Among various approaches for implementing prognostic algorithms data-driven algorithms are popular in the industry due to their intuitive nature and relatively fast developmental cycle. However, no matter how easy it may seem, there are several pitfalls that one must watch out for while developing a data-driven prognostic algorithm. One such pitfall is the uncertainty inherent in the system. At each processing step uncertainties get compounded and can grow beyond control in predictions if not carefully managed during the various steps of the algorithms. This paper presents analysis from our preliminary development of data- driven algorithm for predicting end of discharge of Li-ion batteries using constant load experiment data and challenges faced when applying these algorithms to randomized variable loading profile as is the case in realistic applications. Lessons learned during the development phase are presented.