Caching is a cornerstone problem in competitive analysis, with both unweighted and weighted variants achieving tight randomized competitive ratios of $\Theta(\log k)$, where $k$ denotes cache size. This paper studies weighted caching augmented with a per‑request prediction of the next arrival time of the same page. In the unweighted setting, perfect predictions are sufficient to implement Belady's offline optimal rule. However, for weighted caching, prior work shows that a logarithmic lower bound persists even with perfect predictions on adversarial request sequences, and stochastic assumptions alone do not help because classic lower bounds apply to uniform independent requests. We ask whether combining these two sources of information can break the barrier, and answer affirmatively. Assuming each request is drawn independently from an arbitrary known static distribution, page costs are arbitrary, and predictions are perfect, we devise an efficient eviction policy with a competitive ratio of $15.75$. The analysis hinges on a novel ex‑ante relaxation of the cache covering constraints, yielding a new lower bound on the optimal cost. The algorithm degrades gracefully under prediction errors: we prove a bound of $O(\min\{1+E,\log k\})$, where $E$ aggregates each page's average prediction error and may be unknown to the algorithm.
Review: The result shows that, when request distribution is known and per‑request predictions are available, weighted caching can achieve a competitive ratio far below the traditional logarithmic bound, while also offering robustness to prediction noise, providing valuable insights for practical cache management.