Return value prediction (RVP) is a useful technique that enables a number of program optimizations and analyses. Potentially high overhead, however, as well as a dependence on novel hardware support remain significant barriers to method level speculation and other applications that depend on low cost, high accuracy RVP. Here we investigate the feasibility of software-based RVP through a comprehensive software study of RVP behaviour. We develop a structured framework for RVP design and use it to experimentally analyze existing and novel predictors in the context of non-trivial Java programs. As well as measuring accuracy, time, and memory overhead, we show that an objectoriented adaptive hybrid predictor model can significantly improve performance while maintaining high accuracy levels. We consider the impact on speculative parallelism, and show further application of RVP to program understanding. Our results suggest that software return value prediction can play a practical role in further program optimization and analysis. Categories and Subject Descriptors D.3.3 [Programming