Actively traded assets are among the most widely predicted time series in the world. When it comes to time series predictions, sequence to sequence deep learning models are particularly effective as they take into account the temporal properties of the input features. Yet, due to the competitive nature of the market, there is relatively little publicly available research on deep learning-based financial time series predictions. Many of the published papers promise excellent results. But most of them lack essential information, which makes the reproduction of their results impossible and thus highly questionable. With a transparent approach, in this thesis, I try to find out whether accurate predictions of indices, stocks, or cryptocurrencies are possible. In addition, I investigate if the predictions can be improved by using macroeconomic data as well as the volume of specific search terms on Google, which should reflect the general market sentiment. Neither the queried fundamental data nor the selected sentimental data could improve the predictions. While for most assets accurate predictions over the entire test period were not possible, a thoroughly good predictive accuracy could be observed on the most volatile days. This makes the prediction of high volatility assets like Bitcoin particularly interesting and provides a potential trading opportunity.