Missing values are widespread in many real world applications. It is often preferred to receive in real-time the high quality complete tuples, rather than an incomplete one containing null attribute values. The requirements for high quality and real-time response make the task of missing value imputation much challenging. Most of existing approaches employ off-line processing, thus hardly produce in real-time the high quality imputation for the null values. Inspired by the incremental learning approaches, we devise a model, called OL-MVI, which seeks to produce in real-time high quality candidate on-the-fly for missing values in query results. In OL-MVI, complete tuples are analysed incrementally and summarized into a compact correlation matrix, and a scoring function is devised to guide online imputation in real-time. Additionally, with more data being analysed, more missing values may be imputed with high quality. Our experimental evaluations on real datasets demonstrate the effectiveness of our proposed model.