Time series forecasting via ARCH-type cascade artificial neural network model

Time-series prediction problems can be under two approaches: probabilistic and nonprobabilistic approaches. Probabilistic models are approaches based on statistical approaches such as Autoregressive (AR), Moving Average (MA), Autoregressive Moving Average (ARMA), Autoregressive Conditional Variance Model (ARCH), and Generalized Autoregressive Conditional Variance Model (GARCH). In addition, Artificial Neural Networks (ANN), one of the non-probabilistic time series porecasting models, also known as computational methods, are widely and successfully used as a time series forecasting tool. In this study, a time series prediction that can model linear and non-linear relationships between inputs and outputs together in a single simultaneous process with an ARCH structured artificial neural network model has been introduced. Within the scope of this thesis, the Cascade Forward Artificial Neural Network was transformed into a recurrent structure and the training of the network was studied with a genetic algorithm. To demonstrate the performance of the ARCHCascade Forward-ANN proposed within the scope of this thesis, the time series (BIST100) and CIF2016 (2016) consisting of 106 observations (the highest 100 stocks traded in Borsa Istanbul in terms of market and trading volume) 48-time series used within the scope of the International Time Series Forecasting Competition. As a result of 6 analyzes performed for the BIST-100 time series, it was seen that the proposed model produced satisfactory and superior prediction results. In addition, as a result of 50 iterations performed for each of the 288 analyzes for the CIF2016 time series, it was seen that the proposed ARCH-CF-ANN model produced the best results in 71% of all cases and competitive prediction results in the others.