MANE is a new joint project from NCBI/EMBL-EBI to address this specific question. A beta version of data is now available.

We’re leveraging public deep sequencing datasets to optimize 5’ and 3’ UTR endpoints to more accurately reflect transcriptional processes. To pick representative transcripts, we’ve developed computational methods to evaluate and integrate transcript expression levels, protein conservation, support from archived transcript submissions, clinical relevance, and other factors. Complex genes are subject to review by annotation experts from both groups to agree on a representative transcript and often make improvements to both annotation sets.

Assembly and maintenance process for RefSeq records was published in this handbook in section "how data are assembled and maintained".