Mortality improvement has been a common phenomenon since the 20th century and the human longevity continues to prolong. Post-retirement life receives a lot of attention and the need for modelling mortality rates of the elderly (ages 65 and beyond) is essential because life expectancy has reached the highest level in history. Mortality models can be divided into two groups: relational and stochastic models, but there is no consensus which model is better in modelling the elderly’s mortality rates. In this study, instead of choosing either relational or stochastic models, we propose a synthesis model, selecting and modifying appropriate models from both groups, which not only has satisfactory estimation result but also can be used for mortality projection. We use the data from U.S., U.K., Japan, and Taiwan to evaluate the proposed approach (Data source: Human Mortality Database). We found that the proposed model performs well and is a possible choice for modelling the elderly’s mortality rates.