An AI chatbot–supported hybrid neural network for solving financial problems

In this research, a novel hybrid forecasting framework is introduced, aiming to tackle the intricacies of financial time series prediction, particularly in the context of gold prices, Bitcoin market behavior, and USD/TRY currency fluctuations. The proposed system combines two distinct modeling approaches: a classical linear modeling component, utilized to uncover direct and proportional trends through vectorized inputs, and a deep learning-based component, specifically a Long Short-Term Memory (LSTM) network augmented by dense layers, which is adept at learning nonlinear temporal patterns from structured sequence data. To improve interpretability and facilitate trust in model outcomes, the framework incorporates Shapley Additive Explanations (SHAP), a game-theoretic technique that quantifies each input feature's contribution to the model's predictions. Furthermore, the system is equipped with a conversational AI assistant capable of providing users with real-time insights and textual explanations of forecast results. This interactive feature bridges the gap between complex model outputs and end-user comprehension, making the system both technically rigorous and practically accessible. Empirical evaluations demonstrate that the integrated architecture delivers superior forecasting accuracy compared to standalone models, yielding a mean absolute percentage error (MAPE) of approximately 5.2%. By aligning model complexity with interpretability and user interaction, this approach addresses key challenges in financial prediction systems. Overall, the study offers a comprehensive, explainable, and user-centric solution that can be effectively deployed in financial analytics and decision-support environments.