated coefficients toward each other instead of letting them fight. Checking pairwise correlation between features before modelling costs nothing and prevents most of it.
Related: ols-regression, regularisation, feature-engineering, degrees-of-freedom