Elasticity property of cloud computing enables dynamic provisioning of computing resourcesrnin order to match changes in application demand at run time. Cloud elasticityrnsolution acquires or releases cloud resources automatically using either proactive or reactivernmechanism. Hybrid of the two approaches has been proposed in previous researchesrnto take their advantages and mitigate their drawbacks; however, proposed approaches dornnot address proactive model inaccuracy before resource deployment and after system loadrnbehavior changes. In this research, threshold-based deviation between two-step aheadrnrecursive predictions and observed values is proposed for decision making before scalingrnactions in combining proactive and reactive models. In addition, Stacked Long-Short TermrnMemory Recurrent Neural Network has been implemented as predictive model that canrnlearn linear and nonlinear relationships of time series. The predictive model is retrainedrnautomatically with new observations at run time to adapt system load changes. To evaluaternthe proposed approach, experiments are conducted using CloudSim Plus with realrnsystem CPU usage. Empirical analysis of experimental simulation revealed signiï¬cancernof proposed solution in alleviating proactive model inaccuracy, deployment delay andrnoscillation to improve utilization eï¬ƒciency and application performance. Moreover, performancernof online retraining demonstrated that predictive model can learn new changesrnat run time to enrich proactive mechanism.