Data　prediction　is　proposed　in　wireless　sensor　networks　(WSNs)　to　extend　system　lifetime　by　avoiding　transmissions　of　redundant　messages.　Existing　prediction-based　approaches　can　be　classified　into　two　types.　One　focuses　on　historical　data　reconstruction　and　proposes　backward　models,　which　incur　uncontrollable　delay.　The　other　focuses　on　the　future　data　prediction　and　proposes　forward　models,　which　require　additional　transmissions.　This　letter　proposes　a　hybrid　model　with　the　capabilities　of　both　historical　data　reconstruction　and　future　data　prediction　to　avoid　additional　transmission　and　control　delay.　Two　algorithms　are　proposed　to　implement　this　model　in　real-world　WSNs.　One　is　a　stagewise　algorithm　for　sensor　nodes　to　build　optimal　models.　The　other　is　for　the　sink　to　reconstruct　and　predict　sensed　values.　Two　WSN　applications　are　simulated　based　on　three　real　data　sets　to　evaluate　the　performances　of　the　hybrid　model.　Simulation　results　demonstrate　that　the　proposed　approach　has　high　performance　in　terms　of　energy　efficiency　with　controllable　delay.