Directional Forecasting of Pension Investment Funds: Evidence on the Majority Class Illusion from ARIMA, CNN, and LSTM Models

Demir Ş., Kurt M. A., Erdemir Ö. G.

VII. International Applied Statistics Congress (UYIK-2026) , İstanbul, Turkey, 11 - 13 May 2026, pp.1, (Summary Text)

  • Publication Type: Conference Paper / Summary Text
  • City: İstanbul
  • Country: Turkey
  • Page Numbers: pp.1
  • Hacettepe University Affiliated: Yes

Abstract

Private pension systems play a critical dual macroeconomic role by mobilizing long-term savings and alleviating fiscal pressure on public pay-as-you-go schemes. In Turkey, the Individual Pension System has experienced substantial growth, making the accurate forecasting of pension fund returns both an academic and practical necessity. This study evaluates the directional forecasting performance of traditional statistical methods and deep learning models on Turkish individual pension investment funds obtained from TEFAS, using weekly data from April 2021 to April 2026. Specifically, three pension funds from Allianz Yaşam —ALZ, AZS, and AMZ— are analyzed, representing low-, medium-, and high-risk profiles, respectively. The study adapts and extends the methodology of Van der Burgt (2023), originally developed for the NASDAQ 100 index, to the Turkish pension fund market. Three models were compared: the classical Autoregressive Integrated Moving Average (ARIMA) model and two deep learning architectures, namely Long Short-Term Memory (LSTM) networks and one-dimensional Convolutional Neural Networks (1D-CNN). More than 700 model configurations were trained and evaluated using multiple performance metrics, including Accuracy, Balanced Accuracy, Sensitivity, Specificity, F1-score, and binomial p-values. The results reveal that the “Majority Class (MC) Illusion” is particularly pronounced in low-risk, monotonically trending funds such as ALZ, where high accuracy can be misleading. Importantly, 1D-CNN models exhibit consistently lower MC exposure rates than LSTM models across all risk categories. In the high-risk AMZ fund, the LSTM model with a full feature set achieved statistically significant predictive performance, clearly outperforming the naive baseline. Overall, the findings demonstrate that evaluating forecasting performance in pension funds requires careful consideration of class imbalance and risk structure, and that deep learning models, particularly 1D-CNN, offer more robust alternatives under these conditions.