How it works
From a strait closing to a customer email in minutes.
No black box. Every number on the screen can be traced back to the signal that moved it.
The model
Six steps, in plain English
Signals
Vessel tracking, insurance, freight indices, notices, news
Chokepoint intensity
Scored per passage, 0–100
Lane delay
Summed across the chokepoints a lane crosses
Supplier score
Delay against buffer, weighted by single-sourcing
Order impact
Revised ETA per purchase order
Action
Ranked mitigations and a drafted message
Scoring
The risk score, explained
- 1
Expected delay is the sum of disruption at every chokepoint the lane crosses. A lane through two contested passages carries both.
- 2
That delay is compared against the stock buffer you hold for the part. Fourteen days of delay against thirty days of cover is an inconvenience; against seven days it is a stoppage.
- 3
The result is weighted by how single-sourced the part is. A part with three qualified suppliers absorbs a delay that a sole-source part cannot.
- 4
The output lands in one of three tiers: Stable, Elevated or Critical — with the contributing factors shown alongside, never the tier on its own.
Diversion logic
When a risk becomes a certainty
While carriers still use a route, disruption is probabilistic: some sailings are delayed, some are not. Once carriers stop using it, the added transit of the alternative applies to everything on that lane.
We model that switch explicitly, because the useful window is before it happens. Capacity on the alternative is cheapest and most available while everyone else is still waiting for an announcement.
Humans in the loop
Analysts review what matters
High-impact signals are reviewed by analysts before they move a forecast. Classification by model alone is fast and occasionally confidently wrong, which is an unhelpful combination when someone is deciding whether to air-freight a container.
You can override any chokepoint intensity yourself. Your override is recorded in the audit log alongside the model’s figure, so the disagreement is visible rather than silent.
Accuracy
How well it actually forecasts
Measured against a held-out test set, compared with the obvious alternative: predicting the historical median delay for the same lane.
| Measure | Baseline | Model | What it means |
|---|---|---|---|
| Mean absolute error | 4.6 days | 2.8 days | Average size of forecast error |
| Median absolute error | 3.4 days | 1.9 days | Typical error, less influenced by extreme cases |
| Forecasts within ±3 days | 48% | 71% | Share close to the actual delay |
| Mean signed error | −1.1 days | −0.3 days | Negative values indicate underprediction |
| Coverage of 80% prediction interval | — | 81% | Actual delay falls within the range about four times in five |
| Mean interval width | — | 7.2 days | How broad the forecast ranges are |
- Test shipments
- 1,200
- Test period
- May – 28 September 2026
- Completed shipments available
- 4,800
- Cohort
- 60 production organisations
Baseline: Historical median delay for the same lane, from earlier records. Model: Gradient-boosted regression over lane, normal transit time, disruption indicators, congestion and diversion information. Uncertainty: Quantile predictions giving an 80% prediction interval.
Trained on June 2025 – February 2026, tuned on March – April 2026, and measured on a held-out test set from May – 28 September 2026 that the model never saw during training. Excluded from the evaluation: test accounts, duplicate shipments, missing arrival dates and cancelled shipments.
Two numbers there matter more than the headline. The 81% interval coverage against a nominal 80% means the stated range is honest: when the tool says it does not know precisely, it is right about not knowing. And the mean signed error of −0.3 days shows a slight remaining tendency to underpredict — worth knowing, since underprediction is the direction that costs you.
Three forecasts, followed through
Individual predictions from the test set, with what actually happened.
| Shipment | Organisation | Forecast issued | Predicted | 80% interval | Actual | Error |
|---|---|---|---|---|---|---|
| VAL-001 | Alderbrook Trading | 10 May 2026 | 8 days | 5–12 days | 9 days | 1 day |
| VAL-002 | Bluehaven Components | 18 June 2026 | 4 days | 1–7 days | 6 days | 2 days |
| VAL-003 | Cedarfield Supplies | 7 July 2026 | 10 days | 6–15 days | 17 days | 7 days · outside interval |
VAL-003 fell outside its stated range. At 81% coverage roughly one forecast in five is expected to, which is what a prediction interval means. Showing only the ones that landed would make the interval look better than it is.
Onboarding
Live in three weeks
Data import
Supplier list, lanes, transit times, stock buffers and open purchase orders. A spreadsheet export is enough.
Lane mapping and calibration
We map each lane to the chokepoints it crosses and calibrate expected transit against your own arrival history.
Alerts live
Scores go live, thresholds are tuned to your tolerance, and alerts start reaching the people who act on them.
See it against your own lanes
We load a sample of your suppliers, run a live scenario, and show which orders would slip.