PERBANDINGAN MODEL CMAC DAN ARTIFICIAL NEURAL NETWORK TIME-DELAY UNTUK PREDIKSI RETURN SAHAM BERBASIS LOG RETURN

Abstract

Financial time series modeling faces a fundamental challenge due to the dominance of stochastic components and short-term temporal dependence that cannot be adequately represented by static models. This study aims to compare the predictive capability of static and dynamic models in forecasting daily stock price changes based on log return transformation. The approaches examined include high-order polynomial regression as a baseline model, the Cerebellar Model Articulation Controller with explicit lag structure, and the Artificial Neural Network Time-Delay model. The dataset consists of daily observations of stock code WIFI obtained from Google Finance, covering the period from January 2, 2025 to December 30, 2025 with a total of 240 observations. Model performance is evaluated using a time-based train–test scheme to assess out-of-sample prediction accuracy through Root Mean Squared Error, coefficient of determination, and Durbin–Watson statistics. The empirical results indicate that the polynomial model fails to generalize, producing extremely large prediction errors and strongly autocorrelated residuals. In contrast, the CMAC model with explicit lag structure and the ANN Time-Delay model achieve substantially lower prediction errors and residuals that approximate a random process. The comparable performance of the two dynamic models suggests that explicit temporal memory representation plays a dominant role in improving predictive accuracy, while additional nonlinear approximation complexity provides limited benefit. This study concludes that financial time series modeling is more appropriately formulated as the estimation of a dynamic system with limited memory. The integration of explicit lag structure is shown to be essential for representing stock price dynamics more realistically.