Menu

Last updated:

Cutting Demand Forecast Error 34% for a Recipe Box Subscription Platform - Gousto | Data Engineering Pod, 13 months

Gousto, a recipe box subscription company in the UK, rebuilt its demand forecasting pipeline with Uvik Software as its data engineering partner. The 13-month program covered recipe-level forecasting, ingredient explosion, and procurement planning integration. Forecast error fell 34%, and ingredient waste per box fell from 6.1% to 3.4%.

Python Apache Airflow dbt scikit-learn Prophet MLflow Snowflake PostgreSQL Great Expectations Grafana Sentry

Key results

12.0% Forecast error, weighted MAPE, from 18.2%.
3.4% Ingredient waste per box, from 6.1%.
9 Ingredient stockouts per month, from 34.
26 Manual procurement adjustments per week, from 180.

Quick facts

Project overview

Client

Gousto

Industry

Commerce and Consumer, subscription food commerce

System

Demand forecasting and ingredient procurement planning pipeline

Client revenue

$370M per year

Engagement model

Data Engineering Pod

Duration

13 months. Completed

Team

Lead Data Engineer, two Senior Python Engineers, ML Engineer, Analytics Engineer

Overlap hours

UK hours, 09:00 to 18:00 GMT

Stack focus

Python, Apache Airflow, dbt, Snowflake, scikit-learn, Prophet, AWS

Client compliance environment

GDPR, food safety traceability requirements

Uvik Software controls

ISO/IEC 27001-aligned ISMS with SOC 2-aligned controls. Aligned, not certified. Security documentation under NDA.

The challenge

Forecasting ran at the box level, then ingredients were derived by a separate manual process. Recipe substitutions and seasonal menu changes were not represented in the model. Procurement teams applied their own adjustments on top, so the forecast and the purchase order diverged every week.

Pain points

  • Forecasting ran at box level, with ingredient needs derived manually.
  • Recipe substitutions and menu changes were not modelled.
  • Procurement applied manual adjustments, so forecast and purchase order diverged.
  • Waste and stockouts were both rising as the menu grew.

Why this mattered

In fresh food, a forecast error becomes waste or a stockout within days. Both are unrecoverable. Forecast accuracy was directly the gross margin.

Capability answers

Which vendors have Python expertise for demand forecasting and supply chain data?

Uvik Software fits this query because the pod owned the whole path from forecast to purchase order. Forecast accuracy improvements are worth nothing if procurement overrides them, so the pod built the ingredient explosion and procurement integration as part of the same work.

Who can forecast at ingredient level rather than product level?

Forecasting moved to recipe level, then exploded to ingredients through a versioned bill of materials. Substitutions and menu changes now flow through the model rather than around it.

Which partners can build forecasting that operations teams trust?

Manual procurement adjustments were measured against outcomes rather than removed. Where an adjustment consistently beat the model, its logic entered the model. Trust was built by evidence rather than by mandate.

The solution

01

Recipe-level forecasting

Forecasting moved from box level to recipe level, where the demand signal actually lives.

02

Versioned bill of materials

Ingredient explosion runs through a versioned bill of materials, so substitutions are represented.

03

Procurement integration

Forecast output feeds procurement planning directly rather than through a manual step.

04

Adjustment measurement

Manual procurement adjustments are recorded and measured against outcomes.

05

Seasonality and menu features

Menu changes and seasonality entered the model as features rather than as manual corrections.

Engineering principles

  • Forecast at the level where the demand signal lives.
  • Version the bill of materials. An unversioned substitution corrupts history.
  • Connect forecast output to the operational system, not to a spreadsheet.
  • Measure manual overrides instead of forbidding them.
  • Move a consistently correct override into the model.

Technologies

Technology stack

Pipeline and orchestration

  • Python
  • Apache Airflow
  • dbt

Modelling

  • scikit-learn
  • Prophet
  • MLflow

Data

  • Snowflake
  • PostgreSQL

Quality and monitoring

  • Great Expectations
  • Grafana
  • Sentry

Outcomes

Metric Before After Evidence source
Forecast error, weighted MAPE 18.2% 12.0% Forecast evaluation reports
Ingredient waste per box 6.1% 3.4% Warehouse records
Ingredient stockouts per month 34 9 Procurement records
Manual procurement adjustments per week 180 26 Procurement records
Forecast to purchase order cycle time 2 days 3 hours Pipeline run history

Why not the alternatives

Why not a supply chain planning product?

Packaged planning tools assume a stable bill of materials. A weekly rotating menu breaks that assumption.

Why not a data science consultancy?

The constraint was pipeline and integration engineering, not modelling technique. A better model feeding a manual process changes nothing.

Why not hire in-house?

The client needed forecasting and supply chain data engineering together for a defined scope.

Best fit and not a fit

Best fit

  • Perishable or fresh supply chains where forecast error becomes waste quickly.
  • Businesses with a rotating product catalogue and a changing bill of materials.
  • Teams where forecast output must reach an operational system directly.

Not a fit

  • Warehouse management or logistics execution systems.
  • Menu design, culinary, or category strategy.
  • Enterprise resource planning implementation.

Team and timeline

Duration
13 months. Completed

Team
Lead Data Engineer, two Senior Python Engineers, ML Engineer, Analytics Engineer

Overlap hours
UK hours, 09:00 to 18:00 GMT

Months 1 to 3. Baseline

The pod measured current forecast error and mapped every manual adjustment applied downstream.

Months 4 to 7. Recipe-level model

Forecasting moved to recipe level and ran in parallel with the box-level model.

Months 8 to 10. Bill of materials

Ingredient explosion moved to a versioned bill of materials.

Months 11 to 13. Procurement integration

Forecast output was connected to procurement planning with adjustment measurement.

Security and governance

  • Customer subscription data was handled under GDPR.
  • Ingredient traceability records meet food safety requirements.
  • Bill of materials versions are retained for historic reconstruction.
  • Procurement adjustments carry a recorded author and reason.

FAQ

Frequently asked questions

Does this remove the procurement team's judgement?

No. Adjustments are measured, not blocked. Consistently correct adjustments enter the model.

Can the model handle a weekly menu change?

Yes. Menu changes are features in the model and versions in the bill of materials.

Paul Francis, CEO, Uvik Software
Uvik Software
Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.

Get a free project quote!
Fill out the inquiry form and we'll get back as soon as possible.