Trade-Risk Product
What the project is
I built a monthly exposure forecasting tool for LA/Long Beach (with NY as a benchmark) using U.S. Census International Trade API data. The scope is narrow by design: if vessel-linked flow tightens next month, which import categories create the most national exposure and are hardest to reroute.
Approach
I moved from concentration rankings to port-level masked next-month forecasting on Census data. The product stays split into three lenses: unusual now, next-month exposure, and shock loss. Model selection is walk-forward and leakage-safe. The active stack tests baseline-direct, seasonal-direct, stacked ensemble, and global GBM; auto selects the month’s winner from holdout performance. Feature work is constrained to trade-grounded signals: seasonality, HS4 sub-mix structure, and country-of-origin concentration.
Outcome
On the 24-month masked test window, auto was best on dollar-weighted error: +1.30% lift vs baseline (seasonal: +0.73%; global GBM: -21.82%). Operationally, auto got direction right 63.36% of the time, landed within 2 percentage points 93.43% of the time, and recovered 81.39% of true top-5 risks month to month. The result is useful for exposure ranking and uncertainty-bounded planning, not exact disruption prediction.
Live monthly updates found here: Open monthly port risk snapshot and open benchmark tracking.
Associated Downloads: model backtest summary and snapshot data.