Backpropagation is a learning algorithm in supervised learning using a multilayer model. Selection of initial weight in backpropagation can use the initialization method randomly or by initializing the Nguyen Widrow method. The selection of the initial weights greatly affects the neural network in reaching the global minimum (or maybe only local) on the error value, and affects the speed or failure of the training process towards convergence. Therefore, the aim of this study is to determine the accuracy of the smallest error rate of the 5-7-1 neuron model in the Levenberg – Marquardt algorithm between using the random and the Nguyen Widrow initial weighting method. The method used in this research is to combine qualitative and quantitative research methods or what is called a mixed methods research. The data source used in this study is random data. The data is applied in ANN using Matlab software to generate error values in the Levenberg – Marquardt algorithm. The results of the error value were then analyzed using descriptive statistical tests and Paired t-test. The test results show that based on the Descriptive Test, the initial weighting method is obtained which is more accurate in terms of the smallest error rate, which is the initial weighting using the Nguyen Widrow method with an overall average error value of 0.00018700683. Meanwhile, based on the Paired t-test statistical test with a level of α (alpha) = 5%, it results in a sig = 0.101, so it can be concluded that the mean error value of the initial weighting method of Random and Nguyen Widrow is the same (not different) significantly. Meanwhile, using the level of α = 15% indicates that H0 is rejected and the MSE value with the Random initial weight method is higher than Nguyen Widrow.