Essays, analysis, and technical notes on actuarial science, quantitative finance, data, and tools.
31 articles
If you can't get to your iPhone or Pixel, let someone answer for you. I hooked my phone's call forwarding up to Twilio and OpenAI's voice model, and in under a couple of hours I had an assistant that picks up the way I ask it to and texts me who called. Then I put it to work answering questions about my own work, right here.
Motion Commerce Lab turns marketing ideas into work people can open, review and change: six product ads and a proposed purchase page, built so the ad moment and the product-page moment are worked on together. This post tells how it was built.
The equivalence principle sets the premium so that what the insurer expects to collect equals what it expects to pay in claims. This two-minute video (narrated in Spanish) explains it with images, and I made it with a coding agent: Remotion for the animation and an ElevenLabs synthetic voice.
Ad platform and sales counts differ because each observes a different moment, unit and attribution window, so neither is necessarily broken. A fictional example with 20 attributed conversions against 15 confirmed orders reconciles them and shows which number answers which question.
Verify each step of the journey, from ad click to page, form submission and CRM record, before reading a funnel decline as a user-experience problem. A fictional ad-to-CRM example shows which events, parameters and outcomes to check so an improvement decision rests on evidence.
A dashboard that compares sources refreshing at different times should show a common cutoff, the last hour all of them have complete. A fictional day, with ads complete through 3 p.m., sales through 1 p.m. and web traffic through 2 p.m., shows how a return metric otherwise mixes three clocks.
On 678,013 motor policies XGBoost ranks risk better than a Poisson GLM, but a rate still has to be explained in a technical note and defended to a regulator, which keeps ML as a pricing challenger. The interactive lab shows SHAP explanations and a geographic fairness check.
A small team can have a room-booking app built to its own measure with Cloudflare Pages, a D1 database and a coding agent. This one replaced a cowork's physical whiteboard for its two meeting rooms, where checking availability meant walking over to read it, and put that information on the phone.
Pointing a custom domain at GitHub Pages through Cloudflare comes down to a few DNS records, a decision about www, and HTTPS with a real certificate. This walkthrough follows gonor.me from an empty DNS table to a working site, with the wrong turns left in.
An API is the interface that delivers data to your code, and analysts already use them when pulling Fed rates, Banxico exchange rates or World Bank life expectancy. Understanding it lets you diagnose failures, fix slow queries and build custom tools; the project shows this with real data and interactive labs.
Graph-based credit risk models loans as a network, so guarantee chains, circular guarantees and ownership concentration become measurable, which a relational loan table cannot express. A 500-client portfolio as a Neo4j knowledge graph, processed with PySpark on Databricks and scored with calibrated LightGBM, surfaces the structural risk that SQL keeps hidden.
RAG suits insurance regulation because LISF and CUSF articles cross-reference each other, and Ctrl+F cannot tell the article that defines technical reserves from one that mentions them in passing. The agent indexes each article individually with a cross-reference graph, eliminating citation hallucinations so the model reasons only over the real text of the law and amplifies the actuary's memory without replacing their judgment.
Major medical expenses (GMM) can be priced by splitting 5.1M CNSF claims into three hospitalization levels instead of one average cost. 9,409 diagnoses are classified into those levels, with frequencies and severities by age and sex and an interactive tariff calculator. It began as a UNAM team project.
An auto insurer needs to answer three questions: how much to charge, how much to reserve and where the fraud is. This dashboard answers all three for the Mexican market, where roughly 70% of vehicles are uninsured: a frequency-severity GLM for pricing, IBNR reserves from two complementary methods and anomaly flags on claims before adjustment, on a synthetic portfolio calibrated with public CONDUSEF and AMIS data.
suite_actuarial is an open-source Python library for Mexican actuarial work, covering eight domains (life, P&C, health, pensions, reserves, reinsurance, regulatory and configuration) with EMSSA-09 mortality tables, CNSF circulars and SAT tax articles built in. It installs with pip, deploys with Docker and exposes 28 REST endpoints alongside a bilingual Next.js dashboard; no other open-source library targets Mexican regulation.
Quantitative risk models become credible when implemented in typed Python with automated tests instead of left as formulas in a PDF. Thirteen modules cover portfolio VaR, Monte Carlo simulation, copulas, EVT, deep hedging and graph neural networks for systemic contagion, each pairing LaTeX theory with reproducible results on public market data.
An insurance claims data platform on GCP chains streaming intake, a BigQuery warehouse, Dataform transformations, Terraform provisioning and Tweedie GLM pricing in six connected projects. It frees actuarial teams from spreadsheets and manual processes with an automated, tested and reproducible flow that turns raw claims into regulator-ready inputs.
Use PostgreSQL for indexed point lookups (2.6ms against about 500ms minimum on BigQuery) and BigQuery for full-table analytics at low cost. Both are measured on 5.74M real flight records, from PostgreSQL index optimization through the migration to BigQuery, with real timings, costs and query plans.
The IMSS Pension Simulator compares what a worker would receive under Ley 73, Ley 97 (AFORE) and Fondo Bienestar, implementing all three formulas with current data (UMA, CONSAR tables, EMSSA 2009 mortality). Most Mexican workers do not know their regime or expected pension; interactive sensitivity analysis gives an educational estimate that shows what they control and what they do not.
Life insurance pricing starts by graduating raw mortality data: SIMA takes INEGI and CONAPO deaths and population, smooths them with Whittaker-Henderson and Lee-Carter, and projects forward to premiums, reserves and capital requirements under the LISF. Everything is exposed as an API to connect with other systems, automate sensitivity analysis and meet CNSF requirements; it is open source and built to expand into other lines of business.
Insurance pricing with ML splits the problem into claim frequency and severity, then compares a Poisson GLM against XGBoost and LightGBM on 678,013 policies from freMTPL2. SHAP explainability and a fairness audit check each model, and a cross-border analysis asks what these European pricing techniques mean for Mexico, where roughly 70% of vehicles carry no coverage.
A B-tree keeps data sorted in wide, shallow nodes, so a lookup among 500 million records can touch only about 9 nodes; PostgreSQL, MySQL and SQLite use it for their indexes. Understanding the structure changes how you design queries and judge the real cost of a lookup.
Seven end-to-end projects across six domains: real estate, insurance reserves, e-commerce cohorts, product experimentation, finance (SaaS KPIs and portfolio risk), and public-sector operations. The same statistical toolkit runs through all of them, and that is the argument: a data analyst has to hold several domains in view at once, because range is what makes depth transferable. Each project ships as an interactive dashboard, all on one live domain.
Next.js with Recharts is a serious alternative to Power BI and Tableau when a report must ship as a finished, fast, mobile-friendly product instead of a workspace artifact. Two dashboards, an Airbnb CDMX market analysis and an Olist e-commerce cohort study, test that claim.
Insurance reserves are estimated by developing loss triangles with methods such as Chain-Ladder and Bornhuetter-Ferguson, which produce IBNR estimates. An interactive dashboard on NAIC Schedule P regulatory data applies both actuarial methods across six lines of business, with loss triangles, IBNR and combined ratios, to answer whether reserves are enough.
To find a percentile of a transformed random variable, such as an insurer's payment after a deductible, invert the transformation and use the CDF you already have instead of deriving a new distribution. The post builds this from the ordinary deductible, the way SOA Exam P tends to ask it.
The Central Limit Theorem (CLT) is the mathematical justification for why insurance works as a business: aggregating many risks makes the total behave predictably. This SOA Exam P study guide also covers multivariate random variables, covariance and dependence, and Eve's Law for decomposing variability.
The univariate random variables section, about 45% of SOA Exam P, tests distributions, expected values and payment modifications such as deductibles and limits, and mechanical errors cost more than conceptual ones. This study guide shows how the pieces connect and why they matter in insurance.
The first section of Exam P covers probability axioms, conditional probability and Bayes, the toolkit an actuary uses to classify risk and decide under uncertainty. This study guide explains why the SOA puts them first and how they connect to real actuarial work.
A transformer built from the first matrix multiplication in pure NumPy, with no autograd, shows what happens inside a language model. Trained on all 7 volumes of Proust, it taught me less about the architecture than about what it means that everything is just numbers.
Three interactive SVG visualizations let you manipulate the sin(x)/x limit, the derivative of sine and Euler's formula to see why each holds. They matter in actuarial and financial work: sin(θ) ≈ θ for small angles underpins the linear approximations used in sensitivities.