Artificial intelligence subfields and application domains: A three-layered enriching artificial learning architecture for BIST 100 forecasting via traditional and AI-based methods–TriNet-Fin

Financial time series have complex features such as high volatility, non-linear patterns, and sudden changes in trends. These characteristics make it very difficult to model and predict them correctly. In this context, traditional econometric models have structural limitations in capturing these dynamics, while pure artificial intelligence models often lack contextual information. This reduces their ability to generalize and weakens their prediction performance. The aim of this study is to develop a multi-layered hybrid artificial intelligence learning structure, named TriNet-Fin, to better predict the complex structures in financial time series, and to evaluate the forecasting performance of this structure in comparison with both traditional and AI-based approaches. The proposed model is tested against traditional time series models such as Vector Autoregression (VAR), Vector Error Correction Model (VECM), Seasonal ARIMA with exogenous variables (SARIMAX), and the Kalman Filter; as well as artificial intelligence models including Long Short-Term Memory (LSTM), Multi-Layer Perceptron (MLP), Two-Dimensional Convolutional Neural Network (2D-CNN), Gated Recurrent Unit (GRU), and Vanilla Recurrent Neural Network (Vanilla RNN). The performance of these models is evaluated both in their original forms and in hybrid form integrated with the TriNet-Fin architecture. The models were analyzed using daily data from the period 2002 to 2024. In this context, the dependent variable is the BIST 100 index, while the independent variables include interest rate (TLREF), exchange rate (USD/TRY), Credit default swap (CDS), gold ounce price (USD), Brent oil, VIX index, NASDAQ, and S&P500 indices. Model performances were evaluated on the test set using Mean Squared Error (MSE), Adjusted R², and Accuracy metrics. The empirical findings show that all pure models produced high errors and low R² values on the test data, indicating a failure in generalization. In contrast, the hybrid models supported by TriNet-Fin achieved high positive R² (Adj. R² > 0.70), low MSE, and over 97% accuracy, providing successful and reliable forecasts. These results demonstrate that a multi-layered hybrid structure enriched with contextual information plays a decisive role in financial forecasting by overcoming the limitations of individual methods. This learning structure aims to provide a methodological contribution to the literature and to offer a more reliable decision support framework for practitioners.