Commercial Data Analyst
Plan vs actual · Price-volume-mix · Reconciliation
I build the reporting layer commercial teams run on — plan versus actual, price-volume-mix, margin, and channel efficiency — and the controls that prove it reconciles. Where sell-in disagrees with sell-out, or finance and commercial quote different revenue, somebody has to find out which number is correct and be able to show their working.
Selected work
E-commerce
2026 · Approximately 1 month · Sole contributor
Two source systems reconciled onto one country-month grain. 35 checks across three gates decide what reaches a dashboard, and the forecast was held back until an audit had found two defects in my own code.
Grocery retail
2026 · Approximately 2 weeks · Sole contributor
Own-price elasticity for 147 categories. The campaign evaluation’s first answer said coupons reduced spend and the standard check endorsed it; matched on pre-period spend, every sign flips.
Mobility & delivery
2026 · Approximately 2 weeks · Sole contributor
54 weeks across 8 acquisition channels, reconciling to the company total every week. A scope labelled “All Vietnam” turned out to be exactly two cities. Every published figure is indexed.
Advertising
2026 · Approximately 2 weeks · Sole contributor
852,197 records audited before any effect was estimated. 26,426 users appeared in both arms of a test that assumed one arm each, so the recommendation was a conditional rollout rather than a green light.
Fintech
2026 · Approximately 1 month · Sole contributor
4.26% of scored users are associated with 26.94% of credited promotion cost. Recommended a shadow-mode review queue rather than auto-blocking, because the data carries no fraud labels.
A bank-marketing targeting model — the top-ranked 20% captures 51.8% of responders at 2.59× lift while cutting outreach 80%, with a randomised pilot recommended rather than deployment. The dataset is a widely-used teaching set and credit and risk is not a lane I am pursuing; what transfers is the targeting economics — capture, lift and contact cost against a fixed capacity — rather than the domain.
What it looks like in practice
Each row is a figure that was reported, or about to be, and the check that changed it. Four of the six came out of my own work.
The comparison group spent 7× less before the campaign started. The parallel-trends test passed anyway, because it measures slopes rather than comparability. Matched on pre-period spend, every sign flips.
mart.campaign_effect · 17 campaigns, 6,407 treated householdsExactly Ha Noi plus Ho Chi Minh, to 100.000% in every month. No residual, so no third city. Read as national it would overstate coverage in every slide it reached.
qa.scope_reconciliation · 54 weeks, revenue and commissionIt was — but it was chosen one month ahead and then run eight months ahead, on six folds that split three-three. A coin flip that landed right. Rebuilt to 192 rolling-origin folds.
docs/forecast_review.md · seven checks specified in advanceA sign error. Caught by the reconciliation statement, which has to balance, rather than by an eye scanning a chart.
qa.reconciliation_statement · 292 plan rows submitted17% of customers contributed zero net revenue — everything they bought was cancelled or returned. Excluding them: 0.49. Which population was used is stated wherever the figure appears.
artifacts/advanced_metrics.json · both populations reportedThe randomisation unit could not be confirmed from the data supplied, and 26,426 users appeared in both arms. At user-cohort level the effect is +5.5%.
852,197 records · 36,893 duplicates removed firstA result can be wrong while the standard check that is supposed to catch it reports success. Parallel trends tests whether two groups move at the same rate; it says nothing about whether they are comparable at all. What exposed it was looking at the levels the two groups started from.
Query my portfolio
Run SQL against a curated catalog of my projects, tools and published findings. This interactive demo uses portfolio metadata, not the underlying client datasets.
Demo data: a curated portfolio catalog. Counts have different grains and must not be summed across projects; null means undisclosed or not comparable.
Press Ctrl+Enter to run · SELECT only
How I work
My projects show this working approach at different depths. The decision determines which checks, methods and tools are needed.
01 / Frame the decision
Define the stakeholder, the unit being compared, the observation window and the action the evidence could support. Choose metrics around that decision: a limited contact queue needs capture and lift, not accuracy alone.
The top-20% contact band connects historical responder capture to a realistic campaign workload.
Decision brief
Decision boundary. State the outcome, capacity and evidence that would change the recommendation. Historical ranking does not establish campaign uplift.
02 / Check the evidence
Profile row counts, missing values, duplicates, types and distributions before interpreting the headline. Check feature availability at the time of a decision, and keep exclusions visible.
Duplicate events and users appearing in both groups make the assignment grain a material uncertainty.
Source audit
Decision boundary. Separate a confirmed error from a sensitivity exclusion. If assignment or scoring-time provenance is unresolved, limit the claim.
03 / Build & reconcile
Transform source tables into declared analytical grains. Reconcile totals, joins and denominators; compare independent calculations where available. Keep metric definitions close to the outputs.
2,420,661 events reconcile to 680,862 reconstructed sessions; retail measures feed the Power BI model.
Reconciled analytical tables
Decision boundary. A plausible chart is not a validation. Unexplained duplicate joins or totals must be resolved before interpreting a metric.
04 / Test the interpretation
Use uncertainty intervals and sensitivity views when the design supports them. Separate prediction, association and causal effect. Explain why a metric fits the data and record assumptions beside the result.
Rule comparisons balance credited-cost coverage, review capacity and potential customer impact; signals are not fraud labels.
Evidence with its limits
Decision boundary. An offline lift, flagged cost or model score is not realized value. Unmeasured uncertainty stays unmeasured; it is not a guarantee.
05 / Recommend & hand over
Lead with the recommendation, then its evidence and conditions. Document provenance, run steps and known limits. Specify the next check or controlled pilot; withhold rollout when the evidence is insufficient.
Near-chance held-out ranking led to a non-deployment recommendation and a need for better feature and outcome logging.
Decision report and handoff
Decision boundary. Review capacity, sensitive attributes and customer impact where relevant. A positive point estimate alone is not permission to scale.
Commercial background
Golden Owl’s agency and SaaS clients, Laptop SGN’s electronics retail, Canawan’s Etsy and Amazon marketplace. Campaign delivery, SEO and paid acquisition are where I learned the commercial questions — what a channel costs, what it returns, and when a number is being flattered by attribution. Client reporting belongs to the client, so what is here is the judgement behind the work rather than the files it was delivered in. I am not offering marketing as a service.
What others say
Quoted in full and unedited, from the mentor who ran the data analytics programme I completed in 2025.
I had the pleasure of mentoring Hoai Thu through my Data Analytics course, Thu joined track 2 which specialized for end to end data analytics project, and she is truly one of the most impressive students I’ve worked with.
From day one, Hoai Thu stood out for her self-driven attitude and ability to pick up complex concepts quickly — she never waited to be pushed. What impressed me most, however, was how she communicated her findings. Her project deliverables weren’t just technically solid; they told a clear story, turning data into insights that anyone could understand and act on.
The quality of her work consistently exceeded expectations — structured, thoughtful, and delivered with care.
I have no doubt Hoai Thu will bring that same energy and clarity to any data role she steps into. Any team would be lucky to have her.
What he singled out
Self-driven, never waited to be pushed Communicates findings so anyone can act Technically solid deliverables Consistently exceeded expectationsEducation & credentials
University of Economics Ho Chi Minh City
2018 – 2022TheHau · selected 15-person cohort
Mar – Jul 2025Academic English proficiency
2022Let us talk