Research & Writing
103 data essays, notebooks, and case studies. The essays take public datasets apart and show the gears: the markets, the models, the climate, and the people the numbers describe. The case studies come from systems I built.
Stocks Versus Bonds
Stocks vs bonds since 1871: the long run is real, and it is paid for
The Moving-Average Rule
The moving-average rule halves the worst crash, and in this backtest it also edged the market
Price Versus Earnings
Does the stock market actually follow earnings?
The Dividend Share of Returns
How much of the stock market's return is just dividends
Missing the Best Months
Sitting out the best months is expensive, and the worst months are next door
Cross-Asset Correlation
When the market breaks, stocks and oil stop diversifying each other
The Dollar's Gravity
The dollar's gravity: which currencies track it, and which float free
The Dow's Fat Tails
The Dow's worst month was a 5-sigma event that a bell curve says should never happen
The 2021 Inflation Surge
The 2021 inflation surge was the broadest in decades, not the worst
Gold in Real Terms
Gold made its real high in 1980, and has been losing the rematch ever since
150 Years of Drawdowns
150 years of S&P 500 drawdowns: the market is usually underwater
Market Seasonality
Sell in May, and other things the calendar does not actually tell you
Valuation and Returns
What you pay for stocks sets what you earn, and right now you are paying a lot
The Widow-Maker
Natural gas is the widow-maker, and the winter is where it earns the name
Forty Years of Oil Shocks
Forty years of Brent, and the shocks that define it
Apple Versus the Field
Apple turned $1 into $8.60. Everyone else mostly stood still.
On disabling ML in production
What I learned when my live trading system's ML ensemble silently degraded in production, and the disciplined reintroduction of machine learning that came after.
Anatomy of the VIX
The VIX spends most of its life bored, then loses its mind
Atlas in production: putting a forecasting system in front of real capital
How the Atlas forecasting system handles 542,000 rows/second of market data with sub-second regime detection — async service architecture, dependency-ordered startup, and 10Hz health monitoring.
How Atlas's database got 810× faster: a single-pattern fix
Atlas couldn't start. The trading system's database initialization was taking 6.6 seconds, blocking 37 features from loading. The fix was small.
The Sound of a Hit
What a popular song actually sounds like, according to 28,000 of them
Which Models Make Things Up
Which AI models make things up the most
Brain Signal Redundancy
62 signals, 21 real dimensions: redundancy that does not look like redundancy
Diagnosing on Three Features
Three of thirty features get you within 0.02 AUC of the full model
The Cancer Decision Threshold
The default threshold misses four malignant tumors. I would rather flag thirteen extra benign ones.
Pricing Diamonds
A gradient booster prices diamonds to $276. Then it meets a big one.
Recognizing Digits
Twenty numbers per digit gets you 94% of the way there
The Pixels That Matter
Twelve of the 64 pixels are dead, and the classifier never misses them
Forecasting Air Travel
Holt-Winters beat a one-line forecast by 4.6 points, and the one-liner still taught more
Decomposing Air Travel
The summer bump grew from 44 to 232 thousand passengers, and most of that is scale
Old Faithful's Two Modes
Old Faithful is two geysers wearing a trench coat
The GLUE Leaderboard
The 0.7 points that decide a leaderboard, and where they come from
GLUE and the Transformer Leap
22.67 points: what the field bought by dropping the recurrence
Clustering Penguins
Two penguin measurements beat four
Penguin Dimorphism
A female Gentoo outweighs a male Adelie by 636 grams
Predicting Titanic Survival
The random forest lost. By 0.002 AUC.
Classifying Wine
Three lab readings beat your fancy model
What Makes Wine Good
Wine quality is mostly just alcohol, and even that only gets you so far
Deploying ML in production: a working reference (Part 1)
Serving architectures, containerization, lifecycle management, performance optimization, drift detection, and monitoring — with benchmarks and code from production systems.
A local-LLM scraper for Chamber of Commerce directories
A pipeline that reads Chamber of Commerce directories with a local 7B-parameter model instead of a CSS selector per site. No API costs, and nothing leaves the machine.
ML deployment: a working reference for getting models into production
A field-tested reference for taking ML models from prototype to production — serving patterns, containerization, monitoring, drift detection, and the operational practices that make the difference.
Who Emits the Carbon
Who actually emits the carbon, and the answer depends on how you count
The Accelerating CO2 Curve
The CO2 rise is speeding up, and the curvature is the story
The Warming Rate
The warming rate has sextupled, and the line cannot keep up
The Arctic's Shrinking Ice
The Arctic's September floor is dropping 0.84 million km² a decade
A World Map of Drinking
A world map of drinking, and what the colors give away
Where Energy Comes From
Where the world's energy actually comes from
How Steam Games Fade
The biggest games on Steam, and how fast the rest fade
The Instant Ramen Rankings
The best instant ramen in one reviewer's archive comes from Malaysia, by a nose
How Fast Passwords Fall
The 500 most common passwords all crack in under a minute
Inside the Netflix Catalog
What is actually on Netflix: a catalog built in five years, made mostly of movies, and getting older
What Makes Great Coffee
What actually makes a great cup of coffee
The Candy Power Ranking
The best candy, settled by data
Bob Ross by the Numbers
Every Bob Ross painting, by the numbers
The Bechdel Plateau
The Bechdel test: the pass rate climbed, then stopped
Bird Strikes on Approach
Bird strikes cluster on approach, and big birds do the damage
Premiums and Road Risk
Higher insurance premiums do not mean more dangerous roads
Two Americas of Driving Risk
Two Americas of driving risk, and the insurer charges them the same
The Aging of Congress
Congress got older, but it took a detour first
Disease in the Crimea
In the Crimea, disease killed eight soldiers for every one the enemy did
A Century of Disasters
Five years killed 44% of everyone who died in a natural disaster since 1900
Gas Prices and Driving
Gas prices and American driving look unrelated, until a price spike hits
Money and Life Expectancy
Money buys life expectancy, but the receipt is brutal
When Coders Commit
The 9-to-5 I went looking for was not in the commits
US Health Spending
The US spent $7,194 more per person than Japan and expected to live 7.7 fewer years
The Health-Spending Gap
In 1991 the US health-spending gap crossed $1,000 and never came back
Jobs That Switched Gender
The job that went from all-male to 96% female
College Majors and Pay
Your major sets the odds, not the salary
Box-Office Economics
Spend more, make more, but a third of films still lose money before the lights come up
Aging and Sex Ratios
At 90, there are three women for every man
How People Tip
Tip 18.6% on a small check, 13.3% on a big one
Unemployment by Industry
Construction hit 27% unemployment. Government never cracked 6.
The Jobs That Vanished
Construction lost more than a quarter of its jobs, and never got them back
Who Owns the Economy
Five countries own half the world economy
The Deadliest Mountains
The deadliest mountain is not the one you think
Pulse Under Exertion
Thirty people, ninety pulses, and one mistake everyone makes
The Olympic Body
The body an Olympic sport builds, and how it sharpened over fifty years
A Century of Medals
A century of Olympic medal power, and who traded the lead
Anscombe's Quartet
Four datasets agree on every number and agree on nothing else
Attention and Performance
Focused attention wins by 1.68 points, until the task gets easier
Reading Diamond Prices
The best-cut diamonds are the cheapest, and other lies the data tells
A Diamond's Four Cs
Two of a diamond's measurements are basically decoration
Evidence and Perception
Stronger evidence does not raise the signal; it splits it
fMRI Response Curves
The average brain response is a curve, and averaging is what draws it
Lessons from 30 Datasets
Three numbers off a wine label, and four other things 30 datasets taught me
The MPG Paradox
A bigger engine made these cars more efficient, and other lies of regression
Detroit Versus Japan
Detroit caught up to Japan, but only after you weigh the cars
Mapping the Exoplanets
The exoplanet catalog is a map of our telescopes, not the galaxy
How Exoplanets Cluster
Almost seven orders of magnitude, and the planets pile up at one end
Seattle's Weather
Seattle rains on 43.9% of days. It is sunny on 43.8%.
Who Survived the Titanic
Women and children first, or first class first?
Titanic Families
The Titanic's loneliest passengers, and its largest families, drowned together
Seattle Versus New York
Seattle rains on 10 more days out of every 100 than New York, and New York's rain falls harder
Averaging the Wind
The wind blew north. The average said east-southeast.
America's Farm Exports
Five states ship a third of America's farm exports, and none of them grow the same thing
Cinestyle: matplotlib themes pulled from film
A small Python library of matplotlib themes — Film Noir, Ghibli, Wes Anderson, Blade Runner, Star Wars — applied to 50,000 IMDB reviews.
WebAssembly visualization with Rust: when JavaScript runs out of room
Building browser-side data visualizations in Rust compiled to WebAssembly — particle systems, large-dataset rendering, and the practical wins over a pure-JS implementation.
Database optimization: a working reference
Six database optimization techniques — predicate pushdown, row group pruning, result caching, async I/O, indexes, and where SIMD goes wrong — with impact numbers and code.
Replacing DuckDB with Rust: 10.4× through predicate pushdown
I replaced DuckDB with a custom Rust query engine for a trading-system time-series workload. Five iterations, 10.4× speedup, one optimization that backfired.
Database performance optimization: a reference (Part 1)
A working reference for production database tuning — SQLite PRAGMAs, schema, indexing, transactions, batching, pooling, and the monitoring that proves it's working.