Hult MSc Business Analytics coursework · Business intelligence · 2026

Marriott labour analytics and audit

A Power BI suite that joins timekeeping with scheduling to track labour spend, overtime, attendance and compliance for about 225 hourly employees at a Marriott-managed property. A follow-up audit re-implements every measure in Python and corrects the ones that were wrong.

MethodsStar-schema data modellingDAX measuresPower QueryMeasure re-implementation and auditFuzzy record matchingPseudonymisation (HMAC) ToolsPower BIDAXPower Query MPythonpandaspbixray DomainWorkforce analytics · labour cost · HR compliance

47.4% → 4.4%

overtime's share of pay, after the audit removed paid leave from hours worked

32 · 49

DAX measures and calculated columns in a 2-fact, 3-dimension star schema

225

hourly employees, most of them part-time student workers

Code, model and coded data

The problem

Hotel managers need to see labour spend and compliance risk week by week: who is heading into overtime, who is breaching the 19-hour cap for student workers, and where attendance is slipping. The raw material is two operational exports that don't share a clean key: Kronos timekeeping punches and WhenToWork planned shifts.

The dashboard

Seven pages, one per audience: department managers, supervisors, HR, finance, operations and leadership.

Department manager page: spend, hours and overtime by employee against the 19-hour line — **Figure 1.** The department manager page: fiscal-year, month and week spend, hours and overtime by employee against the 19-hour line. Names and IDs are replaced by codes. Source

Financial and overtime page: pay against overtime by week and department — **Figure 2.** The finance page: pay against overtime by week and department, and planned against actual spend. These are the original (v1) figures that the audit later corrected.

The audit

Re-running the model's logic on its own data showed the headline story was mostly an artefact.

	v1 dashboard	v2 corrected
Overtime share of pay	47.4%	4.4%
Overtime hours	19,852	1,920
Absenteeism	−119% to 20% (unstable)	7.1%
Impersonation cases	212	not measurable
Rounding "gaming"	≈ 5,000 shifts	at chance level

Why:

Paid leave was counted as hours worked, so holidays pushed people over the 40-hour overtime line. About 90% of v1's overtime was leave.
The impersonation measure counted every employee, and the punch-location fields in the export are empty, so impersonation can't be tested at all.
The overall rounding rate matched what random punch minutes produce (25.1% against 24.9%), though 30 employees sit far above it.
The name join lost 44% of planned shifts. A tiered matcher now links 69%.

What survives: Dining is still the largest overtime department, and part-timers routinely breach the 19-hour cap (73 of 196 in at least one week).

Privacy

This is a real HR export. Names and employee IDs are replaced with codes (an HMAC-SHA256 of the ID under a key that never leaves the owner's machine). The Power BI file and raw exports are not published, and the employer is named only as "a Marriott-managed property".

Limitations

One property and one export period.
A third of planned shifts still can't be matched to a timekeeping record, so no-show rates are estimates.