Analytics portfolio

Andrés González Ortega / Data analyst

Behind every number,a story to understand.

I’m Andrés. I’m drawn to statistics as a way of understanding the world: asking questions, testing ideas against evidence, and changing my mind when the data gives me reason to. This portfolio introduces classic problems we can tackle with data: comparing alternatives, understanding customer behavior, and estimating risk.

Explore the storiesWhat will I find here?

No data background needed. Just a little curiosity.

Paper-cut illustration of receipts and a parcel being examined through a magnifying glass, beside a simple chart and notebook.
Everyday information. A clearer picture.

A little orientation

Not just charts. The thinking behind them.

A customer makes a purchase and never returns. Sales fall. An insurance claim still has no final cost. These seven stories begin with situations like these. We follow the clues in the data to understand what is happening and what we might do next.

  1. Start with a question

    A situation you can recognize, with a reason to look closer.

  2. See what the data says

    Visual explanations make patterns—and missing pieces—easier to see.

  3. Understand the decision

    What the findings could change, and where we need to be careful.

Featured study / Insurance

Recent claims have a shorter history. Follow one accident year to see how observed payments become an estimate of the obligation still ahead.

Read the case study

CAS / NAIC · paid Chain-Ladder · historical regulatory aggregates

FIG. 01Retained evidence

Claims payments made and estimated payments remaining

Year1510
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997

Development period →

36.1%

of projected ultimate losses already paid

1997 · 1 observed periods
Paid · 36.1%Projected unpaid · 63.9%
Paid Chain-Ladder · all API companies and lines · 1988–1997. Newer years have less observed development. Projected unpaid losses include case reserves and IBNR; this is not an uncertainty interval. API snapshot · 2026-09-05.

The collection

Seven questions worth investigating.

01—07
01Insurance

How much remains to be paid, and how sensitive is that estimate to the reserving method?

The newest accident years contain the least observed development. Their ultimate cost depends most on the projection assumptions.

Insurance records and simulated claims
SQLPythonActuarial analysis
FIG. 01Retained evidence

Claims payments made and estimated payments remaining

Year1510
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997

Development period →

36.1%

of projected ultimate losses already paid

1997 · 1 observed periods
Paid · 36.1%Projected unpaid · 63.9%
Paid Chain-Ladder · all API companies and lines · 1988–1997. Newer years have less observed development. Projected unpaid losses include case reserves and IBNR; this is not an uncertainty interval. API snapshot · 2026-09-05.
02E-commerce

Does marketplace activity translate into customers who return?

Acquisition, order value, and repeat purchasing answer different questions. A cohort view makes the second purchase visible.

Olist purchase records
SQLPythonCohorts & RFM
Marketplace performanceCustomer retention
FIG. 01Retained evidence

How many customers bought again?

3.0%

of customers placed at least two orders in the retained observation window.

2,801 repeat customers93,358 customers total
Olist customer-retention artifacts · 2016-09-15 – 2018-08-29 · Source counts. This population is separate from the marketplace aggregation.
03Product analytics

Is the treatment effect large and consistent enough to justify a rollout?

This constructed experiment demonstrates how an aggregate result can conceal different effects across user groups.

Simulated data for A/B testing practice
PythonA/B testingBayesian analysis
FIG. 01Illustrative example

Estimated change in conversion and its uncertainty

No effect−3 pp+3 ppBoth improvement and harm remain plausible
Constructed examples: point estimates of +0.4 and +1.6 percentage points, with illustrative 95% intervals [−1.3, +2.1] and [+0.6, +2.6]. These are not results from the experiment. Practical benefit, guardrails, and experiment quality still matter.
04Business reporting

What changed in recurring revenue, and which changes deserve investigation?

Revenue growth and customer health need separate explanations. A revenue bridge shows the contributors; anomaly baselines identify unusual observations.

Simulated data from a fictional business
PythonForecastingPDF reporting
FIG. 01Retained evidence

What changed monthly recurring revenue?

952.2Opening MRR
+8.2New
+34.2Expansion
−3.4Contraction
−21.5Churn
969.7Closing MRR
USD thousands · December 2025 · net movement: +17.5
NovaCRM synthetic scenario, seed 42. This bridge uses the same figures as the December report; displayed components are rounded. Growth is a net movement: expansion can conceal contraction and churn. It does not establish a cause. Source bridge · 2026-09-05.
05Financial analytics

How does the allocation change historical risk and model-based scenarios?

Return, drawdown, and simulated outcomes describe different aspects of a portfolio. The observation window and model assumptions connect them.

Historical prices and simulated scenarios
PythonMonte CarloOptimization
FIG. 01Illustrative example

Two investments with the same final value and different losses along the way

Same final valueStartTime →
Conceptual paths, not historical returns or a forecast. Equal final returns can conceal very different peak-to-trough losses. Compare drawdown, volatility, and concentration before interpreting simulated future outcomes.
06Operational analytics

Where do request volume and resolution delay intersect?

Prioritization requires both volume and delay, with agency and complaint mix held in view. A ranking alone cannot explain a bottleneck.

New York 311 service records
PythonD3Process mining
FIG. 01Illustrative example

How a service request is recorded and its resolution time calculated

  1. 01Request receivedCreated date
  2. 02Agency assignedCompare workload
  3. 03Open or closedStatus at extraction
Resolution time = closed date − created date
Reading guide, not measured stage durations. The 311 snapshot contains current statuses, not a complete event log. SLA comparisons need both a closure date and a due date; unresolved requests must remain visible separately.
07Market research

How do location, room type, and host scale shape the visible listing market?

Listing prices and host concentration describe the supply on display. They do not directly measure bookings, occupancy, or realized revenue.

Airbnb listings in Mexico City
PythonExploratory analysisRecharts
FIG. 01Retained evidence

Listings are concentrated in a few boroughs

Cuauhtémoc12,514
Miguel Hidalgo4,573
Benito Juárez3,087
Coyoacán1,727
Álvaro Obregón978

Listings · five largest borough counts in the retained artifact

Inside Airbnb-derived borough counts. Artifact updated: 2026-03-02. This timestamp is not a verified collection date. Listing supply does not measure bookings, occupancy, or host income.

How I approach the work

I studied actuarial science at UNAM. My work connects quantitative methods with questions about customers, risk, and operations. These studies show the reasoning, assumptions, and limits behind the result.

Public data, synthetic demonstrations, and modeled scenarios are identified in each case. The methods and research are available alongside the visual explanation.