The Taste:
Whether batch 4,817 still tastes like batch 1
A supplier change for cocoa, a small adjustment in conching, a different milk powder batch - and after eighteen months it is a different product. Every batch is within tolerance. The trend is not. The tasting panel read as a time series shows it.
This module is part of a nine-part model calculation on a fictitious mid-sized chocolate manufacturer (€95 million net revenue, 140 items, 4 lines). All euro amounts are derived from the documented assumptions of the sample company. They are not results of real customers and not a performance promise. Which values are achievable on your data is shown only by a pilot.
01 The problem - every batch is fine, the trend is not
People see points, a model sees the line
A trained tasting panel rates important items per batch on scales: cocoa note, sweetness, creaminess, roast note, texture. Every rating lies within the norm. But if the cocoa note drops by a tenth of a point per quarter, after six quarters the item is outside what consumers know - and nobody has measured it.
The retailer reports it as a complaint, consumers as "tastes different from before". Both come late. The panel delivers the same signal months earlier if its data is treated as a time series instead of a release record.
Sensory analysis finds drift. Foreign bodies and microbiological findings remain a matter for quality assurance.
02 The model - outliers and drift considered separately
Panel mean per batch, adjusted for assessor effects via reference samples; outliers by isolation forest, drift by cumulative sum with thresholds fixed beforehand
Batches as outliers: 4 of 180 Drift alarm cocoa note: month 14 (backtest, example values)
Assessors unconsciously shift their scale. The standard for selection, training and monitoring of sensory assessors therefore requires reference samples and regular performance checks. The model uses exactly these reference samples to separate assessor drift from product drift. Without this adjustment the model would alarm on the assessor, not the product.
03 Business impact - early correction instead of late rework
Two levers, both assumptions of the sample company
Every assumption comes from the sample company and is stored centrally. With your figures the input changes, not the method.
| Position | Value |
|---|---|
| Sensory complaints in retail / year | €80,000 |
| Lever: Fewer complaints through early correction (25 %) | €20,000 |
| Cost of one rework per batch | €120,000 |
| Lever: Avoided rework: one case every two years | €60,000 |
| Result: savings / year | €80,000 |
All amounts of this module are assumptions of a sample company without a published anchor.
Almost every manufacturer has a panel. Hardly any treats its data as a time series. The step from release to monitoring costs no new sensor technology, only a different look at existing records. The amounts are assumptions; whether at your company a rework is avoided every two years or every five is known only to your complaint files.
04 Next steps in your factory
From release record to time series
The ratings per batch and assessor are transferred from the records into a table - including the reference samples per session.
For the three most important items the time series per attribute is built. You see whether drift occurred and when the model would have reported it.
A monthly report per item: attribute, direction, strength of the drift, suspicious raw material batch - as a template for the recipe review.