Forecasting and risk#
Forecasts with calibrated uncertainty and risk rankings that say where to look first, for infrastructure regulators, city governments, grid operators, retailers and manufacturers. Eight projects, 2019 to 2026.
Cost and duration forecasting for water infrastructure#
Problem · A UK water regulator needs independent cost and duration estimates for the schemes it reviews, across reservoirs, desalination, water treatment, wastewater treatment and water pipelines, each with its own cost drivers.
What I did · Built a forecasting tool with separate cost and duration models per asset class, trained on thousands of projects from dozens of countries collected with agents. Model settings were chosen by cross-validation over a thousand candidates, and accuracy was measured by leave-one-out testing on every completed project. Handed over as a CLI with YAML pipelines, a web dashboard and a security report.
Result · Every cost and duration test beat the naive baseline, with cost error cut by more than a third in the strongest classes. Ready for operation by the regulator.
Stack · Python, gradient boosting, YAML pipelines, FastAPI dashboard.
Tree-fall risk scoring for urban forests#
Problem · Cities manage thousands of trees with limited inspection capacity. Falling trees put people and property at risk, so inspections should go where the risk is highest, and a model trained in one city should ideally transfer to another.
What I did · Built an explainable risk-ranking model for two Italian cities from tree inventories, maintenance logs and reanalysis weather data, using gradient boosting in learning-to-rank mode and evaluating it by how many falls the top-ranked trees capture. Delivered a technical report through eight revisions and a reproducible repository, and tested cross-city transfer.
Result · Inspecting a quarter of the trees in ranked order covers most of the falls, several times better than random inspection. A small set of genus and species combinations accounts for the majority of falls.
Stack · XGBoost, Python, ERA5-Land, GIS data.
Railway construction cost forecasting#
Problem · A high-speed railway under construction in the UK, with cost projections in the tens of billions of pounds, has become controversial over delays and cost increases. Planners need up-to-date estimates for monitoring and assurance.
What I did · Trained forecasting models on empirical data from current and past projects, calibrated to UK data with quantile mapping, and reported forecast progress during mid-delivery against existing reference-class estimates. A second round added schedule-forecasting scenarios. The models combine conformal prediction, transfer learning and machine learning.
Result · The forecasts are used by senior leadership to direct strategy.
Stack · Python, conformal prediction, quantile mapping, Jupyter.
Risk Management Lead
"Thanks to Michele and the team for preparing the work; it presents a powerful new way of forecasting."
Early-warning forecasts for a public works portfolio#
Problem · A city government oversees hundreds of active capital projects. Its surveillance system needed better forecasts of cost and schedule overruns, sound handling of outliers, and spend forecasts at project and portfolio level, to flag issues early.
What I did · Over two phases, built six forecasting tasks: overrun ratios, overrun labels, and project and portfolio spend. Trained gradient-boosted models on hundreds of completed projects, modeled spend curves first as Beta distributions and then with smooth interpolation, added conformal-prediction uncertainty in the second phase, and packaged everything as a reusable Python library with a handover presentation. Co-authored the client reports. The work was presented at ML Week Europe.
Result · A double-digit reduction in cost and schedule overrun error, label accuracy above four in five, and a further cut in error in the second phase.
Stack · XGBoost, conformal prediction, Python, Jupyter.
Client management, public works department
"Can we extend the models to offer additional capabilities?"
Cost benchmarking for an Australian grid operator#
Problem · A transmission operator needed to benchmark the cost of transmission and substation projects against comparable schemes, with costs made comparable across countries and years.
What I did · Built a reference-class forecasting study: a benchmarking tool, Power BI prototypes, and cost normalization with World Bank and national price indices.
Result · Benchmarks delivered for the operator's investment planning.
Stack · Excel/VBA, Power BI, reference-class forecasting.
Contact center forecasting for an online furniture retailer#
Problem · Contact centers forecast demand per channel to schedule agents. Padding against uncertainty is expensive, so the forecasts for the current week and four weeks out had to get better.
What I did · Researched and deployed driver-based forecasting of weekly phone volume: seasonal ARIMA, Prophet, random forests, gradient boosting, elastic net and ensembles, fed by order forecasts, promotions, holidays and shipping drivers. Took the best models to production on Vertex AI pipelines with weekly retraining, evaluation, CI and alerting.
Result · Error cut by about a third, with most weeks within a few percent in production, and estimated savings in the millions per year.
Stack · BigQuery, Vertex AI pipelines, Prophet, XGBoost, scikit-learn, Python.
Senior Manager, Machine Learning
"I wholeheartedly endorse Michele and his efforts, as his contributions have been instrumental in the successful implementation and deployment of the forecast models."
Performance forecasting for megaprojects#
Problem · A megaproject is a large, complex venture that takes years to build, involves many stakeholders and affects millions of people. Managing it means monitoring progress, forecasting delays and costs, and adjusting budgets in time.
What I did · Developed an AI-based early-warning system for spotting high-risk projects, including outlier prediction and forecasts of future portfolio spending.
Result · The system draws on thousands of project-years of combined construction activity and tens of billions in aggregate cash flow.
Stack · Python, gradient boosting, time-series modeling.
Head of Data Science
"This is excellent."
Demand forecasting for the textile industry#
Problem · Textile production runs through pre-treatment, dyeing, printing and finishing over several months, so a manufacturer must produce the right quantity at the right time to keep delivery times and unsold stock low.
What I did · Examined the historical data through exploratory analysis, identified data-quality issues and additional data sources that would improve accuracy, developed and evaluated a demand forecasting model, and presented the findings in a detailed report.
Result · Report and model delivered to the consulting agency and its client.
Stack · AWS Forecast, JupyterLab, Pandas, Matplotlib.
CEO, consulting agency
"Thank you very much from the entire team for helping us."
Working on something similar?
The quickest way to find out whether I can help is a short call. Prefer email? Write to [email protected].