I'm interested in using our historical project data to build a predictive model. Is it possible to use Machine Learning to identify early warning signs of a budget overrun based on past performance, and what features should I include in the dataset?
3 answers
Absolutely, this is a fantastic use case for Regression analysis. You should build a dataset that includes features like: initial budget, team seniority levels, project duration, number of stakeholders, and "Requirements Volatility" (how often the scope changed). In a project I worked on last year, we found that "Time Since Last Milestone" was the strongest predictor of a cost spike. If a milestone was delayed by more than 10%, there was an 80% correlation with a 20% budget overrun. Using a Random Forest model can help you see which of these factors is the most "important" in driving costs up.
How are you planning to handle the "Dirty Data" from your older projects? In my experience, historical budget logs are often incomplete or filled with inconsistent categorization, which can ruin a model's accuracy.
Start simple with a Logistic Regression to predict a "Binary" outcome: Will this project go over budget? Yes or No. Once that is accurate, then move to predicting the exact amount.
Great advice, Lisa. A binary classifier is much easier to train and still provides immense value for early risk assessment.
That's my biggest hurdle, Kevin. I’m currently manually cleaning the last two years of data. Do you think I should focus only on the last 12 months where our logging was more disciplined, or is more data (even if slightly messy) better for training the model?