Statistician & data scientist · credit risk and statistical modeling · Mathematics, UNAM · Mexico City
I build statistical models that other people can check and rely on: explicit assumptions, honest uncertainty, and code anyone can rerun. My recent projects are in credit risk and loan pricing (survival models for default, out-of-time validation, portfolio loss simulation against the Basel formula); my paid work and publications are in clinical trials, health data and national surveys.
- Statistical consultant (part-time, remote) for a Canada-based clinical-epidemiology research team: I fit and validate Bayesian longitudinal models in Stan on individual-patient trial data (convergence diagnostics, cross-validated model comparison), build R pipelines that turn 12 years (2010–2021) of Mexican health-system administrative records into analysis-ready panel data, and estimate national prevalences from ENSANUT under its complex survey design.
- Collaborating statistician for clinical research at Hospital Infantil de México Federico Gómez (co-author, Nutrients 2026).
- Earlier, statistics intern in INEGI's research directorate: weighted estimation on national household and employment microdata (ENIGH, ENOE), and a job-quality index for the Ministry of Labour (STPS) built as a latent variable with principal components, complemented with regression and cluster analysis.
- Teaching assistant at UNAM's Faculty of Sciences since 2023: 13 course sections (about 360 students) in Probability, Statistical Inference, Nonparametric and Regression Models, Multivariate Analysis, Survival and Time Series Models, and Stochastic Simulation. I teach Statistical Inference and Regression around real-data case studies (pedestrian crashes, mortgage pricing).
| Project | Question | Result |
|---|---|---|
| Mortgage credit risk · code | Why did 2006 mortgages default almost six times more than 2017 ones? 100,000 Freddie Mac loans followed for up to 20 years (Python, SQL). | About 86% of the gap is the economy, not the borrowers. Competing-risks survival (naive Kaplan-Meier nearly doubles the default probability), a Cox model validated out of time (ranking transfers, calibration does not), a WoE scorecard that matches XGBoost, LightGBM and a neural network out of time, and a Monte Carlo of portfolio losses that matches the Basel capital formula, with importance sampling and MCMC. |
| Do mortgage brokers charge more? · code | What sets the rate on a US mortgage? 46,190 loans originated in 2024. | Not the channel: brokers −0.5 bp vs retail (95% CI −1.9 to +0.9) for the same risk, while a FICO of 660–679 costs +23 bp. Nonparametric tests, regression diagnostics, ANCOVA and MANCOVA show brokers write larger, higher-LTV loans at the same rate; a raw comparison gets the correspondent gap backwards. |
| Randomised trial: diet during cancer treatment · code | Does an anti-inflammatory diet beat the standard low-residue diet for women receiving chemoradiotherapy for cervical cancer? (Frontiers in Nutrition 2026) | I directed the statistical analysis of the trial: 136 women randomised, attrition handled with intention-to-treat and per-protocol models, multiple imputation (35 datasets) and mixed models. Complete response 70.7% vs 54.8% (RR 1.29, 95% CI 0.92–1.8). |
| Pedestrian crashes in Mexico City · code | Is someone hit by a car every 30 minutes at rush hour? Tested with two years of emergency-call records (C5). | Every 36.6 minutes (95% CI 34.6–38.8): the claim overstates the frequency by about 22%. Getting there meant catching three data traps: a window bias that passed every diagnostic, crashes confirmed twice, and seasonal heterogeneity. A negative binomial frequency GLM with exposure is then validated out of sample. |
| Quick count and survey estimation · code | How should 900 polling stations be sampled to estimate a national vote share, using all 163,666 stations of the 2021 election? | Stratifying by district gives a margin of error of ±0.87 points; concentrating the sample in 30 districts, ±3.9. The survey part reproduces INEGI's official housing estimates exactly. |
| Imputing counts with excess zeros · code | Which imputation model works best for a zero-inflated variable in 583,678 patient records? (BSc thesis) | Zero-inflated Poisson had the lowest error, but only because its imputations were too narrow: it kept about half of the true variance. |
| Maternal diet and congenital heart defects · code | Is a mother's combined B-vitamin intake associated with complex heart defects in her child? (Nutrients 2026) | Each unit of a PCA-based B-vitamin index: adjusted odds ratio 0.707 (95% CI 0.586–0.843). |
- Vilchis-Gil J, Gómez-Jiménez AM, et al. Maternal Dietary B-Complex Vitamin Pattern and Risk of Complex Congenital Heart Defects in a Mexican Population. Nutrients 2026; 18(16): 2693. doi:10.3390/nu18162693
- Luvián-Morales J, et al., Gómez-Jiménez AM, et al. Benefits of an anti-inflammatory diet compared to a low-residue diet during concurrent chemoradiation therapy for patients with locally advanced cervical cancer: a randomized clinical trial. Frontiers in Nutrition 2026. doi:10.3389/fnut.2026.1835417
Methods: credit risk (PD, competing risks, Cox models with out-of-time validation, Vasicek/Basel portfolio loss simulation) · generalised linear models, including frequency models with exposure · ANOVA/ANCOVA/MANCOVA · survival analysis · Bayesian models and MCMC · multiple imputation · survey sampling and complex designs · clinical-trial analysis · meta-analysis · machine learning (gradient boosting with XGBoost and LightGBM, neural networks, SHAP, PCA, clustering, discriminant analysis, SVM, k-NN)
Tools: R · Python · SQL · Stan · Git · Quarto · LaTeX