The latest studies have compellingly argued that Neural Networks (NN) classification and prediction are the right direction for forecasting. It has been proven that NN are suitable models for any continuous function. Moreover, these methods are superior to conventional methods, such a Box–Jenkins, AR, MA, ARMA, or ARIMA. The latter assume a linear relationship between inputs and outputs. This assumption is not valid for skimmed milk powder (SMP) forecasting, because of nonlinearities, which are supposed to be approximated. The traditional prediction methods need complete date. The non-AI-based techniques regularly handle univariate-like data only. This assumption is not sufficient, because many external factors might influence the time series. It should be noted that any Artificial Neural Network (ANN) approach can be strongly affected by the relevancy and “clarity” of its input training data. In the proposed Convolutional Neural Networks based methodology assumes price series data to be sparse and noisy. The presented procedure utilizes Compressed Sensing (CS) methodology, which assumes noisy trends are incomplete signals for them to be reconstructed using CS reconstruction algorithms. Denoised trends are more relevant in terms of NN-based forecasting models’ prediction performance. Empirical results reveal robustness of the proposed technique.