Effect of Representation of Information in the Input of Deep Learning on Prediction Success

Stock price prediction has been always a hot topic in financial mathematics. There are lots of studies with various techniques on this area. Prediction studies are based on the assumption that the next price has a relation with the past values. On the other hand, artificial neural networks and especially deep learning is one of the best tools for extraction of the relations between inputs and outputs. Naturally, deep learning is becoming a widely used tool day by day. There are some developed neural nets especially for time series like RNN and LSTM. Besides the structure of the neural net, the type of the input is also important for the success of the prediction. Same information can be represented in different formats without any loss. This study investigates the effect of representation of the information in the input on the prediction success. Experiments are done with the USDTRY day close values starting from 1994 till 2017. 4 different representations are tested namely time series itself, Real DFT, DWT and DCT. The transforms used are lossless and time series can be reconstructed exactly so they are just different representations of the information. 64 tests with different time spans are conducted. The same neural net structure with 512 input neurons, 3 hidden layers, 10 output neurons is used for all tests except a slight difference for Real DFT. The results show that DCT gives the best results due to its high energy compaction property. Energy compaction results a sorted array in terms of significance therefore results stable weight calculation in neural nets.