The catch-up TV (CUTV) service allows users to watch video content that was previously broadcast live on TV channels and later placed on an on-line video store. Upon a request from a user to watch a recently missed episode of his/her favourite TV series, the content is streamed from the video server to the customer’s receiver device. This requires that an individual flow is set up for the duration of the video, and since it is hard to impossible to employ multicast streaming for this purpose (as users seldomly issue a request for the same episode at the same time), these flows are unicast. In this paper, we demonstrate that with the growing popularity of the CUTV service, the number of simultaneously running unicast flows on the aggregation parts of the network threaten to lead to an unwieldy increase in required bandwidth. Anticipating this problem and trying to alleviate it, the network operators deploy caches in strategic places in the network. We investigate the performance of such a caching strategy and the impact of its size and the cache update logic. We first analyse and model the evolution of video popularity over time based on traces we collected during 10 months. Through simulations we compare the performance of the traditional least-recently used and least-frequently used caching algorithms to our own algorithm. We also compare their performance with a “perfect” caching algorithm, which knows and hence does not have to estimate the video request rates. In the experimental data, we see that the video parameters from the popularity evolution law can be clustered. Therefore, we investigate theoretical models that can capture these clusters and we study the impact of clustering on the caching performance. Finally, some considerations on the optimal cache placement are presented.