ABSTRACT Precisely deconstructing the numerical composition of advertising cost‐conversion efficiency (ACTE) and uncovering its microlevel associative mechanisms are critical for enterprises to optimize marketing decisions. To overcome the limitations of existing methods regarding modeling accuracy and result interpretability, this study constructs an analytical framework integrating machine learning with interpretable artificial intelligence. Using 174,584 real‐world microlevel advertising records from a Chinese telecommunications company in July 2024, we systematically compared six representative machine learning models, rigorously evaluating them against a historical average baseline. The findings reveal that the XGBoost model achieves optimal fitting and generalization performance, with a test set coefficient of determination of 0.8236 ± 0.2061, significantly outperforming the baseline and other models. Furthermore, SHAP‐based interpretability analysis not only precisely quantifies the contribution of each feature to the ACTE value but also systematically reveals underlying nonlinear dependencies and threshold effects. Specifically, the advertising conversion rate plays a dominant role; exposure frequency shows a step‐like positive association; production cost exhibits a reversal effect demarcated by the median; click‐through rate and play volume display discrete stratified structures; and average display cost contributes minimally. Specification tests indicate that ACTE interpretability primarily stems from algebraic relationships of constitutive features, with limited independent contributions from non‐constitutive features. Theoretically, this study provides microlevel evidence for nonlinear advertising associations; methodologically, it establishes a rigorous modeling‐to‐interpretation paradigm; and practically, it offers empirically grounded decision support for advertising exposure regulation and cost optimization.
ABSTRACT This paper develops a class of models to capture the joint dynamics of asset returns and multivariate volatility. Owing to its two‐stage construction, this class is referred to as the multivariate two‐stage volatility‐return (MTS‐VR) model. The first stage of the MTS‐VR model focuses on modelling the dynamic trend of the multivariate stochastic volatility process. This stage is constructed by calibrating the model using time‐varying covariance matrices, which are estimated from multiple financial asset return series. Within this component of the MTS‐VR framework, we propose two new classes of dynamic volatility models formulated under a Bayesian paradigm. The first class directly models the stochastic dynamics of the covariance matrix, whereas the second separately models each asset's stochastic volatility and the dynamics of the correlation structure. The second stage of the MTS‐VR model involves the modelling of the multivariate asset returns profiles, conditional on the trend model of the multivariate volatility dynamic obtained from Stage 1. This involves a class of multivariate Student‐ t stochastic models, conditionally driven by the volatility dynamic of Stage 1 and a return trend process. We evaluate the performance of the proposed models using both simulated and financial data. The models are implemented in Rtan within a Bayesian framework. The simulation study examines the effect of the degrees of freedom parameter of the Wishart distribution in the data‐generating process. In the empirical application, the second model specification is preferred based on its superior performance.
ABSTRACT This study examines the predictive value of tourist‐generated images and text for daily tourism demand forecasting. Building on established L 0 ‐regularized extreme learning machine (ELM) formulations and ADMM‐based sparse optimization, we construct an ensemble forecasting framework, termed ES‐ L 0 ‐ELM, that combines image‐derived and textual predictors with numerical variables. The framework integrates sparse output‐weight estimation with model selection and weighted aggregation. Experiments on Jiuzhaigou and Mount Siguniang show that ES‐ L 0 ‐ELM achieves lower forecasting errors than the evaluated benchmarks. On the Jiuzhaigou dataset, it reduces MAPE by 18.66% and 14.33% relative to recent tourism forecasting models BERT‐ELM and TFIDF‐Tran‐ELM, respectively. Image features provide greater predictive gains than text features, whereas their combination yields the best performance. This study provides a sparse ensemble forecasting framework and empirical evidence of the complementary value of tourist‐generated images and text.
ABSTRACT Decomposition methods are widely used to process the nonlinear and nonstationary characteristics of tourism‐demand time series by decomposing the original data into different components with different timescales. However, the new issue is that some components after the first decomposition are still unstable and irregular, and there remains limited evidence on how to further process these unstable and irregular components effectively. Therefore, in this study, a novel secondary decomposition–ensemble approach, including empirical mode decomposition (EMD) and variational mode decomposition (VMD), is first proposed to forecast tourism demand. Correspondingly, the point and interval forecasting results show that the proposed approach generally improves forecasting performance relative to the selected benchmark models, particularly in the normal‐period and political‐unrest datasets. The COVID‐period results are more mixed, suggesting that the proposed framework remains competitive under extreme disruption but does not uniformly dominate all one‐stage decomposition alternatives across all markets and horizons.
ABSTRACT This paper introduces machine learning methods, combined with the diversity of investors' behavior, to form a comprehensive momentum indicator to measure the momentum effect in the Chinese stock market. We find that neural networks are superior in capturing the nonlinear relationships of financial data. Furthermore, we empirically verify the existence of momentum effects generated by heterogeneous investor behaviors. Our results reveal a robust multi‐time‐horizon linkage momentum effect after controlling for risk factors, with medium‐ and long‐term sentiment serving as key determinants in shaping momentum indicator. Through the spanning regression test, the paper finds that the traditional momentum returns cannot fully explain the deep‐learning‐based momentum LS returns. Economically, information uncertainty and investor attention constraints partially account for the formation of multi‐term momentum. Moreover, the momentum effect is amplified under up‐trending markets, high‐sentiment regimes, expansionary business cycles, and low economic policy uncertainty. This study deepens the understanding of the existence and economic sources of the momentum effect across multiple time horizons.
