Week 6 – Machine Learning Models and Results

31 Jul 2026 - Sneha Dharmeche - climate risk, data science, feature engineering, NHS ERIC

Project overview

This week I implemented a baseline Random Forest regression model to predict total backlog maintenance cost. A Random Forest was chosen as an initial model to capture nonlinear relationships and handle the complex interactions between the variables and allow a good comparison to gradient boosting model.

The Random Forest model gave an R² score of 0.79, with a Mean Absolute Error (MAE) of approximately £2.9 million and Root Mean Squared Error (RMSE) of £10.9 million.

Feature importance analysis showed that estate size, energy consumption variables and age vulnerability were among the strongest predictors of backlog maintenance costs. This suggests that larger estates, higher operational demands and older infrastructure are important factors associated with increased maintenance requirements.

Following the Random Forest, I modeled an XGBoost regression model. The XGBoost model improved performance compared with Random Forest, increasing the R² score from 0.79 to 0.84 and reducing RMSE from £10.9 million to £9.6 million.

This improvement suggests that the boosting approach was better able to capture the nonlinear relationships between NHS estate characteristics and backlog maintenance costs.

Next Stage

The next stage of the project will focus on further model evaluation and interpretation. This includes analysing feature contributions in more detail and investigating whether incorporating additional climate-related variables improves predictions. I will also explore comparing machine learning results with a Bayesian modelling approach to understand uncertainty.