ABSTRACTReduction in child mortality is one of the United Nations Sustainable Development Goals for 2030. In Brazil, despite recent reduction in child mortality in the last decades, the neonatal mortality is a persistent problem and it is associated with the quality of prenatal, childbirth care and social-environmental factors. In a proper health system, the effect of some of these factors could be minimized by the appropriate number of newborn intensive care units, number of health care units, number of neonatal incubators and even by the correct level of instruction of mothers, which can lead to a proper care along the prenatal period. With the intent of providing knowledge resources for planning public health policies focused on neonatal mortality reduction, we propose a new data-driven machine leaning method for Neonatal Mortality Rate forecasting called NeMoR, which predicts neonatal mortality rates for 4 months ahead, using NeoDeathForecast, a monthly base time series dataset composed by these factors and by neonatal mortality rates history (2006-2016), having 57,816 samples, for all 438 Brazilian administrative health regions. In order to build the model, Extra-Tree, XGBoost Regressor, Gradient Boosting Regressor and Lasso machine learning regression models were evaluated and a hyperparameters search was also performed as a fine tune step. The method has been validated using São Paulo city data, mainly because of data quality. On the better configuration the method predicted the neonatal mortality rates with a Mean Square Error lower than 0.18. Besides that, the forecast results may be useful as it provides a way for policy makers to anticipate trends on neonatal mortality rates curves, an important resource for planning public health policies.Graphical AbstractHighlightsProposition of a new data-driven approach for neonatal mortality rate forecast, which provides a way for policy-makers to anticipate trends on neonatal mortality rates curves, making a better planning of health policies focused on NMR reduction possible;a method for NMR forecasting with a MSE lower than 0.18;an extensive evaluation of different Machine Learning (ML) regression models, as well as hyperparameters search, which accounts for the last stage in NeMoR;a new time series database for NMR prediction problems;a new features projection space for NMR forecasting problems, which considerably reduces errors in NRM prediction.