Client case study
27% less waste with per-flavor forecasting.
Demand forecasting The product-by-product notebook
An ice cream manufacturer replaced a one-size-fits-all forecasting approach with model selection for each flavor, reducing waste and saving planning time.
Reduction in production waste
Model architectures evaluated per flavor
Manual planning time saved monthly
01 / Context
72 flavors. 72 different demand patterns.
An ice cream manufacturer produces for a network of stores across multiple locations. Its lineup covers 72 individual flavors, and each one sells differently through the year.
Production planning relied on a single forecasting approach for the whole lineup. The result was overproduction of some flavors and underproduction of others, which meant waste on one side and lost revenue on the other.
- Industry
- Food manufacturing · Retail
- Service
- Machine learning demand forecasting
- Scope
- 72 flavors across a multi-location store network
- Deliverables
- Model pipeline · Automated selection · Forecasting API
The challenge
One forecast could not serve every flavor.
The client was overproducing and underproducing across their flavor lineup, leading to significant waste and lost revenue. A single forecasting model couldn't capture the distinct seasonality and demand curves of each flavor.
A single forecasting approach missed the distinct demand patterns of each flavor, leading to overproduction and underproduction.
Each flavor used its best-performing model, selected by backtesting against historical data.
02 / Approach
Select a forecasting model for each flavor.
Select models at the flavor level. Historical production and sales data provided the basis for backtesting five candidate architectures per flavor, rather than applying one model to different demand patterns.
Model by flavor
Built a dedicated forecasting model for each individual flavor, instead of one model for the whole lineup.
Compare five architectures
Evaluated DNN, XGBoost, LightGBM, random forest and CNN models for every flavor.
Backtest against history
Backtested every candidate against historical production and sales data.
Select and serve automatically
Automated model selection and tuning across all 72 flavors, and delivered the forecasts through a production API.
03 / Process
From historical data to a production forecasting API.
Historical production and sales feed five candidate model types per flavor. Backtesting picks the best performer for each flavor, and a Dockerized API serves the forecasts.
Candidate models compared for each flavor
- DNN
- XGBoost
- LightGBM
- Random forest
- CNN
- Historical production & sales→
- Compare candidates per flavor→
- Backtest & select→
- Serve forecasts through the API
04 / Reported results
Less production waste. Less manual planning.
Reduction in production waste
Reduction in production waste reported after introducing per-flavor forecasting.
Model architectures evaluated per flavor
Architectures evaluated per flavor, as part of model selection.
Manual planning time saved monthly
Manual production planning time recovered each month.
“We’d been guessing for years. Software Sushi built models that actually understand our flavors — strawberry behaves nothing like chocolate chip. We cut waste by 27% in the first season and finally have confidence in our production runs.”
About these results
These outcomes are reported in the published Software Sushi case study. The published account does not include the baseline dataset or measurement methodology.
- Reduction in production waste: Published case study · Results
- Model architectures evaluated per flavor: Published case study · Results
- Manual planning time saved monthly: Published case study · Results
05 / Delivery
What we delivered.
Model-comparison pipeline
Five candidate model types evaluated for every flavor.
Automated selection and tuning
Applied across 72 individual flavors.
Dockerized forecasting API
Delivered as a production-ready container.
Tools and methods
- DNN
- XGBoost
- LightGBM
- Random forest
- CNN
- Docker
Let each product choose its model.
Demand is rarely uniform across a product range. Comparing candidates per product, against the company’s own history, beat any single model applied to everything, and automation keeps that choice maintainable at 72 flavors.
A useful first conversation
Does demand vary across your product range?
Let’s discuss your forecasting process, historical data and production-planning challenges.
