Machine Learning
20 articles
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.
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
I would rather flag thirteen benign tumors than miss four malignant 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
Losing to Holt-Winters by 4.6 points was the good news
Decomposing Air Travel
44 passengers in 1949, 232 in 1960, the same summer bump
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 numbers off a wine label beat your fancy model
What Makes Wine Good
Wine quality is mostly just alcohol, and even that only gets you so far
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.
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.
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.