Skip to content
View Sbusiso-Mdingi's full-sized avatar

Highlights

  • Pro

Block or report Sbusiso-Mdingi

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Sbusiso-Mdingi/README.md

Hey, I'm Sbusiso Mdingi 👋

Actuarial science graduate · Data scientist · Quantitative and risk analyst

I'd rather show you the model than tell you about it.

I'm an actuarial science graduate from the University of the Free State, finishing my BSc Honours in Risk Analysis and taking MSc Financial Engineering modules at WorldQuant University. I build the things I talk about, in Python, R and SQL:

  • Actuarial pricing, valuation and capital models (SAM, IFRS 17)
  • Risk and quantitative models: VaR, stress testing, GARCH, copulas
  • Machine learning and statistical models, validated on data they haven't seen
  • Forecasting systems with live data behind them
  • The software and pipelines that keep them running

How I got here

I started in actuarial science because I liked the idea of putting an honest number on uncertainty. Somewhere along the way, "how much should this cost" turned into "why do people price risk the way they actually do," and that question is what keeps me building. My honours research, recovering how investors trade off risk and return across JSE sectors, was me trying to answer it with data instead of intuition.

I care more about finishing the whole pipeline

data → model → decision → something a non-technical person can actually use

than about a clever result stuck in a notebook. If it can't survive contact with a real user or a real out-of-sample period, it doesn't count yet.

What I've built

🏦 Mdingi Life A fictional South African life insurer with 5,000 annuities and 50,000 funeral policies, modelled end to end on real data. Lee-Carter, CBD and Renshaw-Haberman mortality on Stats SA deaths, yield curves on 4,610 National Treasury bond auctions, key-rate ALM, the SAM standard-formula SCR (R423m), a 1,000-scenario economic scenario generator, five stress tests and an IFRS 17 CSM roll-forward. Key-rate matching lifts solvency cover from 0.94 to 1.15, but only 60.1% of scenarios stay covered. Live dashboard.

🏥 Sequrin Technologies My startup. Claims-integrity software for South African medical schemes: claims analytics, explainable risk signals and investigator workflows, running on AWS with Docker services and infrastructure as code. Behind it sits an internal workforce of 6 production and 7 pilot AI agents on Amazon Bedrock AgentCore, with least-privilege IAM and human approval gates. Private, but the story is on mdingi.co.za.

📈 Copula-GARCH Pairs Trading Standard Bank against its peers, with ARMA-GARCH margins and a Student t copula. The honours version of this looked great: 6.2% CAGR and a 0.90 Sharpe. Then I rebuilt it properly, with rules fixed on 2010 to 2019, one validation run on 2020 to 2026, and costs on every trade, and it came out at 0.3% and 0.08. That second number is the one I trust, and it's scored live on days the model has never seen.

📉 JSE Portfolio Risk An old Shiny app that picked weights from random draws by in-sample Sharpe, rebuilt in Python and tested out of sample. Six weight rules, four VaR models, stress scenarios. Only one of the four VaR models passed its coverage backtests, which is exactly the kind of thing worth knowing. Live dashboard.

✈️ Pickup Alerts for FlySafair My mom never knows when to leave to fetch my aunt from the airport, so I built the fix. A scheduled pipeline follows each aircraft through its day, predicts the landing, adds the walk to the kerb and live Google traffic, and tells the person fetching when to leave. Replayed on 474 arrivals: passengers waiting outside for over 5 minutes went from 5.5% to 0.2%. Try it.

🌍 Spatial Statistics of SADC Development Global and local Moran's I over seven World Bank indicators for the 16 SADC countries, with a linked map and parallel-coordinates plot. Urbanisation clusters regionally (Moran's I = +0.425, p = 0.0001). Plus a kriging study in R. Explore it.

🔐 Secure Steganographic Messaging The odd one out: AES-256-GCM encryption hidden in image pixels, then attacked with the statistical tests used to catch it. Invisible to the eye, visible to RS steganalysis. Turns out "quant" and "security" scratch the same itch: both are about finding what's hidden in the noise.

Also: a health insurance pricing model in R (negative binomial frequency and Gamma severity GLMs on 100,000 members, holdout Gini 0.369, live calculator), and cross-line tail risk between life and property catastrophe losses with copulas.


Right now

  • 🎓 Wrapping up my BSc Honours in Risk Analysis (dissertation submitted)
  • 📚 MSc Financial Engineering modules at WorldQuant University
  • 🏗️ Building Sequrin
  • 🗺️ Learning a lot of spatial statistics, mostly by pointing it at real data

Toolbox

R · Python · SQL · JavaScript · VBA · LaTeX · Git Pricing & profit testing · SAM / Solvency II · IFRS 17 · Mortality modelling · VaR & stress testing Machine learning · GARCH · Copulas · GLMs · Spatial statistics & kriging · Discrete choice modelling · Monte Carlo simulation · Forecasting AWS (Bedrock AgentCore) · Azure · Google Cloud · Docker · Linux · FastAPI · Shiny · React


Find me

🌐 mdingi.co.za

💼 linkedin.com/in/ssmdingi

🔬 orcid.org/0009-0003-4417-968X

Pinned Loading

  1. Copula-Cross-Line-Insurance-Risk Copula-Cross-Line-Insurance-Risk Public

    Tail dependence between life and property catastrophe losses: copulas fitted by maximum pseudo-likelihood with GPD tails, and how far each model understates 99.5% VaR economic capital. Python.

    Python

  2. health-insurance-pricing-model health-insurance-pricing-model Public

    Health insurance pricing in R: negative binomial frequency and Gamma severity GLMs on 100,000 synthetic members, monthly risk premiums with 95% intervals, holdout Gini 0.369, live premium calculator.

    R

  3. jse-risk-management jse-risk-management Public

    JSE portfolio risk: walk-forward tests of 6 portfolio rules and 4 VaR/expected shortfall models with trading costs, coverage backtests and stress tests, with a live risk dashboard.

    Python

  4. jse-statistical-arbitrage jse-statistical-arbitrage Public

    Copula-GARCH pairs trading on JSE banks: ARMA-GARCH margins and a Student t copula, rules fixed on 2010-2019, validated once on 2020-2026 with costs, scored live on unseen data.

    Python

  5. mdingi-life mdingi-life Public

    A fictional South African life insurer modelled end to end on real data: Lee-Carter/CBD mortality, annuity pricing, bond yield curves, ALM, SAM SCR, IFRS 17 CSM, 1,000-scenario ESG and stress tests…

    Python

  6. stego-messenger stego-messenger Public

    AES-256-GCM encryption hidden in images with keyed LSB matching, tested against RS, chi-square and HCF-COM steganalysis on a reproducible benchmark. Live demo on Azure.

    Python