Aug 2026· Journal of Artificial Intelligence and Technology· 0 citations· 30 references
TL;DR
This study compares the predictive performance of the standard LC model with the proposed extensions: LC-ARIMA, LC-ANN, and LC-RF and suggests that integrating neural networks into the LC framework effectively captures the time-component patterns.
Abstract
Accurate mortality predictions are important for pension sustainability and life insurance valuation. Existing extensions of the Lee–Carter (LC) model typically use a fixed fitting period and rely on a single forecasting approach to predict the time component. This study proposes a hybrid mortality forecasting framework based on the LC model, with a particular focus on improving the estimation of its time component. The approach integrates an optimized selection of the fitting period with both traditional time-series modeling with auto-regressive integrated moving average (ARIMA) (p,d,q) and machine learning techniques, namely artificial neural networks (ANNs) and random forests (RFs). The objective is to assess whether these enhancements improve forecasting performance. Using 45 years of Malaysian age-specific mortality data (1980–2024), this study compares the predictive performance of the standard LC model with the proposed extensions: LC-ARIMA, LC-ANN, and LC-RF. Results showed that, while the LC ARIMA version minimizes prediction residuals by using fitting periods of 1980–2002 for males and 1980–2004 for females, the LC-ANN version achieves the highest aggregate predictive accuracy when averaged across genders. These findings suggest that integrating neural networks into the LC framework effectively captures the time-component patterns. Our projections through 2038 indicate a continuous decline in mortality rates among Malaysians, with greater improvement among females. Overall, the proposed hybrid framework offers a more accurate and flexible approach to mortality forecasting. These improvements are particularly relevant for applications such as pension planning and population projections, where reliable mortality estimates are essential for long-term policy decisions.
Background: The Accuracy of mortality rate forecasting plays an important role in various decision-making processes in the life insurance sector, including determining premium amounts. In addition, it also contributes to assessing the readiness of human resources to support national defense, as well as to conducting risk assessments aimed at maintaining demographic stability. The United States mortality data was selected as the study object due to the availability of comprehensive and high-quality.
Aims: This study explores five hybrid approaches that combine stochastic models and machine learning, along with one non-hybrid approach to assess their potential to improve forecasting accuracy.
Method: In this study, the Lee-Carter–ARIMA, Lee-Carter–Random Forest, Lee-Carter–ANN, Lee-Carter–ARIMA–Random Forest, Lee-Carter–ARIMA–ANN, and ANN models were evaluated. These models were applied to mortality rate data from nine divisions in the United States (US), stratified by gender, using training data from 1966 to 2005 and test data from 2006 to 2015. The best model is determined based on the smallest Mean Absolute Percentage Error (MAPE) value while also considering the interpretability of the model.
Result: The study's results show that, across the number of divisions, the Lee-Carter–ARIMA–Random Forest model produces the smallest MAPE values most often. However, in terms of average MAPE, the Lee-Carter–ARIMA–ANN model performs better, with MAPEs of 9.66% for females and 9.28% for males. Furthermore, neither of these models yields a substantial improvement in predictive accuracy compared with the Lee-Carter–ARIMA model.
Conclusion: Considering the relatively small decrease in MAPE and the difficulty of interpreting machine learning models due to their black box nature, the Lee-Carter–ARIMA model demonstrates the best overall performance relative to the other models. Nevertheless, the Lee-Carter–ARIMA–Random Forest and Lee-Carter–ARIMA–ANN models show potential as alternative approaches that merit further investigation and may contribute to national defense planning and support the maintenance of demographic stability.
Vita Nuarini, Mahmudi, Nina Fitriyati et al.· International Journal of App...· 0 citations
This research tests the predictive power of selected classical time-series and machine learning models for Cambodia’s LNCPI. Monthly observations from January 2008 to April 2026 were split sequentially into an 80% training set and a 20% test set. An automatic seasonal autoregressive integrated moving-average model, ARIMA(3,1,1)(1,0,0)[12] with drift was used as the benchmark and compared to Random Forest, Extreme Gradient Boosting (XGBoost), and Support Vector Regression (SVR). For forecasts within the test period, we recursively used lagged values of LNCPI and a deterministic time trend, as well as indicators for month to capture any seasonality. Root Mean Squared Error (RMSE), with the raw values characterizing forecast accuracy like Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE). The ARIMA model outperformed the others on all three out-of-sample accuracy metrics, producing an RMSE of 0.0170, MAE of 0.0145, MAPE of 0.2726%. Random Forest, XGBoost, and SVR yielded RMSEs of 0.0398, 0.0369 and 0.0474 respectively. Variable-importance results showed that most predictive information rested in the time trend and the first two LNCPI lags, with monthly seasonal indicators contributing little. As the forecast horizon increased, machine-learning forecasts began to flatten out and increasingly under-predict LNCPI. The results show that using a parsimonious seasonal ARIMA specification based on a relatively small univariate dataset has more accurate forecasts than the selected machine-learning models.
Mara Mong, Siphat Lim· International Journal of Eco...· 0 citations
Traditionally, the techniques used in forecasting life insurance sales have been based on traditional statistical models, including ARIMA and VAR, but now they have moved to data-driven artificial intelligence methods. This paper is based on a systematic and bibliometric review aimed at tracking this methodological shift, combining results from literature spanning 2000 to 2025 indexed in Scopus. The review combines quantitative bibliometric mapping with VOSviewer and qualitative synthesis based on the PRISMA approach. Findings indicate that the trend is inclined toward machine learning, deep learning, and hybrid ensemble models, which are more accurate in prediction, flexible, and have risk assessment potential. Major contributors, thematic groups, and relation networks were determined with a primary focus on hybrid AI structures and explainable modeling techniques. The paper ends by providing a roadmap that highlights domain-specific data integration, multimodal AI systems, and ethical explainability as the key to the next generation of forecasting in the insurance sector.
Dayanand Lal, Gurpreet Kaur· The Journal of Theoretical A...· 0 citations
AbstractConventional Vector Autoregressive (VAR) models are widely applied for multivariate time series analysis, their performance deteriorates in high-dimensional settings due to inefficient parameter estimation, unstable forecasts, and difficulties in interpreting temporal dependencies. This study conducts a comparative study on conventional Vector Autoregressive (VAR) models, Multivariate Random Forest for VAR (MRF-VAR) models and Multivariate Extreme Gradient Boosting for VAR (MXGB-VAR) models, validated using simulated and real-life dataset for Nigerian financial time series. Augmented Dickey–Fuller (ADF) test was adopted to test the stationarity of the data. Forecast accuracy across short-term and long-term horizons for the models were measured using Mean Absolute Deviation (MAD) and Root Mean Square Deviation (RMSD). Results for simulated data show that the conventional VAR models achieved the best short-term forecast performance (MAD = 1.642, RMSD = 2.016), while the MRF-VAR models performed best in long-term forecasting (MAD = 0.947, RMSD = 1.197). In the case of the real-life dataset for Nigerian financial time series, the MRF-VAR model performed well in short-term forecasts (MAD = 108.84, RMSD = 149.53), whereas MXGB-VAR model provided better results in long-term forecasts (RMSD = 730.57). Policymakers and financial analysts should be encouraged to apply machine learning approaches to VAR models in macroeconomic and financial forecasting to improve decision-making.
N. Isah, S. Doguwa· Indian Journal of Industrial...· 0 citations
This study proposes the Dynamic Environment-Adaptive Risk Prediction Model, an attention-based neural network architecture that integrates an environmental perception module with a dynamic indicator generation mechanism to address multi-period financial risk forecasting. Comparative experiments were conducted using multi-source datasets comprising China’s A-share listed companies’ financial reports and macroeconomic stress indices. Evaluation against baselines including LSTM, XGBoost, and Altman's Z-score using AUC and interpretability scores reveals that DEARPM achieves an AUC of 0.89 and 0.93 in recession and expansion periods, respectively—significantly outperforming benchmarks. Dynamic indicators significantly enhanced class separation, increasing the Mahalanobis distance between high-risk and standard class centroids by a factor of 3.2 and substantially improving the Fisher discriminant ratio compared to static features. Combined with SHAP-based feature attribution, this clear separation provides robust and objective interpretability for risk management under dynamic market conditions. The model demonstrates strong robustness by maintaining an AUC> 0.88 under 50% missing data and 25% noise, offering high practical value for risk management under dynamic market conditions.
The presence of linear temporal dependencies and non-linear market behaviour poses a significant challenge in the field of financial time-series forecasting. The following study presents an in-depth comparison between statistical time-series models, ensemble machine learning models and a hybrid forecasting framework trained on historical NIFTY-50 stock price data. Baseline and statistical forecasting models, including Random Walk, AutoRegressive (5), and Moving Average, were initially used to understand the linear structure. Machine learning models, Random Forest and Gradient Boosting, were later implemented to learn the non-linear patterns that arose in the data. This led to a hybrid residual-learning approach, developed in conjunction with testing, where AR(5) captured linear structures in the data and machine learning models corrected the prediction errors caused by AR(5). The models used were evaluated using error metrics like RMSE, MAPE, MAE and directional accuracy by implementing a walk-forward validation framework. The experimental results indicate that the hybrid forecasting framework demonstrated comparatively stable forecasting behaviour under walk-forward validation while producing meaningful forecasts under realistic sequential prediction settings.
Kashyap Madhav Yeleswarapu, K. B.· International Conference Com...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
