Financial analysts' earnings forecast is one of the most critical inputs for security valuation and investment decisions. However, it is challenging to utilize such information for two main reasons: missing values and heterogeneity among analysts. In this paper, we show that one recent breakthrough in nonlinear tensor completion algorithm, CoSTCo , overcomes the difficulty by imputing missing values and significantly improves the forecast accuracy in earnings. Compared with conventional imputation approaches, CoSTCo effectively captures latent information and reduces the tensor completion errors by 50%, even with 98% missing values. Furthermore, we show that using firm characteristics as auxiliary information we can improve firms' earnings prediction accuracy by 6%. Results are consistent using different performance metrics and across various industry sectors. Notably, the performance improvement is more salient for the sectors with high heterogeneity. Our findings imply the successful application of advanced ML techniques in a real financial problem.