Analysis of the accuracy of artificial neural networks in financial time series: An application on BIST 100

Predicting the path to be followed by the financial series is an important field of research for both investors and researchers. In the effective market hypothesis, which stands out in the financial literature, it is noteworthy that the future path of the financial series cannot be explained from the path that they have followed in the past. In this study, it is aimed to use artificial neural networks (ANN) methods to test the validity of effective market hypothesis for Borsa Istanbul 100 (BIST100) index and to compare it with linear time series models in predicting daily BIST100 index. The sample consists of daily series for the period 6.3.2001-18.12.2017. In the first stage, the series was estimated with the fourth order autoregressive model, and in the second stage, the baseline model was compared with the ANN models with different neuron numbers. In the selection of ANN model architecture, the steps to be taken to select the optimum model architecture are suggested as follows: i. creating the input layer with the selected information criterion, ii. repeatingly estimating the number of neurons of the individual latent layer model by increasing it individually; iii. repeating the previous step for the two-layer ANN, iv. selection of the model with the best performance in training and validation subsamples, v. future forecasts are produced within the framework of the optimum model. The findings of the study point out to significant non-sample success in modeling BIST100 with ANN models and hence the effective market hypothesis may not be valid for BIST100.