Cumhuriyet Üniversitesi İktisadi ve İdari Bilimler Dergisi

Öz

In this study, a novel forecasting method is proposed by integrating classical type-1 fuzzy functions with deep learning models based on Long Short-Term Memory (LSTM) networks, aiming to enhance forecast accuracy and systematically address uncertainty in time series data. In the developed approach, an independent LSTM model is constructed for each fuzzy cluster, and these models are trained using lagged variables derived from the time series. During the forecasting phase, the output of each model is weighted according to the degree of membership of the corresponding observation to its respective cluster, and the final prediction is computed accordingly. In this way, the proposed Deep Fuzzy Forecasting Functions (DBF) approach simultaneously captures both the dynamic temporal patterns of the series and the structural uncertainty among observations. The training process employs the Adam optimization algorithm, and various combinations of hyperparameters—including the number of epochs, the number of hidden units, the number of clusters, and the α-cut threshold—were experimentally tested. The optimal configuration was determined based on empirical performance. The proposed DBF approach was evaluated on time series datasets with distinct characteristics, including financial and meteorological data, and benchmarked against widely used forecasting methods. Experimental results demonstrated that the proposed DBF model achieved the lowest prediction errors in many scenarios and delivered high accuracy, particularly for short- and medium-term forecasts. These findings indicate that the developed approach is flexible, generalizable, and effective as a forecasting tool across various domains.