01 // BUSINESS PROBLEM · PROFITGUARD

ProfitGuard

Where is an e-commerce business losing profit?

ProfitGuard connects revenue, refunds, delivery, customer, product, payment and marketing data to find where profit is leaking.

FOLLOW THE DATA
A parcel moving through a bright e-commerce fulfillment centre.
FULFILMENT SIGNALORDER → DELIVERY → REFUND RISK
Illustrative fulfillment path · project data is synthetic
10,000ORDERS
£6.58MNET REVENUE
8.81%REFUND RATE
08DATASETS

02 // WHAT I FOUND

Why are refunds
happening?

In the project data, refund rates were higher for late-delivery orders. That association made delivery performance the clearest place to investigate.

01 / ON TIMEDelivery completed as expected
5.83%REFUND RATE
02 / LATEDelay flags appear along the route
24.97%REFUND RATE
RATE GAP+19.14ppDESCRIPTIVE DIFFERENCE
STATISTICAL TESTTwo-proportion z-test
SIGNIFICANCEp < 0.001

Late delivery was associated with higher refund risk in this dataset.

A parcel delivered at the entrance of a home.
ON-TIME DELIVERYArrives as expectedILLUSTRATIVE SCENE
A parcel waiting at a busy sorting facility.
DELIVERY DELAYA moment to interveneILLUSTRATIVE SCENE

03 // WHAT I BUILT

Eight datasets.
One connected workflow.

From cloud data to practical refund-risk recommendations.

  1. 018 DATASETSCONNECTED INPUTS
  2. 02AWSS3 · GLUE · ATHENA
  3. 03SQL / PYTHONDATA ANALYSIS
  4. 04STATISTICSHYPOTHESIS TESTING
  5. 05REFUND-RISK MLLOGISTIC REGRESSION
  6. 06POWER BIBUSINESS REPORT

03 // REAL POWER BI · 5 REPORT PAGES

From analysis
to a shared view.

These are the project’s real report screenshots. Select a page to bring it forward, or enlarge it to read the details.

REPORT PAGE 01 / 05

Start with the business picture.

See order volume, net revenue, refunds and the headline operating measures together.

Use ← → keys or choose a report page
PROFITGUARD / POWER BI01 / 05
ProfitGuard Power BI Executive Overview with headline order, revenue and refund metrics.
Executive Overview

Actual ProfitGuard Power BI exports · synthetic project data · dashboard screenshots are not mockups.

· PROJECT SCREENSHOT

Use browser zoom or pinch to inspect fine details.

INSPECT ANALYSIS Technical details, checks and supporting screenshots
METHODS AT A GLANCE

Amazon S3 · AWS Glue · Athena · SQL · Python · Logistic Regression · hypothesis testing · historical discount comparison.

THE HANDOFF

Eight datasets become queryable tables, then statistical and machine-learning analysis feeds the Power BI report and recommendations.

03 / DATA QUALITY
01───08

Keep the record.
Flag the anomaly.

29 orders (0.29%) had discounts greater than gross order value. Those rows were retained, flagged, and their net order value was floored at zero for downstream analysis.

29ORDERS FLAGGED0.29% OF ORDERS

04 // HISTORICAL CONTROL VS TREATMENT

Did the larger discount
raise order value?

This is an A/B-style analysis of historical orders, not a randomized experiment.

CONTROL · LOWER DISCOUNT TREATMENT · HIGHER DISCOUNT
CONTROL
£662.40
TREATMENT
£647.15
DIFFERENCE−£15.25 · ~−2.3%
STATISTICAL RESULTp = 0.235

More discount ≠ higher order value

Historical groups can differ in other ways. This comparison does not establish that discount level caused the order-value difference.

05 // REFUND-RISK SCREENING

Can we spot orders
at risk of a refund?

ProfitGuard compares order signals to help identify cases that may deserve earlier attention. This is an early-warning screening model, not an automated production decision system.

LATE DELIVERYDELIVERY DELAYDISCOUNT AMOUNT
SELECTED MODELLOGISTIC
REGRESSIONORDER-LEVEL SCREENING
OUTPUTREFUND-RISK
RANKINGNo fixed score shown
MODEL COMPARISON

DUMMY BASELINEReference point

LOGISTIC REGRESSION SELECTEDBest-performing model in this project comparison

RANDOM FORESTWeaker minority-class performance

LOGISTIC REGRESSION · HELD-OUT METRICS
0.25PRECISION
0.51RECALL
0.34F1-SCORE
0.700ROC-AUC

Metrics are moderate: a useful screening signal for investigation, not a production-ready automated decision.

06 // ANALYTICS DEPTH

Ask business questions
with SQL and Python.

Reusable query patterns and analysis move from data checks to revenue, refund and marketing questions.

CTEsWindow FunctionsLAGRANKCASEAggregationsMulti-table joinsDate functions
PROFITGUARD / ANALYSIS MAP
ANALYSIS AREAS 08

01Revenue→

02Products→

03Customers→

04Refunds→

05Delivery→

06Payments→

07Marketing→

08Data Quality→

ATHENA SQL → PYTHON · PANDAS · SCIPY / STATSMODELS

04 // BUSINESS IMPACT

Delivery is the clearest
place to protect profit.

Three practical recommendations from the evidence.

01

Reduce late deliveries

Late deliveries were associated with a 24.97% refund rate, compared with 5.83% for on-time orders.

RECOMMENDED ACTION

Prioritise delay prevention and proactively support orders at risk of arriving late.

02

Use refund-risk scoring as an early-warning tool

Logistic Regression reached 0.700 ROC-AUC in this project.

RECOMMENDED ACTION

Use risk scores to help prioritise review or outreach, with people in the decision loop.

03

Test discounts properly before scaling

The higher-discount group had lower average net order value; the historical difference was not significant.

RECOMMENDED ACTION

Run a properly randomised experiment before changing discount strategy at scale.

PROFITGUARD // E-COMMERCE PROFIT LEAKAGE ANALYTICS

Make the evidence
useful to the business.

10,000 orders, eight connected datasets and one clear place to start: delivery performance.