This project develops an end-to-end machine learning risk prediction platform designed to identify construction activities that are likely to become overdue.
The objective was to transform traditional reactive project monitoring into a proactive risk management approach by using historical project data to predict potential delays before they occur.
The platform combines data preprocessing, feature engineering, machine learning classification, model evaluation, and risk categorisation to generate actionable insights for project decision-making.
Construction projects generate large amounts of operational data, including task statuses, actions, priorities, workflow information, and historical delays.
However, project teams often identify risks only after delays have already affected project timelines.
The key business question:
Can historical project activity data be used to predict which activities are likely to become overdue before the delay occurs?
The goal of this project was to build a predictive early-warning system that helps project teams:
- identify high-risk activities earlier
- prioritise resources effectively
- improve project visibility
- reduce reactive problem-solving
The dataset contains historical construction project activity records.
Dataset size:
12,424 records
Features include:
- Status
- Type
- Priority
- Cause
- Task Group
- Days Open
- Days Since Status Change
- Project Total Forms
- Project Total Actions
- Project Open Actions
- Project Overdue Forms
- Has Image
- Has Comment
- Has Document
This project was designed as a binary classification problem.
Target:
Overdue
Classes:
- 0 = Activity is not overdue
- 1 = Activity becomes overdue
The dataset contained a significant class imbalance, with overdue activities representing approximately 7% of total observations.
Because of this imbalance, model evaluation focused on risk detection metrics rather than accuracy alone.
The raw project dataset was transformed into a machine-learning-ready format.
Steps included:
- data quality checks
- datatype validation
- missing value analysis
- feature selection
- numerical and categorical feature processing
- preprocessing pipeline creation
Analysed:
- overdue activity distribution
- feature relationships
- project workflow patterns
- operational risk indicators
Key finding:
A model predicting only the majority class would achieve high accuracy but fail to identify genuine project risks.
Therefore, the focus was placed on correctly identifying overdue activities.
Created predictive features representing project risk behaviour.
Examples:
- Days Open
- Days Since Status Change
These captured activity ageing and workflow stagnation.
Examples:
- Project Total Actions
- Project Open Actions
- Project Total Forms
These represented workload and operational pressure.
Examples:
- Has Image
- Has Comment
- Has Document
These represented supporting information available for each activity.
XGBoost Classifier
XGBoost was selected because it performs well on structured tabular datasets and captures complex relationships between operational variables.
Advantages:
- handles nonlinear relationships
- captures feature interactions
- performs strongly on classification problems
- widely used in risk modelling applications
Because the dataset was imbalanced, performance was evaluated using:
- Precision
- Recall
- F1-score
- ROC-AUC
ROC-AUC:
0.991
The model demonstrated excellent ability to distinguish between overdue and non-overdue activities.
Overdue activity detection:
Precision:
0.59
Recall:
0.96
F1-score:
0.73
The high recall indicates that the model successfully identified the majority of activities that were likely to become overdue.
The most influential features were:
- Task Group
Certain project workstreams demonstrated higher delay risk patterns.
- Days Open
Long-running activities showed increased probability of becoming overdue.
- Type
Different activity categories displayed different risk profiles.
- Days Since Status Change
Activities with limited recent progress were more likely to become delayed.
Model predictions were converted into business-friendly risk categories.
Risk Levels:
2,117 activities
146 activities
222 activities
This enables project stakeholders to quickly identify where intervention may be required.
- Python
- Pandas
- NumPy
- Scikit-learn
- XGBoost
- Matplotlib
- Seaborn
- Jupyter Notebook
- VS Code
- Git
- GitHub
- Predictive modelling
- Risk analytics
- Classification algorithms
- Feature engineering
- Handling imbalanced datasets
- Model evaluation
- Machine learning pipelines
- Business-focused data interpretation
- Version control and documentation
Potential improvements include:
- implementing time-based validation
- adding SHAP model explainability
- deploying the model through an API
- creating an interactive risk dashboard
- integrating live project management data
- implementing model monitoring and retraining pipelines
This project demonstrates how machine learning can enhance traditional project risk management by providing early-warning signals, helping teams identify potential delays earlier and supporting proactive decision-making.