Artificial Intelligence & Automation
Predictive models, generative AI and agents, trained on your existing data and delivered in the dashboards your teams already use.
Learn moreSupply chain · International industrial equipment manufacturer
Predicting on-time delivery, order by order
Turning seven years of orders accumulated in the Operations datamart into a forecast of On Time Delivery for the months ahead, delivered in the dashboard the teams already use.
Scope and set-up
| Source | Operations datamart — customer orders |
|---|---|
| Volume | 5.6 million lines |
| Depth | 7 years of orders |
| Delivery format | BI dashboard, star schema |
| Delivery model | Nearshore team on a time-and-materials basis, based in Casablanca, embedded in the client's data teams and managed from France. |
Context
The on-time delivery rate (On Time Delivery, OTD) is the supply chain performance indicator tracked by end customers and tied to contractual penalties. At this international industrial equipment manufacturer, the Operations datamart consolidates seven years of customer orders.
Problem
OTD was only measured after the fact: a drift was only noticed once the delays had been suffered. The question put to SOFALI: can the years of orders accumulated in the datamart anticipate OTD for the months ahead, rather than merely record it?
Objectives
Anticipate OTD for the months ahead from the order history, order by order, and show it next to the actual indicator, in the dashboard the supply chain teams already use.
SOFALI’s intervention
Data preparation
This is the step that really made the difference. Justified, traced exclusion rules were defined with the business: order category, credit notes and returns, inconsistent dates, orders far ahead of or far behind schedule. Explanatory variables were then built from the order history: confirmed date reconstructed by cascade, delivery cycles, multi-currency conversion, amount categorisation, geographical grouping, gap in days.
Customer segmentation
RFM clustering (recency, frequency, monetary value) produced a customer behaviour score, fed back into the predictive model as an explanatory variable.
Modelling
Several formulations of the target were compared before selection. Several families of models were trained comparatively on the same training and test sets; the selection was made on the test set, independent of training, and the selected model was delivered with its limitations.
Delivery
The prediction table was integrated into the existing star schema. The dashboard puts actual OTD and predicted OTD side by side on the same time axis, with the usual business filters and configurable tolerance thresholds.
Data and technology
Python for data preparation and modelling; RFM clustering for segmentation; SQL, star-schema modelling and a BI dashboard for delivery.
Results
An OTD predictive model, order by order, whose accuracy was measured on an independent test set, delivered in the existing dashboard next to the actual indicator.
Gains identified
- Avoid late-delivery penalties by anticipating them.
- Negotiate a rescheduling with the customer ahead of time.
- Anticipate lead times in production planning.
- Optimise inventory and the load / capacity ratio.
- Make the revenue estimate more reliable.
- Keep lead times competitive.
Related expertise
BI & Data Analytics
Data warehousing, modelling and reporting: a 360° view of your business and your customers, built on solid foundations.
Learn moreQuantified results correspond to achievements measured on the scopes and datasets described; they do not constitute a performance commitment on any other scope. Offers flagged as "to be scoped" describe a proposed approach, not an existing achievement.
A similar project? Let's talk.
A first conversation to frame your needs, with no commitment, in Paris or Casablanca.