Research Article
A Hybrid VAR-LSTM-GARCH Model for Multivariate Volatility Forecasting
Charles Chege*,
Martin Kithinji,
Peter Gachoki
Issue:
Volume 11, Issue 4, August 2025
Pages:
99-113
Received:
25 April 2025
Accepted:
19 May 2025
Published:
14 July 2025
Abstract: Financial markets show persistent volatility, creating barriers to achieving exact financial predictions. The forecasting of multivariate financial data requires forecasting models like the Vector Autoregressive (VAR) model for modeling linear dependencies, the Long Short-Term Memory (LSTM) model for modeling non-linear patterns, and the Generalized Autoregressive Conditional Heteroscedastic (GARCH) model that is capable of modeling volatility clustering. Each of these models fails to handle complete data complexity on its own, as they specialize in unique properties of the data. Recent studies have been carried out that enhance forecasting accuracy by combining two models. The first case is the VAR-GARCH model, which can model linear and volatility clustering aspects but fails to model non-linear dependencies. Another case is the LSTM-GARCH model that can explain non-linear dependencies and volatility patterns, but fails to explain linear dependencies. A third instance is the VAR-LSTM model that can explain the linear and volatility aspects, but fails to model the non-linear patterns. However, there is a need to have a model that can combine the three models to explain the linear, non-linear, and volatility aspects in financial time series data collectively. This research fills this gap by combining VAR, LSTM, and GARCH into a VAR-LSTM-GARCH hybrid model, which provides improved forecasting. This study uses historical five-year daily data for VIX, US Dollar Index, and S&P 500 E-mini futures obtained from Yahoo Finance. The model-building process involves constructing a VAR (9) model selected using AIC criteria to reveal linear dependencies. The residuals from the VAR are used to train an LSTM model to capture nonlinear trends. The residuals of the LSTM are then used to fit an M-GARCH (1, 1) model, which generates volatility cluster estimates. The VAR-LSTM-GARCH hybrid model demonstrates superior performance with substantial improvements across all evaluation metrics compared to individual models, showing consistently lower prediction errors and enhanced forecasting accuracy. The progressive three-stage modeling approach demonstrates that each component contributes incrementally to forecasting performance, with the incorporation of volatility modeling through GARCH being particularly effective in enhancing predictive accuracy. The research suggests using this hybrid model for volatility prediction on multiple portfolios and emphasizes future development of real-time diagnostic processes. The new approach delivers an advanced instrument that helps financial analysts work efficiently by effectively capturing the complex interdependencies in multivariate financial time series data.
Abstract: Financial markets show persistent volatility, creating barriers to achieving exact financial predictions. The forecasting of multivariate financial data requires forecasting models like the Vector Autoregressive (VAR) model for modeling linear dependencies, the Long Short-Term Memory (LSTM) model for modeling non-linear patterns, and the Generalize...
Show More
Research Article
Predicting Wealth Index Using an Ensemble Model of Random Forest and Multilayer Perceptron
Issue:
Volume 11, Issue 4, August 2025
Pages:
114-124
Received:
30 July 2025
Accepted:
15 August 2025
Published:
12 September 2025
DOI:
10.11648/j.ijdsa.20251104.12
Downloads:
Views:
Abstract: Wealth inequality remains a significant challenge in Kenya, exacerbated by the limitations of traditional wealth measurement methods. This study develops and evaluates an ensemble wealth index model combining Random Forest (RF) and Multilayer Perceptron (MLP) algorithms to improve prediction accuracy using socio-economic data from the 2019 Kenya Population and Housing Census (KPHC). The models were assessed using performance metrics such as accuracy, precision, sensitivity, specificity and ROC-AUC. The Random Forest model, configured with 500 trees and a split variable count of 2, achieved 34.3% accuracy, 58.54% balanced accuracy, 65.98% out-of-bag error and performed best on the ”Poorest” class with a 13.7% class error. It further recorded 41.13% precision, 34.27% recall, 83.17% specificity, and a 67.31% AUC. The MLP model, using sigmoid activation in hidden layers and softmax in the output layer, achieved 33.4% accuracy, 57.7% balanced accuracy, 30.6% precision, 32.54% recall, 82.86% specificity and AUC of 69.12%. The RF-MLP ensemble model outperformed the individual models with a 34.4% accuracy, 37.42% precision, 34.05% recall, 83.2% specificity and AUC of 68.55%. Despite modest overall accuracies, the ensemble model showed enhanced balanced accuracy and specificity, particularly in extreme wealth categories. However, classification of middle and poor wealth levels remains challenging due to feature overlap and class imbalance.
Abstract: Wealth inequality remains a significant challenge in Kenya, exacerbated by the limitations of traditional wealth measurement methods. This study develops and evaluates an ensemble wealth index model combining Random Forest (RF) and Multilayer Perceptron (MLP) algorithms to improve prediction accuracy using socio-economic data from the 2019 Kenya Po...
Show More
,
Anthony Waititu
