Prediction of Investment Alternatives with Artificial Neural Network

Abstract

Purpose – Since investment decisions are made in an uncertain environment, it is vital to develop prediction models that enable investors to make the right decision on time. Artificial neural network (ANN) method is one of the most widely used for this purpose. Thus, in the study, USA dollar Exchange rate, BIST 100 index, gold price in ounce and TL deposit interest rate are determined as alternative investments and their future values are predicted by using ANN models. The lag values of each investment alternative with other investment alternative values are considered as influencing variables. Hence, it is aimed to develop multidimensional prediction models. Design/methodology/approach – In the study, a multilayer artificial neural network model was used. As a data set obtained from the Central Bank database, 284 weekly data for the period of January 2015 and June 2020 were included in the analysis. Of this data set, 238 were used for training and 46 for testing. In the models, the lagged values of each variable and the influencing variables are included in the model as independent variables. Model trials were carried out over the hyperbolic tangent and logistic activation functions for each variable. As the error function, the sum of the squares of the error was chosen. The fast back propagation algorithm was used as the learning algorithm. Findings – ANN models were built with the dataset and processed with algorithm specified in the method part. Prediction values for each investment alternative were obtained by choosing the model with the smallest mean squares error among constructed all the models. The fact that the chosen prediction model results in very low error rates reveals that the prediction performances of the models are quite well. In addition, obtaining over 93% R2 values indicating the explanatory power of these models implies the validity of the them. Discussion – Predicting the future value of alternative investments for investors minimizes the possible risks they may encounter. Developing models such as neural networks by identifying appropriate influencing variables provides investors to do this. This study developed multidimensional models by analyzing both the relationship between alternative investments and their own lagged values. As a result, the lagged values of investment alternatives were found to be effective. This result reflects the situation that supports the assumption that the structure shown in the past will continue in the future. In fact, this imply that the effects of causal variables have been already reflected in their past values. Hence, it can be stated that ANN models based on time series data might be more preferable than models using influencing data. However, in this case, the break points or periods specific to time series analysis should be taken into account and the autocorrelation problem should be considered. Thus, it will be ensured that the prediction performance of the models emerges with better results. In addition, the better prediction values can get in shorter time by generating models with optimal parameter values.