Case studies About us Results How we work Talk to Us

THE WORK / DATA & ANALYTICS

Data that explains.
Decisions that
move you forward.

Real challenges. Connected thinking. Clear outcomes.

Four connected investigations for one enterprise-scale US telecom business. Follow the signals, the breakthroughs and the thinking behind them.

Discover the work ↘
Journey analyticsBigQuery · ARIMA_PLUS
QLIKMATRIX / ANALYTICS●
FROM SIGNAL TO STRATEGY ↗
3.53%

Held-out forecast error · 10 weeks

Model visualizationIllustrative ↗
✓
One business. Four breakthroughs.Diagnose · Validate · Understand · Predict
THE SCALE
OF THE QUESTIONEnterprise.
Every day.
70,000+daily funnel sessions
115weeks of channel data
700,000+weekly sessions in forecast scope

An anonymized top-tier US carrier · Direct-to-consumer shopping experience.

OUR WORK, UNDER THE MICROSCOPE

Real questions.
Rigorous answers.

Explore each question, the evidence behind it
and what the investigation made possible.

01
Journey diagnostics45%purchase drop investigated
↗
01 / CHECKOUT DIAGNOSTICS

A 45% purchase drop.
A fault in the journey.

01 The challenge

Completed device purchases fell 45% against their normal trend. Was it lower traffic, a site defect or friction in the buying journey?

02 How we investigated

Mapped product-to-purchase paths and compared behavior before and during the issue. Repeat payment visits rose from 19% to 41%, pointing to the payment-processing step.

THE RESOLUTION

The web team traced a payment-confirmation bug. Purchases returned to normal; two Power BI funnel measures enabled daily monitoring.

WHERE SESSIONS WENT AFTER CHECKOUT

Before issue During issue
Purchase complete
51%
34.3%
Back to checkout
21.3%
46.9%
Financing sign-up
20.9%
12.4%

Reported session shares. Source comparison: 9,712 pre-issue hits and 7,200 during-issue hits. Selected destinations shown.

02
Measurement audit28.2% → 1.9%reported conversion decline
↗
02 / MEASUREMENT INTEGRITY

The dashboard crashed.
The business hadn’t.

01 The challenge

Reported checkout-to-purchase conversion fell from roughly 30% to 2% in two days, across 70,000–83,000 daily sessions.

02 The diagnostic clue

Journeys showed payment without checkout, or checkout without a product view. Several funnel steps declined together. A GTM audit found Page View tags firing on manual refresh but failing during normal navigation.

THE REPORTED OUTCOME

The team identified a data-quality issue and avoided unnecessary marketing or pricing changes in response to a false alarm.

NORMAL DAY → WORST DAY

May 28 vs May 30

Device to detail61.1%→13.6%
Shopping to checkout42.3%→7.1%
Checkout processing to purchase88.4%→1.0%
Checkout to purchase28.2%→1.9%

Detailed source-chart rates; headline uses rounded narrative figures. Reported dip: May 29–June 9.

!

Validate the measurement
before changing the business.

03
Spend & traffic analysis115weeks of channel evidence
↗
03 / CHANNEL ANALYSIS

Channel signals.
Better budget questions.

01 The challenge

Across 17 channels and 115 weeks, weekly spend was compared with website visits at channel and sub-channel level.

02 What the evidence can tell us

Shopping / PMAX / Demand Gen showed the strongest reported positive association among the displayed channels. TV and Display also had positive associations.

DECISION USE

Prioritize further testing of the Search Group; assess TV against awareness goals and Non-Branded Search against efficiency and revenue.

Correlation does not establish incremental impact. Shared spend patterns can confound results. Branded Search was excluded by the source because its negative reading was considered inconsistent with branded intent.

CORRELATION WITH WEEKLY WEBSITE VISITS

Shopping / PMAX / Demand Gen
+0.448
TV
+0.302
Display
+0.275
Audio
+0.057
AARP
+0.027
Email
-0.061
Digital Video
-0.139
Non-Branded Search
-0.252

Bar length shows association strength; signed values show direction. These are associations, not causal effects.

04
ARIMA_PLUS / BigQuery3.53%held-out forecast MAPE
↗
04 / PREDICTIVE MODELING

Build a baseline.
Spot the exceptions.

ARIMA_PLUS ran inside BigQuery. Weekly forecasts proved more stable than direct daily forecasts and were allocated using typical daily traffic shares.

3.53%

Mean absolute percentage error
across 10 held-out weeks

THE SELECTED INPUTS

Search Group spend + adjusted TV. A consent-rate input was later added to account for changes in measured traffic.

The source also describes a quarterly forecast layer. Forecast error is validation performance, not a guarantee of future accuracy.

FROM INPUTS TO ANOMALY REVIEW

  1. 01
    Historical signals

    Traffic, Search Group spend and adjusted TV spend.

  2. 02
    Weekly forecast

    Recurring patterns and holiday effects in BigQuery.

  3. 03
    Daily allocation

    Weekly predictions split by usual daily traffic shares.

  4. 04
    Anomaly review

    Observed traffic compared with the baseline.

TV CARRYOVER FEATURE

(Current week TV × 0.90)
+ (Prior week TV × 0.40)

THE MODELING STUDIO

Many signals.
One connected model.

Explore three tested input combinations.
The validation score updates with your selection.

BIGQUERY / ARIMA_PLUS10 HELD-OUT WEEKS
01 / INPUT SIGNALS
◷
Historical trafficWeekly session patterns
✓
▤
Adjusted TVCurrent + prior-week spend
✓
02 / THE ENGINEARIMA_PLUSPatterns. Seasonality. Signals.BigQuery
03 / VALIDATION3.53%

Mean absolute
percentage error

Weekly forecast↓Daily allocation
● Model inputs from the source study

Lower MAPE means lower error on the held-out weeks. It does not guarantee future accuracy.

MODEL VALIDATION

Built to predict.
Designed to adapt.

All 11 predictor combinations.
Lower MAPE is better.

THE MODEL LEADERBOARDMAPE ↓
Search Group + adjusted TV
3.53%
Adjusted TV
4.23%
Search Group + adjusted TV + Display
4.26%
TV spend
4.45%
Search Group
4.62%
Search Group + adjusted TV + rest
4.8%
Display
5.49%
SBPA group
5.5%
ASSMBP group
5.66%
Display + PSMDT + ASSMBP + SBPA
5.69%
PSMDT group
6.9%

BRING US THE QUESTION

Your data has
a story.
Let’s find it.

A sales drop, a dashboard anomaly, unclear channel value or an unexpected traffic shift. Let’s work out what the data is telling you.

Talk to our analytics team ↗cs@qlikmatrix.com
About the evidence

Source: QlikMatrix Data Analytics Case Study, redesigned portfolio (2026), supplied by QlikMatrix. All four investigations concern the same anonymized enterprise-scale US telecom business. Figures and outcomes are reproduced from that document. Revenue uplift and a quantified post-fix recovery period were not supplied. Channel correlations are not causal estimates. Model-group abbreviations follow the source.