The overall ambition of this bachelor thesis is to evaluate and develop an analysis procedure which is able to valuate time series. The focus is on measuring the quality, type and best prediction method with a learning algorithm. In advance, general methods
to predict those time series are presented.
Therefore, various capabilities of prediction will be compared and result in a universal approach to prediction (neural network). Possible configurations of the forecasts are the methodology (e.g. initialization, learning algorithms) and topology (e.g. amount of input/hidden/output neurons, activation functions).
In the practical part, a basic prototype for predictions of time series based on Python and Tensorflow is implemented. After evaluating and comparing different frameworks, the detailed code base is explained and some test results are presented.
By accomplishing a series of experiments an optimized configuration for a neural network based on the Mini Dow Jones is explored. The model reaches a maximum accuracy of 66{ab9b61ccb0ae1881e9c37503622f2b6f7f2d532ca90e49b70a50c40b4c036eea} for predicting short-term trends (rising/falling) within the next 10 seconds.