Data Management Office Specialist | MIS Ph.D. Candidate, Sakarya University | M.Sc. Statistics
I work at the intersection of data management, applied statistics, and actuarial risk modeling. I am a Ph.D. candidate in Management Information Systems at Sakarya University and hold an M.Sc. in Statistics. In my professional work, I focus on trustworthy data, clear reporting, and dependable workflows; in my research, I build quantitative prototypes whose assumptions and limitations are explicit.
My projects combine statistical reasoning with practical Python: make the inputs traceable, validate each step, and make results straightforward to reproduce.
- Actuarial risk: solvency-oriented analysis, tail risk, and scenario modeling.
- Research: stochastic processes, regime analysis, and applied statistical methods.
- Data practice: quality controls, reproducible reporting, and workflow automation.
- Machine learning: interpretable classical ML and compact PyTorch experiments with clear train/validation/test splits.
| Project | What you can explore |
|---|---|
| Motor pricing engine (freMTPL2) | Frequency × severity pricing on 678k open French motor policies: Poisson GLM vs gradient boosting (Gini 0.27 → 0.32), Gamma severity and a Generalised Pareto large-loss layer (ξ ≈ 0.87). Written up as a Turkish article for Café Poliçe. |
| Where to set an XL priority | Excess-of-loss reinsurance on the Danish fire losses: Hill and POT/GPD tail, Monte Carlo capital, priority chosen by reinsurance margin + cost of capital, and a time hold-out showing that untrended models understate the lower layers. Café Poliçe article #2. |
| Large hail vs the damage record (NOAA) | 30 years of US storm data: hail ≥ 2 in grows 2.5% a year and a trend model predicted 2016–2025 within 0.2%, while the damage column of the same database under-records recent catastrophes. 2030 projection with uncertainty. Café Poliçe article #3. |
| Biogas RS-PIM Engine | Actuarial research pipeline combining OU-SDE calibration, HMM/GMM regimes, extreme-value analysis, and solvency metrics. Research prototype; inspect documented assumptions before interpreting results. |
| Demand Tail Analysis | Peaks-over-threshold tail modeling with an emphasis on risk interpretation and reproducible analysis. |
| Project | What you can explore |
|---|---|
| QMASE: Martial-Arts Analytics Engine | Evidence-first research engine: deny-by-default dataset registry, consent-gated video pipeline, Hawkes and semi-Markov models of BJJ position flow (no momentum once session rates are separated; Weibull dwell shape k ≈ 1.6), tests in CI. Exploratory: one public dataset, two athletes. |
| BJJ Semi-Markov Analysis | Transition matrices, dwell-time summaries and Kaplan-Meier sojourn curves from timestamped BJJ position labels. |
| Project | What you can explore |
|---|---|
| HMM Regime Analysis | Hidden Markov model analysis for demand regimes, with an inspectable workflow. |
| Bootstrap Inference Lab | Bootstrap confidence intervals and permutation tests for hands-on statistical inference. |
| Breast Cancer Classification | A scikit-learn classification workflow with held-out evaluation metrics; an educational dataset exercise, not a diagnostic tool. |
| CIFAR-10 CNN Baseline | A compact PyTorch computer-vision baseline with a documented dataset pipeline. |
| Task | Held-out result | |
|---|---|---|
| fremtpl2-pricing-engine | Motor claim frequency and pure premium | Gini 0.32; predicted / observed 0.999 |
| danish-fire-xl-priority | Excess-of-loss priority (EVT) | ξ = 0.50; lowest-cost priority 37.5–100 mDKK |
| noaa-hail-trend-2030 | Large-hail hazard trend and 2030 projection | 2016–2025 hold-out within 0.2% |
| noaa-storm-events-hail-1996-2025 | Dataset: 1.78M US storm events, CPI-restated damage | NOAA, public domain |
| uci-har-activity-baseline | Activity recognition from smartphone sensors | 95.5% on unseen subjects |
| bjj-position-family | BJJ position family from a 3D skeleton | 0.67 ± 0.12 (grouped CV) |
| bjj-bottom-posture | BJJ bottom-player posture | 0.80 ± 0.05 (grouped CV) |
| grapplemap-bjj-graph | Dataset: 725 BJJ positions, 1,485 transitions | from GrappleMap (public domain) |
| oxford-pets-cat-breeds | 12 cat breeds, frozen ConvNeXt features | 89.6% accuracy |
| simba-or-not | Is this my cat? (photos never uploaded) | 41 / 42 found; 1 false alarm in 2,371 |
These repositories are independent research and learning projects. Read each README for its data source, run instructions, evaluation setup, and limitations. Model scores on public benchmark datasets do not establish production or regulatory suitability.
Konya <-> Sakarya <-> Istanbul <-> Tashkent - the places connected to my work, study, and collaborations. The graphic is a personal portfolio illustration, not a live network or representation of system access.
Martial arts keep me grounded in patience, discipline, and steady practice: Taekwondo | Muay Thai | Kickboxing | Brazilian Jiu-Jitsu.
Simba 🐾 - chief morale officer and keeper of break time.
Stats cards and the contribution animation are provided by external services and GitHub Actions; they may occasionally be unavailable. Language statistics reflect public repository contents, not proficiency.
- GitHub: @000MAK
- LinkedIn: Mehmet Ali Kaya
- Hugging Face: makactuary
- ORCID: 0000-0002-7905-3337
Think rigorously. Build reproducibly. Document the limitations.