Projects

An interactive marimo notebook that explores trend and seasonality in sales.

Question: How do trend, seasonality and noise combine to shape a monthly sales series?
Approach: Generate a synthetic series and adjust trend, seasonal amplitude and noise with sliders.
Result: With the default settings, average sales peak in April and bottom out in October.
Data notes: Synthetic data generated in the notebook; no real data. Runs in the browser via /wasm.

A Streamlit app that clusters customers and lets you inspect each segment.

Question: Can customers be grouped into distinct segments by spend, visits and basket size?
Approach: K-means clustering on standardised features, with the number of segments and customers adjustable in the app.
Result: With k = 3 the app recovers three distinct, similar-sized groups: high spenders, frequent visitors and low-spend shoppers.
Data notes: Synthetic data generated in the app, with no personal information. Well-separated groups make this easier than real data.

A marimo notebook that compares simple baseline forecasts with a trend-plus-seasonality model.

Question: How much does a trend-plus-seasonality model improve on simple baselines?
Approach: Backtest naive, seasonal-naive and trend-plus-seasonality models on a held-out period, scored by mean absolute error (MAE).
Result: With the default settings, the trend-plus-seasonality model has an MAE of 5.4, against 14.5 for seasonal naive and 21.1 for naive.
Data notes: Synthetic data generated in the notebook, so results will differ on real series. Runs in the browser via /wasm.

A reproducible Quarto article that uses simulation to illustrate the central limit theorem.

Question: How quickly do means of a skewed distribution become approximately normal?
Approach: Simulate 5,000 sample means at several sample sizes and compare them to the normal curve.
Result: Skewness falls from about 2.0 at n = 1 to about 0.1 at n = 200.
Data notes: Synthetic exponential data with a fixed seed. Reproduce with uv sync && uv run quarto render.

A short Typst paper on confidence intervals for a mean, published as a PDF.

Question: How is a confidence interval for a mean built, and how does its width change with sample size?
Approach: A worked t-interval example, typeset in Typst and compiled with uv.
Result: The interval width shrinks in proportion to 1/√n, so quadrupling the sample size halves it.
Data notes: Illustrative numbers only. Reproduce with uv sync && uv run typst compile papers/confidence-intervals.typ.