Knowing How Good the Data Is

Every analysis built on time records inherits their accuracy, and almost nobody measures it. These four checks take an afternoon and are worth repeating quarterly.

Recording delay

Time between the work and the entry, as a distribution.

Proportion made within one working day is the headline.

Segment by team, which shows where the process fits badly.

This is the strongest single predictor of accuracy and the one most responsive to design.

The tidiness check

Plot daily and weekly totals.

Real work produces a messy distribution. Reconstructed work produces spikes at round numbers and at the expected total.

A sharp peak at exactly the contracted hours means reconciliation, not a well-run week.

Compute the proportion of days landing exactly on a whole or half hour. A high figure is diagnostic.

This check is unusual, cheap and immediately informative, and it is the one that tells you whether to trust anything downstream.

Category discipline

Proportion of time in a catch-all or general code.

Agreement between people on the same work, tested by giving three people the same ten scenarios.

Entries whose category does not match the project, which indicates a confusing list rather than careless people.

A rising catch-all rate is the early signal that the structure has stopped fitting the work.

Completeness

Compare recorded hours against contracted hours, in aggregate.

A persistent shortfall means work is not being recorded, usually the unbilled kind.

Compare against another source where one exists: access control, calendar, the ticket system. Not to police anyone — to find out what the timesheet systematically omits.

Unbilled effort as a proportion, trended. A falling figure usually means it stopped being recorded rather than stopped happening.

Reporting quality alongside the numbers

Put these four on the same page as the analysis, so a reader knows how much weight to apply.

State the recording method. Weekly reconstruction and timer capture are not the same data and should not be presented identically.

Where a category is known to be under-recorded, say so rather than letting the figure stand as complete.

What to do about poor quality

Reduce delay first, which addresses most of it.

Fix the category list if the catch-all rate is high.

Remove friction rather than adding reminders.

Make the data visibly useful, which is the lever that costs nothing and works.

And adjust the claims you make from it. Data accurate to a wide band supports statements about shape and trend, not about individual differences, and reporting it otherwise is where the credibility goes.

Comparing against another source

A check that finds what the timesheet systematically omits, without policing anyone.

Compare aggregate recorded hours against contracted hours.

Against access control totals, where they exist.

Against calendar time for meeting-heavy roles.

The purpose is not to catch individuals and should be stated as such before anyone runs it.

The purpose is to find the shape of the omission, which is almost always the unbilled and internal work, and which then explains why project costs look better than they are.

Reporting quality with the analysis

The discipline that keeps the numbers usable.

Put the four checks on the same page as the findings: delay, tidiness, catch-all rate, completeness.

State the recording method.

Name the known omission, which is always the unbilled work.

Adjust the claims to match. Data accurate to a wide band supports statements about shape and trend, not about individual differences.

A report that states its own limits is trusted; one that does not is tested and then discounted entirely.

Check the difficult case

Use review this data-flow example to frame one representative test for this issue. The useful evidence is not the marketing description itself but the record created when a worker corrects an entry, a manager reviews it and an administrator exports it.