•  
  •  
 

Corresponding Author ORCID Identifer

Muhammad Syahid Pebriadi https://orcid.org/0000-0001-7403-2142

Section

Regular Issue

Abstract

Stock price forecasting with a uniform approach is a challenging process due to nonlinearity, regime changes, and seasonality. The following paper would like to address the practical problem of selecting an appropriate model through comparing three widely used models such as Auto Regressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM), and Prophet based on how they are applied to different attributes of data and forecast-related tasks. The structured direct comparison will be our proposal, as to be able to identify the strengths and shortcomings in abiding by each model. Each of the model is methodologically trained and tested on the same data stock price series and parameterized by the time of standard errors: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE) and the Mean Absolute Percent Error (MAPE). Experimental design is founded on the canonical setups and reasonable hyperparameters to an algorithm and the results are measured based on the hold-out samples, which have volatile intervals. The results show that under the conditions of intricate dynamics and large volatility, LSTM delivers the greatest general results with MAE = 277.50, MSE = 165,244.61, RMSE = 406.50 and MAPE = 3.09%. ARIMA performs well on stable, approximately linear and stationary segments but struggles with nonlinearity and long-term dependencies. Prophet is effective when strong seasonal patterns are present and is robust to missing values and outliers. We conclude that ARIMA suits stable series, LSTM is preferable for complex and rapidly changing markets, and Prophet is advantageous under pronounced seasonality. These findings offer practical guidance for model selection aligned with data properties and forecasting goals.

Download

Share

COinS