Contents
In short
- 48 methods, each with a spec sheet: required data, formula, assumptions, limitations and minimum data.
- 480 automated tests, run before every build on Linux and on Windows. A build is not released if a single test fails.
- 48 reference answers — one per method — calculated independently on a fictional training warehouse.
- Blind test on 137 public datasets from 20 countries (214 files): 0 false results, 0 crashes, 18 of 18 figures matched an independent recalculation.
- No AI in the calculations. Every result comes from a fixed, documented formula. The same data and parameters always give the same result.
Where the formulas come from
The methods implement the standard, widely used definitions of warehouse and supply chain practice — the ones taught in supply chain courses and used in WMS and ERP reports. Examples:
- ABC analysis — the Pareto principle: items sorted by their share of value, classes by cumulative share (80 % / 15 % / 5 % by default, adjustable).
- XYZ analysis — coefficient of variation of monthly demand,
CV = σ / μ. - Reorder point —
ROP = d·L + SS: demand over the lead time plus safety stock. - OTIF — share of orders delivered on time and in full; fill rate — share of the ordered quantity shipped.
- Control charts — Shewhart p-charts with control limits at three standard deviations.
- Forecast accuracy — MAE, MAPE, WAPE and bias against actual demand.
Each method’s exact formula, assumptions and limitations are shown on its Methodology tab in the program and on the Methodology sheet of every report. LOQMETRA does not claim certification by any standards body; it makes every calculation visible so that you can check it.
Rules that protect the result
- A method runs only if your data really supports it. If a table or field is missing, it says which one — it never fills gaps with invented values.
- Generic column names (“Quantity”, “Date”, “Value”) are never interpreted silently: you confirm their meaning.
- No row is removed or changed without your click; every exclusion is listed in the report.
- Missing months are not treated as zero demand unless the data confirms it; short histories are reported as a limitation.
- The report records the program version, method versions, parameters, field mapping and exclusions, so the result can be reproduced.
Reference answers on a training warehouse
A separate generator creates a fictional warehouse with two years of history in 23 tables: items, suppliers, customer orders, shipments, receipts, inspections, stock, cycle counts, operations, shifts, costs, returns and more. The generator does not use any code of the program.
A second, independent script calculates the expected answer of every one of the 48 methods directly from that data. The automated tests then compare the program’s results with these reference answers — figure by figure, and table by table where a method returns a table.
Blind test on real public data
Neat test data is not enough, so the program is also run on data it has never seen: 137 public datasets from 20 countries — 214 files from Kaggle, the open-data portals of the United Kingdom, Canada, Australia, Ireland and Quebec, the EU data portal (France, Germany, Austria, Italy, Greece, Poland, Slovakia, Spain, the Netherlands), Zenodo, Harvard Dataverse, the Northwind sample database and Microsoft samples.
Before each run it is written down which methods can honestly be calculated from each dataset. Results of the latest round (October 2026):
- 133 of 150 required methods completed (89 %) when the user follows the program’s suggestions.
- 0 false results on 87 tables that are not warehouse data — the program correctly refused to calculate.
- 0 crashes on all 214 files.
- 18 of 18 checked figures (ABC and XYZ classes, demand history, supplier lead time and on-time rate, freight cost, order lines) matched an independent recalculation from the source files.
Every build
- The full test suite runs on Windows on GitHub’s servers; the program is built only if every test passes.
- The published ZIP is checked: the file downloaded from the public link must match the build byte for byte (SHA-256 checksum, published next to the download).
- Before release the program is also started on a real Windows computer and the full analysis of the training warehouse is compared with the reference figures.
What testing cannot promise
Tests show that the formulas are implemented correctly and that the program behaves honestly on a wide range of files. They cannot guarantee that every export from every system is read the way you mean it. That is why you confirm column meanings, why limitations are always stated — and why you should check important results before acting on them (see Terms of use).
Found a number that looks wrong? Write to support@loqmetra.com — every confirmed error becomes a new test.