Publications

Overview: This paper addresses a significant gap in econometric theory by formally establishing the asymptotic properties of the Conditional Maximum Likelihood Estimator (CMLE) for ARMAX-GARCH models. While CMLE is widely employed in financial time series to model conditional heteroskedasticity with exogenous regressors, rigorous proofs of its efficiency have historically been limited. This research provides explicit, sufficient probabilistic conditions under which the CMLE is strongly consistent, asymptotically normal, and efficient—ultimately attaining the Fisher information bound. Key technical contributions include establishing geometric convergence rates for the recursive approximations of initial residuals and volatility seeds, alongside demonstrating the almost sure consistency of the outer product of score contributions.

Overview: This paper bridges the gap between deep learning architectures and econometric inference in high-frequency trading by introducing the Two-Time-Scale Transfer Learning (TTSTL) framework. To overcome the non-stationarity and computational bottlenecks of pure neural networks, TTSTL couples a high-capacity CNN-LSTM “backbone”—which extracts complex, non-linear representations from event-driven Limit Order Book (LOB) data—with a lightweight ARMAX-GARCH “adapter” that rapidly recalibrates to local market dynamics. Regulated by an interface for dynamic feature selection via Flexible adaLASSO, this hybrid approach transforms deterministic neural network outputs into rigorous probabilistic forecasts. Validated on real-world Intel (INTC) Market-by-Order data, the TTSTL framework significantly outperforms standalone deep learning baselines in predictive accuracy.

Overview: This research investigates the primary drivers of membership and market penetration within Medicare Advantage (MA) health plans by analyzing a high-dimensional dataset of plan benefits, costs, and CMS quality metrics. To overcome the severe multicollinearity and overparameterization inherent in complex healthcare data, this study deploys Principal Components Regression (PCR) coupled with forward variable selection and rigorous residual diagnostics. By orthogonalizing the feature space through PCA before executing the regression, the model successfully isolates the precise statistical impact of specific plan features. The findings quantify the elasticity of market share with respect to financial constraints and reveal the outsized impact of non-financial drivers such as drug coverage, nutritional benefits, and star ratings.

Overview: This paper presents a structural overhaul of the Colombian Health Insurance Program’s (SGSSS) risk adjustment framework, addressing the severe financial and operational inefficiencies of its legacy age-sex-territory compensation model. By processing over 18 million person-years of exposure and mapping native diagnostic codes to 255 distinct clinical categories, this research engineers a unified, prospective, condition-based risk adjustment model. The proposed system replaces fragmented, retrospective legal compensations with a rigorous predictive framework calibrated to handle heavy-tailed claim distributions. Evaluated via tolerance curves, predictive ratios, and Area Under the Curve (AUC) metrics, the new model demonstrates superior accuracy in forecasting healthcare expenditures.