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.

• One header row with descriptive column names helps most. Title lines above the table, totals rows below it and empty columns are skipped for you, and the tool says so.
• Dates may be written in many ways. Day/month versus month/day is settled from the data where possible and stated when it is a guess; you can flip it with one tap.
• Units written inside cells or headers (“1.2k”, “€ 1.234,50”, “Delay (days)”, “Temp °C”) are read and shown beside the numbers. Keep one currency per column.
• Yes/No columns become rates, such as the share of shipments that were on time. A missing value stays missing; it is never turned into zero.

CSV, TSV, Excel, JSON, JSONL, XML and YAML are supported, by dropping the file anywhere on the page, choosing it, or pasting a table. The 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 how it was read before trusting a headline.

The title bar states how many rows and columns were found, the period covered and its grain (hourly, daily, weekly, monthly). The Data view lists every column with the role it was given (time, measure, category, identifier), and the notes say what was cleaned.

A quality grade summarizes completeness, parsing and type consistency. It cannot tell you whether the source system recorded the right business event. Repeat purchases and repeated dates are not treated as duplicate errors.

An unusual value is judged against what ordinary variation would produce for that kind of data; a quiet series and a busy one are not held to the same bar. Investigate the corresponding source record: it may be a mistake, a valid large order or a change in the business.

3. Separate a pattern from an explanation.

The Overview shows the key numbers with their change since the previous period, then the findings ranked by how clear and how large they are. Each finding opens into the facts behind it. Trends lets you pick any key number; a period that is still in progress is drawn dashed so a half-finished month is not mistaken for a drop. Breakdown splits a number by category, Connections shows what moves together, and Data shows the columns and a preview.

Findings must clear a significance and an effect-size test before they appear; on data with no pattern in it, almost nothing is reported, and “Steady and unremarkable” is a valid result. Correlation describes how paired values vary together. It does not establish why they move, and the tool never words it as cause.

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 periods or a stale endpoint can suppress the estimate.

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. A forecast is shown only when it beats that baseline on this historical check. Even then it does not guarantee the next value; the uncertainty band describes past one-step errors, not a promise of future coverage.

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

5. Save the result with its context.

Use Export for an HTML report with all charts, a PDF (through the browser’s print dialog), an Excel workbook of the findings and the tidied data, a CSV of the tidied data, or JSON of everything found. Any chart can also be saved as an image. Keep the source period, column definitions and any notes with the report so another reader can understand it.

Your source datasets and full charts are not saved automatically. A free account lets you preview and explicitly save dashboard summaries, findings, notes and KPI targets. Use Save snapshot to keep a summary, or Export for the full report 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.

6. Go further with specific questions.

Once the file looks right, three short guides cover the questions people ask most, each with a working example.

• How to tell whether two columns are related — correlation, strength and the usual traps
• How to compare groups — does a category really change a number?
• Breakdowns and where a total comes from — shares, flows and drivers