Heteroskedastic errors have to do with the variance of errors fitting the regression model are large. This descriptor is used for regression modeling of all types, from models on portfolios to academic studies.
Imagine a graph with a bunch of data points, randomly placed, but generally skewing in a line from the bottom-left to the upper-right. Now draw a “best fit” line, minimizing the distance between the line and all the points. The farther all the data is from the line, the more heteroskedastic it is. The closer the data points are to the line, the more homoskedastic the variance is.
The more heteroskedastic the variance, the more likely there’s a better model for the data, and the less likely the model is correlating with the data.