Cyprus ICU Forecasting, Uncertainty & Ensemble Lab
Real ECDC COVID-19 ICU data → ARIMA + LSTM → validation-weighted ensemble → P10/P50/P90 → calibration → 1-, 2- and 4-week forecasts
Research portfolio demonstrator showing practical experience in time-series forecasting, model validation, ensemble modelling and uncertainty-aware prediction using real healthcare data.
Why this demonstrator matters
This Space compares a classical statistical model (ARIMA) with a deep-learning model (LSTM), learns ensemble weights only from validation data, calibrates uncertainty on a separate calibration period, and evaluates the final system on unseen held-out data.
It represents the forecasting foundation for my current research. More advanced methods such as POMDP-based decision-making and risk-constrained reinforcement learning are part of my current methodological development rather than claimed prior expertise.
What this demonstrates
- Existing capability: time-series preprocessing, ARIMA, LSTM, validation, ensemble forecasting, MAE/RMSE, Python
- Developing capability: P10/P50/P90 probabilistic forecasts, uncertainty calibration, coverage, interval width and WIS
- Next research step: POMDP-based sequential decision support and risk-constrained reinforcement learning
Research/teaching demonstrator only — not for clinical or public-health decisions.