This repository compares several word embeddings methods and Machine Learning techniques in order to construct a suitable model for computing the sentiment of the National Bank of Romania (BNR) quarterly inflation reports. The purpose is to identify if the narrative contains information that can be utilized to improve the precision of numerical forecasts. Our approach shows that deep learning models can be applied effectively for financial sentiment analysis, and that a recurrent neural network is the best model for predicting financial sentiment. The results show that FinBERT outperforms the BERT base model when used in combination with neuronal network such as DNN, CNN and LSTM.