<p>Extreme multi-label text classification (XMTC) aims at assigning multiple labels to a text data point from an extremely large vocabulary of a few hundreds of thousands or millions of labels. In the last decade, many XMTC techniques have been proposed that achieve compelling performance. However, there is a dearth of simple baselines that can justify the need for complex models. In this paper, firstly we introduce a simple baseline technique for XMTC that utilizes low-dimensional random feature embeddings obtained from high-dimensional sparse features to find nearest neighbours of a given data point. The labels are then assigned using a weighted label transfer mechanism. Secondly, XMTC researchers have perpetually advocated the significance of accurate predictions of both frequent and rare labels, and have adopted conventional and weighted (propensity-scored) metrics to measure these. However, there is no single metric that can combine both and enable head-to-head comparisons of overall performance. To this end, we also introduce new evaluation metrics that are based on harmonic means of the corresponding conventional and weighted metrics at multiple retrieval ranks. Extensive experiments validate the promise of our approach, where it achieves compelling performance along with reasonably fast training and prediction time while using just one CPU core.</p>