The NAR-NN model will be applied in time series forecasting, namely data on confirmed cases of Covid- 19 in East Kalimantan Province. The use of time series data as the basis for forecasting so that it can recognize patterns that occur which can then be used as a reference to predict the number of cases that will occur. This research data is 300 daily data for the time period from October 23, 2020 to August 18, 2021, which follows a nonlinear pattern and experiences an upward trend. In this study, the best architecture was determined for the NAR-NN model using the sigmoid activation function and the Levenberg-Marquadt Backpropagation training algorithm. The NAR-NN architecture consists of three layers, namely the input layer, the hidden layer, and the output layer. The evaluation model used is the Mean Absolute Percentage Error (MAPE). The results of this study by experimenting with the number of hidden neurons showed that the model with the best architecture at the time of delay was 4 and the number of hidden neurons was 8 with the MAPE value forecast with actual data of 7.5083%.