Notes from the bench.
Studies, class notes, computational and statistical experiments, implementations and tests. The posts below hold part of what I have been thinking since this blog was created. They may help you learn, apply or correct something, but the real purpose is the simple act of writing.
03 / articles
The p-value, demystified
The most used and most misread number in statistics, explained with a coin and two A/B tests: what the p-value measures, where it is used, what it does not say, the multiple-testing problem, and how it compares with confidence intervals, Bayesian methods and information criteria.
statistics12 min read
Hyperparameters for statisticians
The ridge λ, the bandwidth and the loess span are hyperparameters, and statistics has been choosing them for 50 years. What they are, why they cannot be estimated with the parameters, how CV, GCV, AIC and REML compare in a simulation, and how grid, random search, Bayesian optimisation and successive halving tune many at once — with the bias of picking the best of many.
statistics · machine learning30 min read
Quasi-likelihood in GLMs
The equations of a GLM only use the mean and the variance function. Wedderburn drew the conclusion in 1974: specify just those two and forget the distribution. The origin, the proofs, the estimators, the algebraic advantages and three examples — with simulations of when quasi models are the right choice and when they are not.
statistics · glm35 min read
Docker from scratch: what it is, why it exists and how to use it
For people who have never used Docker: why “works on my machine” happens, how a container fixes it and how it differs from a virtual machine. Then hands-on — a first container, your own Dockerfile, layers, volumes and an application with Docker Compose.
engineering · docker25 min read
Claude skills: what they are, when to use them and how to build one
What a skill is, how Claude decides to use it and how to build one from scratch — demonstrated with a skill that teaches Claude to make charts with the principles of Storytelling with Data.
ai · claude skills16 min read
From Poisson to the negative binomial
Write the count in two layers — a Gamma rate for each unit, then Poisson given the rate — and the iterated expectation, the law of total variance and one integral show where overdispersion comes from and why the answer is the negative binomial. Then a fit and a simulation measure the cost of ignoring it.
statistics · glm30 min read
Non-parametric vs parametric bootstrap
From zero to rigour: what the bootstrap is, how the non-parametric and parametric versions work, how to build each confidence interval and, in simulations where the answer is known, when each one gets it right and when it breaks.
statistics · bootstrap35 min read
Local LLMs: a guide from scratch
What a language model running on your own computer is, when it beats the cloud, what you need to run one and which model to pick for each task — no AI background assumed.
ai · local llms15 min read
Why the house wins
A statistical view of betting: why the margin is in the price, why luck runs out after about 1,800 bets, and what it would take to actually win.
probability · betting12 min read
Is τ = 0.24 really a cost ratio?
Checking the caveat of the previous note: are the scores calibrated, does recalibration help, and how noisy is the optimal threshold?
calibration · resampling7 min read
A threshold is a decision
Why 0.5 is not neutral, what τ = 0.24 implies about error costs, and how to choose it without touching the test set.
statistical decision theory6 min read
fmt / format — Each note: a question, the math, real data, and what it changes in practice.