Large language models (LLMs) have shown strong reasoning capabilities in knowledge graph question answering (KGQA), yet their predictions are not always supported by explicit and verifiable reasoning paths. Existing methods often retrieve or generate plausible candidate paths, but path plausibility alone does not establish that a prediction actually depends on the retrieved evidence. To address this issue, we propose FBV-KGQA, a forward-backward verification framework for reliable and interpretable KGQA. FBV-KGQA first derives executable reasoning paths through relation-level planning and evaluates their forward support for candidate answers. It then identifies critical entity and relation pivots and applies schema-constrained counterfactual perturbations to assess their influence on model predictions. Reverse questions are subsequently constructed from candidate answers to verify whether they can be consistently traced back to the original question entity. Finally, an adaptive selection strategy integrates forward support, counterfactual effects, and backward consistency. Experiments on WebQSP and CWQ show competitive performance against representative KGQA baselines, while human evaluation indicates improved traceability, clarity, and usefulness of the resulting reasoning paths.