Predicting demand of a product or service can be crucial for the well-being of most companies. Inventory planning, production scheduling, cash flow planning, decisions concerning staffing levels and other kinds of decisions can all depend on the precision of forecasts. Making these predictions as precise as possible leads to better customer service and higher satisfaction levels making the customers more likely to buy again. In addition, some kinds of costs get lower due to the prevention of unplanned emergency restocking and there is a lower possibility of excessive stock levels and unsold products. However, despite the importance of accurate demand forecasts, surveys have shown that the most used method for demand forecasting (48%) between companies in the USA depends on spreadsheets, while at the same time only 11% of the companies use specialized software. Available forecasting software uses a big variety of methods that suit the nature and category of the product and accuracy metrics are used as a way to compare different methods. The purpose of this undergraduate thesis is to develop and explore the appropriateness of an improved hybrid model, which was initially proposed to solve the problem of univariate timeseries predictions, in the case of predicting number of sales. The proposed model uses a modern neural network which has shown promising results in the lasts years compared to a classic feedforward neural network. In addition, an extra feature is used with the hope to further improve the predictions produced by the neural network.