Profile photo: Jerry in a cream and grey plaid shirt, head and shoulders, looking at the camera. A wheelchair arm and wheel are visible at the left edge, with pale curtains behind him.

Jerry Bannister Zachary

Statistician · Data scientist

GitHub Portfolio Chess.com Lichess Letterboxd Email

My ideas, AI’s implementation — I fuse the two, and I am a transhumanist about where that leads.

I am a statistician at Abubakar Tafawa Balewa University in Bauchi, Nigeria, where I am a Senior Statistician in the Academic Planning & Quality Assurance unit — building the institutional databases that sixty-odd departments report into, and advising the research projects that pass through them. Since 2017 I have also run JJMB Datascience Hub, an independent consultancy doing much the same work for clients in public health, finance and academia.

I am currently a PhD student in Statistics at ATBU, working on hybrid Bayesian deep learning for spatio-temporal risk forecasting. My MSc is also from ATBU and my BSc from the University of Maiduguri, finishing in at least the top two of my class in both instances. I put it that way on purpose. I am strongly against the GPA system: it compresses years of uneven and largely incomparable work into one number, and then invites everybody to treat that number as a finding rather than as the very lossy summary it is. Rank is the least misleading version I can give, and the actual evidence is the work — the posts here and my projects.

My interest is in statistics as a whole — every method in it, not a chosen subset of favourites. The fuller range is below.

This site is where I check that work instead of assuming it. A simulation that disagrees with what I expected is the most useful thing I can publish, so most of what is here is written to be run again rather than read once.

Education and honours

PhD Statistics, Abubakar Tafawa Balewa University (2026–present) — in progress: hybrid Bayesian deep learning for spatio-temporal risk forecasting.

MSc Statistics, ATBU (2024) — best in class, top 1%. Thesis Close-Knit Regression, supervised by Prof. A. Abdulkadir.

BSc Statistics, University of Maiduguri (2013) — second best in class, top 2%. Undergraduate project M/M/c queuing analysis of automated teller machines at Fidelity Bank, Maiduguri, supervised by Prof. Haruna Rann Bakari.

MTN Foundation Scholarship (2009–2012), with a place on the Dean’s List every year of that period; commendation from ATBU’s Directorate of Academic Planning (2021). Registered Statistician, CISON (member ID 20260732), since September 2026.

Seven peer-reviewed papers in health and applied statistics — Google Scholar · ORCID.

Selected work

  • Tobacco price intelligence — XGBoost demand forecast at 10.9% MAPE against a 13.8% naive baseline; fine-tuned FinBERT at 76.8% against a 71.0% majority class.
  • Crude blend optimizer — yield surrogates at R² ≥ 0.982 drive differential evolution to within $0.005/bbl of the exact LP optimum, with VaR/CVaR over 10,000 price scenarios.
  • Disability rights assistant — quotes the 2018 Act and the Constitution word for word with a citation on every excerpt; held-out strict recall 0.63–0.65, zero false refusals, WCAG 2.2 AA with no axe violations across 33 views.
  • Urban agriculture, where it concentrates — Moran’s I = 0.300 across 775 LGAs, only 19 significant after FDR control; density falls 66% per doubling of area, and the effect survives an explicit control for uneven OpenStreetMap coverage (ΔAIC ≈ 20).
  • curb-cut — arcsine Fay–Herriot small-area estimates of adult disability prevalence, with a bias-corrected back-transform and a parametric-bootstrap MSE.

Toolbox

  • Languages — Python · R · SQL · TypeScript/JavaScript
  • Statistics and ML — PyMC, scikit-learn, XGBoost, TensorFlow/Keras, SciPy optimisation, survival, spatial and time-series models
  • Apps and reporting — Quarto, Shiny, Streamlit, Gradio, React, Tableau, Power BI
  • Engineering — Git, GitHub Actions, Docker, Supabase, Cloudflare Pages, Render
  • Accessibility — WCAG 2.2 AA, keyboard navigation, screen-reader layouts, axe audits

Areas I work in

A map of the ground, not a claim to have covered all of it. Where something here has a post behind it, the post is the evidence; where it does not, read it as interest rather than track record. Each domain opens only if you want it.

Statisticsinference, Bayesian, regression, survival, time series, spatial, causal
  • Foundations — probability and distribution theory, asymptotics, estimation theory (maximum likelihood, moments, M-estimation), sufficiency, information and decision theory
  • Inference — hypothesis tests, confidence and credible intervals, likelihood-ratio/Wald/score, power and sample size, multiple comparisons and FDR, equivalence testing
  • Bayesian — priors, elicitation and conjugate analysis, MCMC (Metropolis–Hastings, Gibbs, HMC/NUTS), variational inference, ABC, hierarchical and multilevel models, posterior predictive checks, LOO-CV, WAIC, Bayes factors, model averaging
  • Regression — linear, GLM, GAM, mixed-effects and GLMM, quantile, robust, ridge/lasso/elastic net, splines, nonlinear regression, overdispersion; logistic, ordinal and multinomial; Poisson, negative binomial, zero-inflated and hurdle models
  • Survival — Kaplan–Meier, log-rank, Cox, parametric AFT, competing risks, frailty, time-varying covariates, recurrent events, joint longitudinal–survival models
  • Time series — ARIMA/SARIMA/ARIMAX, exponential smoothing, state-space and Kalman filtering, structural time series, VAR/VECM and cointegration, GARCH, spectral analysis, change-point detection, probabilistic forecasting and backtesting
  • Spatial — geostatistics and kriging, variograms, CAR and SAR, INLA, point processes, geo-additive models, disease mapping, small-area estimation
  • Multivariate — PCA, factor analysis, canonical correlation, discriminant analysis, MANOVA, cluster analysis, multidimensional scaling, correspondence analysis, SEM
  • Causal inference — potential outcomes, DAGs and do-calculus, propensity scores and matching, inverse probability weighting, doubly robust estimation, instrumental variables, regression discontinuity, difference-in-differences, synthetic control, mediation, sensitivity analysis
  • Design and surveys — RCTs, factorial and fractional factorial, blocking, split-plot, crossover, response surface, optimal design, A/B and sequential designs; sampling designs, stratification and clustering, complex weighting, raking, non-response, design-based variance, questionnaire design
  • Missing data and resampling — MCAR/MAR/MNAR, multiple imputation, EM, weighting, sensitivity analysis; bootstrap, jackknife, permutation tests, cross-validation, Monte Carlo, simulation study design
  • Nonparametrics and applied fields — rank-based tests, kernel density estimation, nonparametric regression, empirical likelihood; biostatistics and epidemiology, econometrics and panel data, psychometrics and IRT, reliability and extreme values, statistical process control, meta-analysis, demography, actuarial and risk modelling
  • Specialised data — compositional, functional data analysis, directional and circular statistics, network data, longitudinal and panel data
Machine learningsupervised, ensembles, unsupervised, probabilistic, interpretability, MLOps
  • Supervised and ensembles — linear and logistic regression, k-NN, naive Bayes, SVM, discriminant analysis, decision trees; bagging, random forests, extra trees, gradient boosting (XGBoost, LightGBM, CatBoost), AdaBoost, stacking and blending
  • Unsupervised and representation — k-means, hierarchical clustering, DBSCAN/HDBSCAN, Gaussian mixtures, spectral clustering, association rules; PCA and kernel PCA, ICA, NMF, t-SNE, UMAP, autoencoders, manifold learning, learned embeddings
  • Probabilistic ML — Gaussian processes, Bayesian optimisation, probabilistic graphical models, Bayesian networks, hidden Markov models, Markov random fields
  • Model development — feature engineering and selection, encoding strategies, pipelines, hyperparameter search (grid, random, Bayesian, Optuna), nested cross-validation, learning curves
  • Uncertainty and hard cases — probability calibration, conformal prediction, prediction intervals; class imbalance (SMOTE, cost-sensitive learning), concept drift, distribution shift, small-sample regimes, censored targets
  • Interpretability — SHAP, LIME, partial dependence and ICE, permutation importance, counterfactuals, surrogate models
  • Other paradigms — semi-supervised, self-supervised, active, online and streaming, transfer, multi-task, meta- and federated learning; bandits, MDPs, Q-learning, policy gradients, actor–critic, offline RL
  • Specialised systems — recommenders (collaborative filtering, matrix factorisation, hybrid), learning to rank, anomaly and fraud detection, survival ML, uplift modelling, forecasting at scale
  • Responsible ML and MLOps — fairness metrics, bias auditing, differential privacy, model governance, reproducibility; experiment tracking, model registries, feature stores, containerised deployment, batch and real-time serving, drift monitoring, CI/CD, shadow and A/B deployment
Deep learningvision, transformers and LLMs, generative, graphs, efficiency
  • Fundamentals — backpropagation, activation and loss design, initialisation, optimisers (SGD, momentum, RMSProp, Adam/AdamW), learning-rate schedules, batch/layer/group normalisation, dropout, weight decay, early stopping, gradient clipping, mixed precision
  • Vision — CNNs, ResNet, DenseNet, EfficientNet; detection (YOLO, Faster R-CNN, DETR); segmentation (U-Net, Mask R-CNN); augmentation; vision transformers
  • Sequence and attention — RNN, LSTM, GRU, sequence-to-sequence, attention mechanisms, temporal convolutional networks
  • Transformers and NLP — BERT, GPT, T5, tokenisation, embeddings and vector search, NER, text classification, summarisation, translation, question answering
  • Large language models — pretraining, supervised and parameter-efficient fine-tuning (LoRA, QLoRA), instruction tuning, RLHF and DPO, prompt engineering, retrieval-augmented generation, evaluation and hallucination measurement, quantisation, distillation, agentic workflows and guardrails
  • Generative and graphs — VAEs, GANs, diffusion models, normalising flows, autoregressive generation; GCN, GraphSAGE, graph attention networks, node classification and link prediction
  • Deep forecasting and scientific ML — DeepAR, N-BEATS, N-HiTS, temporal fusion transformers, probabilistic deep forecasting; PINNs, neural ODEs, operator learning, surrogate modelling
  • Uncertainty and multimodal — Bayesian neural networks, Monte Carlo dropout, deep ensembles, evidential learning, classifier calibration; contrastive learning (SimCLR, MoCo), CLIP, masked modelling, speech and ASR, vision-language models
  • Adaptation, efficiency and interpretability — transfer learning, fine-tuning, domain adaptation, few- and zero-shot, curriculum learning; pruning, quantisation, distillation, ONNX and TensorRT, edge inference, distributed GPU training; saliency maps, Grad-CAM, integrated gradients, attention visualisation, probing, mechanistic interpretability

Chess

I play as BayesianRules on both Chess.com and Lichess, with a rapid peak of 2011 on Chess.com and a blitz peak of 2072 on Lichess. It is the one place I get to reason about an opponent’s uncertainty rather than a distribution’s, which turns out to be less of a change of subject than it sounds.

Hobbies

Films, television and books, mostly in the evenings. What I watch goes on Letterboxd — the profile is the log, so it is a fairer account of my taste than anything I would write here.

Getting in touch

batesthommie@gmail.com
+234 706 167 4831
Bauchi State, Nigeria

Consulting enquiries, collaborations, and questions about anything here are all welcome by email. Corrections especially — particularly the ones that show a simulation here was set up wrong. Those can go by email too, or as an issue on the repository if you would rather they were public.