In this paper, a likelihood based evidence acquisition approach is proposed to acquire evidence from experts' assessments as recorded in historical datasets. Then a data-driven evidential reasoning rule based model is introduced to R&D project selection process by combining multiple pieces of evidence with different weights and reliabilities. As a result, the total belief degrees and the overall performance can be generated for ranking and selecting projects. Finally, a case study on the R&D project selection for the National Natural Science Foundation of China is conducted to show the effectiveness of the proposed model.The data-driven evidential reasoning rule based model for project evaluation and selection (1) utilizes experimental data to represent experts' assessments by using belief distributions over the set of final funding outcomes, and through this historical statistics it helps experts and applicants to understand the funding probability to a given assessment grade, (2) implies the mapping relationships between the evaluation grades and the final funding outcomes by using historical data, and (3) provides a way to make fair decisions by taking experts' reliabilities into account. In the data-driven evidential reasoning rule based model, experts play different roles in accordance with their reliabilities which are determined by their previous review track records, and the selection process is made interpretable and fairer. The newly proposed model reduces the time-consuming panel review work for both managers and experts, and significantly improves the efficiency and quality of project selection process. Although the model is demonstrated for project selection in the NSFC, it can be generalized to other funding agencies or industries.