This study introduces predictive models for automated algorithm performance improvement during runtime, overcoming the traditional prediction approaches that focus only on the final performance outcome. Leveraging sequential iteration data from the run, including solution locations and values (window size), such models forecast performance improvements after a predefined number of iterations (predictive window). The real-time prediction capability enables early evaluation of algorithm hyperparameters, potentially saving substantial computational time. A long short-term memory (LSTM) network, chosen for its efficacy with longitudinal data, is tested and analysed in the scenario where personalization is made to runs from a single problem instance. Utilizing the CEC2014 benchmark suite with 10-dimensional problem instances and various DE configurations, our approach using exploratory landscape analysis features shows promising results by surpassing a standard baseline prediction model.