The researchers used a computer model to study forecast errors in weather predictions. They found that when comparing forecast errors to a more accurate "Nature Run," errors were underestimated for up to 48 hours. The model also overestimated error growth rates and misjudged where errors were strongest. By adding synthetic observation errors, they created a different version of the model and found that forecast errors became more similar during the early forecast period, with correlations increasing for up to 72 hours before decreasing.