Central AsianResearch Index

Forecasting Tunisian Stock Market Dynamics Using Neural Networks

Predicting stock market dynamics remains challenging, especially in emerging and frontier markets, where high uncertainty, limited liquidity, and information asymmetries can reduce the robustness of predictive models. This study aims to compare the forecasting performance of Artificial Neural Networks (ANN) and Long Short-Term Memory (LSTM) models for the TUNINDEX20 against a naïve persistence benchmark. The information base consisted of daily data on the TUNINDEX20 index of the Tunis Stock Exchange from February 6, 2012, to December 31, 2024, including 1,915 observations. The sample was divided chronologically into a training set of 1,532 observations and a test set of 383 observations. The results showed that the artificial neural network slightly outperformed the long short‑term memory model in terms of key accuracy metrics: for the first model, MAE was 12.06, RMSE was 15.92, MAPE was 0.31%, MASE was 1.0769, R² was 0.9912, and the accuracy of predicting the direction of index movement was 53.91%. At the same time, the naive model showed slightly lower forecast errors (MAE = 11.76; RMSE = 15.82), and the Diebold–Mariano test did not reveal a statistically significant advantage of ANN over the naive model (p = 0.7285). The results obtained show that a more complex architecture does not automatically provide an advantage in forecasting TUNINDEX20. Future research could include adding technical indicators and macro‑financial variables, testing models in other emerging markets, and applying adaptive and hybrid forecasting methods.

Сведения

Опубликовано
2026 · Т. 21, № 3, С. 89-103
Язык
Английский
ISSN
1997-9967 · 2663-550X
Издатель
The economy: strategy and practice, Institute of Economics Science of the Republic of Kazakhstan
Аффилиации
University of Sousse