Arbi Haza Nasution, Anggi Hanafiah, Winda Monika, Rajalingam Sokkalingam and Mohd Sham Mohamad
Adv. Artif. Intell. Mach. Learn., 5 (2):3809-3833
1. Arbi Haza Nasution: Universitas Islam Riau
2. Anggi Hanafiah: Universitas Islam Riau
3. Winda Monika: Universitas Lancang Kuning
4. Rajalingam Sokkalingam: Universiti Teknologi PETRONAS
5. Mohd Sham Mohamad: Universiti Malaysia Pahang Al-Sultan Abdullah
DOI: 10.54364/AAIML.2025.52216
Article History: Received on: 02-Apr-25, Accepted on: 10-Jun-25, Published on: 17-Jun-25
Corresponding Author: Arbi Haza Nasution
Email: arbi@eng.uir.ac.id
Citation: Arbi Haza Nasution, et al. Assessing Lag-Llama in Probabilistic Time Series Forecasting for the Indonesian Stock Market. Advances in Artificial Intelligence and Machine Learning. 2025;5(2):216.
Accurately predicting stock prices is crucial for investors and policymakers. This paper presents the first empirical evaluation of Lag-Llama, a novel probabilistic time series forecasting model, for predicting stock prices on the Indonesian Stock Exchange (IDX). By applying Lag-Llama to both univariate and multi-time series forecasts of key IDX stocks, we assess its ability to capture temporal patterns and market volatility, particularly in comparison to state-of-the-art models like DeepAR (RNN) and Temporal Fusion Transformer (TFT). Our results show that in fine-tuning scenarios Lag-Llama achieves a Continuous Ranked Probability Score (CRPS) of 0.0195 on a combined dataset of three major stocks (BBCA, BMRI, and AMRT), closely matching TFT (CRPS 0.0179) and outperforming DeepAR (CRPS 0.0270). However, forecasting across broader stock groups (Top 1–9 and Top 10–18 by market capitalization) proves more challenging, with CRPS values rising (e.g. 0.0517 for the Top 1–9 stocks). This study demonstrates Lag-Llama’s potential as a robust tool for stock price prediction—particularly for select, closely-related stock groupings—offering improved precision and reliability compared to traditional methods.