Multicollinearity

When two columns carry nearly the same information, many coefficient pairs fit equally well, so estimates swing between resamples and signs can flip. Only their combined effect is identified. Check pairwise correlations or variance inflation factors; remedies are dropping or combining columns, domain-driven aggregation or ridge shrinkage. It matters for interpretation more than for prediction.