Predicting the Quality of Life Index: A Comparative XGBoost–LSTM Study with SHAP-Based Explainable AI

Author:
Open Access
Publisher website
Google Scholar
Abstract
In this study, the prediction performance of different artificial intelligence algorithms was examined using quality of life data from 2016 to 2025. The analysis compared gradient-boosted tree-based XGBoost with LSTM, which has the capacity to model time series and sequential dependencies. In addition, SHAP analysis was applied to ensure the model's explainability and to identify the key factors affecting quality of life. The findings show that both models successfully capture quality of life patterns, with the LSTM model achieving higher out-of-sample accuracy than XGBoost (higher R² and lower MAE, RMSE, and MAPE). SHAP analysis revealed that Purchasing Power and Pollution are the factors with the strongest impact on quality of life. The decisive effect of Purchasing Power indicates that macroeconomic conditions such as real income level, price stability, and Purchasing Power Parity -adjusted welfare indicators directly reflect quality of life. Other factors, such as cost of living, housing price/income ratio, security, healthcare services, climate, and commute time, were found to have varying degrees of importance across countries. These findings emphasize the priority of designing macroeconomic frameworks targeting income/wage policies and price stability alongside policies aimed at improving environmental conditions. The results obtained indicate that policy makers should focus on the efficient allocation of resources. The results obtained provide policymakers with an evidence-based roadmap for the efficient allocation of resources and demonstrate that more detailed analyses can be conducted using different explainable artificial intelligence methods for future research. Additionally, to test the robustness of the model, different training/testing splits, alternative error metrics, and hyperparameter sensitivity analyses were performed; the direction and magnitude of the main findings were found to be consistent across these scenarios. Finally, SHAP-based findings provide a starting framework for policy simulations, enabling the quantitative prediction of potential welfare gains from targeted improvements in specific sub-indices.

This publication has 23 references indexed in Scilit: