Manufacturing & Logistics
Data & MLOps Foundations
Ferrovia Logistics
Forecasts that say when
they are losing accuracy.
Forecast error (MAPE)
16 weeks
12.5%
Was
18.9%
−6.4 pts
−22%
Safety-stock cost
22
Warehouses
16 wks
Kickoff to production
02 · The engagement
Published
01 · Challenge
What was actually broken
Not the symptom, the cause.
Ferrovia’s demand-forecasting models were accurate at launch and steadily degraded over the following months, forcing planners to fall back on manual overrides and wider safety-stock buffers to cover the gap. Nobody owned monitoring, so degradation was only caught when a stockout or an overstock report surfaced it.
02 · Approach
How we scoped it
And what we deliberately chose not to build first.
Rather than retrain the existing models and repeat the same failure mode, we built the foundations that were missing: a versioned feature store shared across all 22 distribution centers, an evaluation pipeline that runs against held-out data on a fixed schedule, and drift monitoring that alerts the planning team before forecast error crosses a defined threshold — not after a stockout report does.
03 · Architecture
The system, in one diagram
What it connects to, what a person still reviews, where the guardrails sit.
Solution architecture
Data & MLOps
22 warehouses
Sales, promos, stock
Shared, versioned
Feature store
Demand by SKU
Per-centre models
Triggers retraining
Drift threshold
Pass · Confidence per SKU
Planner dashboard
Else · Prior model version
Instant rollback
Drift monitored per centre · retrain on thresholds, not the calendar
A central feature store feeds per-center forecasting models, with automated retraining triggered by drift thresholds rather than a fixed calendar. Deployment tooling supports instant rollback to the prior model version, and a monitoring dashboard gives planners visibility into forecast confidence per SKU per center, not just an aggregate accuracy number.
04 · Result
What changed
Measured in production, not in the pilot.
Sixteen weeks after kickoff, forecast accuracy across the network improved by 6.4 points of MAPE, and safety-stock levels came down as planners regained confidence in the forecast instead of padding around it.
Result at a glance
Client
Ferrovia Logistics
Industry
Manufacturing & Logistics
Service
Data & MLOps Foundations
Time to production
16 weeks
Headline
22%
reduction in safety-stock carrying costs
03 · What changed
In the client’s own operating numbers
22%
reduction in safety-stock carrying costs
6.4 pts
improvement in forecast accuracy (MAPE)
16 weeks
from kickoff to production
−6.4 pts
Forecast error (MAPE)
18.9%
→
12.5%
04 · In their words
Operators, not buyers
We'd been burned before by a model that was great in the pilot review and quietly wrong six months later. What Northwind built is the first forecasting system we've had that tells us when it's losing accuracy, instead of us finding out from an empty shelf.
Tom Baptiste, VP of Supply Chain, Ferrovia Logistics
Next case study
Let’s find out
what it’s worth in production.
Have a workflow like this one? A 30-minute call is enough to tell whether it is ready to build.