A useful report starts with a well-understood file.

1. Know what one row represents.

Before opening a file, identify its unit of observation: one order, one customer, one day or one monthly total. Mixing daily records with monthly totals can make summaries misleading even when every calculation is correct.

• Use a single header row with descriptive column names.
• Keep dates in a consistent format, preferably an unambiguous format such as YYYY-MM-DD.
• Keep units consistent within a column. Separate currencies or convert them using a documented method before comparing them.
• Leave missing observations distinguishable from genuine zeros. Zero sales and an unrecorded day are different facts.

CSV, TSV, Excel, JSON, JSONL, XML and YAML are supported. The current limits are 5 files per session, 20 MiB per file and 50 MiB in total. Analysis accepts up to 50,000 rows, 128 columns and 2,000,000 parsed cells per file. Larger row sets produce a notice; oversized files or grid shapes are rejected.

2. Check the interpretation before the headline number.

Open the file, read the overview and inspect the source preview. Use Configure to choose the time axis, confirm date conventions and exclude columns that do not belong in the analysis.

A quality grade summarizes completeness, parsing and type consistency. It cannot tell you whether the source system recorded the right business event. Declare a unique key only when a column really should be unique; repeat purchases and repeated dates are not automatically duplicate errors.

An outlier is a value outside the tool’s quartile-based fences. Investigate the corresponding source record: it might be a mistake, a valid large order or a change in the business. The label alone does not tell you which.

3. Separate a pattern from an explanation.

The analyst brief summarizes the file. The signal rows let you compare values and small trend charts; selecting a signal opens its column profile. Chart controls let you choose the axes and metrics rather than relying on a single default view.

Confirm the date range and units before comparing changes. A percentage change from a zero baseline is undefined, even when the direction of movement is clear. Percentage-valued inputs retain their percentage-point units.

Correlation describes how paired numeric observations vary together. It does not establish why they move, and a result based on few pairs is weak evidence. Constant columns and insufficient paired observations are shown as unavailable, not as a fabricated zero relationship.

4. Ask whether a simple baseline would do better.

KPI Master uses Holt smoothing with a median error correction for a one-period estimate. It requires a usable time axis with regular, unambiguous observations. Duplicate or irregular dates, missing metric periods or a stale endpoint can suppress the estimate; the tool explains the reason.

With at least 19 observations, the forecast finding compares up to 12 recent one-step predictions with a simple baseline: repeat the previous value. Each prediction uses only earlier observations and earlier errors. The reported mean absolute errors are in the metric’s own units. Lower is better on that historical check.

If the last-value baseline did better, treat the model estimate cautiously. Even a lower historical error does not guarantee the next value. The uncertainty band describes past one-step errors; it is not a promise of future coverage. Short histories do not support the same checks as long ones.

For background on held-out forecasts and baseline comparisons, see Forecasting: Principles and Practice.

5. Save the result with its context.

Use Export report to choose the focused dataset or all datasets, then download a workbook, data export or standalone HTML report. PDF uses the browser’s print dialog. Keep the source period, column definitions and any exclusions with the report so another reader can understand it.

Your datasets and results are not saved automatically. Export before closing or reloading. Installing the app makes it easier to reopen the tool; the portable offline edition contains the tool and its examples, not a backup of your current work.