Skip to main content
Loading Capital Risk portfolio data…
NYC DDC
DDC's CIO asked whether a model trained on the agency's 10+ years of structured capital-project records — timelines, milestones, contractor performance, change orders — could flag early warning signs of delay or cost overrun across DDC's 600+ active capital projects, since project managers currently rely on experience alone. The desired outcomes were a validated model that flags at-risk projects early, a ranked list of the most predictive variables, and a prototype risk dashboard for project managers.
Sponsor: Zeeshan Anwar, CIO · NYC DDC
Capital Project Risk is a public, read-only risk explorer built on the public shadow of DDC's internal records: original-versus-current milestone dates, budget trajectories, linked Checkbook NYC contracts, and labeled historical delay reasons. A trained, published model scores 1,323 of the 1,665 stitched active DDC projects with delay and cost-overrun probabilities and plain-language per-project risk factors; the other 342 projects display an honest unscored state with a reason code instead of a fabricated number. The portfolio ranks and filters by borough, phase, and category, and drills into each project's milestone and budget history alongside its contracts. Every headline metric, calibration bin, and label definition is published on the model card, so a skeptic can audit each number rather than take it on faith.
Ingest scripts land all seven public sources into Neon Postgres (`nyc.capital_risk_*`), and a stitching pass links each project across datasets — 66.2% legacy-history coverage and 10.7% Checkbook coverage. A Python training stage (LightGBM gradient-boosted classifiers plus a Cox proportional-hazards survival model, temporal holdout validation) publishes runs back into Postgres; the live pages read the latest published run. The current published run (v1.0.0, trained July 2026) reports delay PR-AUC 0.692 against a 0.658 per-category historical base-rate baseline, cost-overrun PR-AUC 0.176 against 0.133, and Cox time-to-completion concordance 0.608 — modest, honestly-reported lifts on public data alone.
Live verification rated this product TRUSTWORTHY: 8 criteria passed outright, 4 were cleanly gated pending a model run (a published run has since landed and now feeds the live pages), 1 minor warning was noted, and zero fabrication was found.