d3rr1ck27/sa-youth-unemployment-model

"Machine learning model predicting youth NEET status in South Africa using QLFS microdata | Python · scikit-learn · Streamlit # 🎓 SA Youth NEET Risk Predictor **Live Demo** South Africa has a serious problem, about 33% of young people aged 15-24 are not in employment, education or training. I built this to quantify exactly how socioeconomic factors drive that risk, using real government survey data from Stats SA. Input a youth profile, get a NEET probability score, see how it compares to the national average, and understand which factors are actually driving the result, based on what the data says, not assumptions. --- ## What it does - Predicts NEET risk probability for any youth profile - Compares against the 33% national average - Explains which factors matter and why — backed by actual survey patterns - National stats panel showing NEET rates by province, education level, and gender --- ## What I learned building this The data tells a different story than the conventional wisdom. Work history is by far the strongest predictor — youth who previously worked but aren't currently employed show an 88.8% NEET rate. "Never worked" is actually low risk (32.1%) because most 15-19 year olds who've never worked are still in school. Provincial differences are also much smaller than expected (28-42% range across all 9 provinces). Getting the model right was only half the work. Making the explanations honest and data-driven was the other half. --- ## Tech Python · XGBoost · scikit-learn · Streamlit · Plotly · pandas --- ## Data QLFS microdata from Statistics South Africa: 5 survey years (2019–2023 Q2), ~49,000 youth records aged 15-24. Available at statssa.gov.za. --- ## Run locally ```bash git clone github.com cd sa-youth-unemployment-model pip install -r requirements.txt streamlit run app.py ``` --- ## Limitations The model reflects historical patterns in survey data, not causal relationships. It tells you who has been NEET, not necessarily who will be. That distinction matters and I tried to build the explanations with t …