We address the problem of computing the infimum accumulated weight for the reachability of some goal location in weighted timed automata. While there already exist efficient techniques to solve this problem, we propose here a backwards symbolic algorithm computing the accumulated weight to the goal, instead of the accumulated weight from the initial state. Going backwards has in itself a few advantages: most notably it does not require any extrapolation operation to ensure termination. Also it may be more efficient than going forward if the set of co-reachable states is smaller than the set of reachable states. Backwards algorithms are also instrumental in several problems beyond reachability, like control problems for instance. We obtain our backward algorithm by proposing extensions of the classical action and time predecessor operations on zones to account for weights. We have implemented the approach and report on its performance.