This paper proposes a new sea surface wind speed retrieval modeling algorithm based on the empirical orthogonal function (EOF) analysis for observations acquired by the global navigation satellite system reflectometry (GNSS-R). As a nonparametric modeling algorithm, it is simpler compared with the nonlinear methods. The influence of wind speed and incident angle on the modeling error is analyzed for the first-time using spectrum analysis. Three types of data from 80% CYGNSS 2019–2020 observations [Delay Doppler Map Average (DDMA) and Leading Edge Slope (LES)], signal incident angle and ERA5 (European Centre for Medium-range Weather Forecasts Reanalysis V5) reference wind speed are used in the EOF analysis to establish two retrieval models. The remaining 20% of the data are used for accuracy evaluation after getting the final wind speed by the minimum variance (MV) estimator. As a result, when using three 0–20 m/s wind speeds of ERA5, Advanced Scatterometer (ASCAT) and the Modern-Era Retrospective Analysis for Research and Applications V2 (MERRA2) as contrasts, the RMSEs are 1.51, 1.45, 1.43 m/s respectively. Compared with CYGNSS wind product, the performance of this algorithm is closer to L2 CDR (Climate Data Record) V1.1 product than V1.0. The results demonstrate that the EOF algorithm has a good performance in retrieving sea surface wind speed and can better retain the influence of the incident angle on the observations.