Spatium Novum — Models, and where they break
Different partial models can agree. The real work is showing when their agreement tells us something relevant about the world.
I build models for a living and read about them the rest of the time: commodities and finance, machine learning, pharmacology, economics, cognition. I have not settled on which of those I will end up working in, so for now I write about all of them, and about the part they share.
Five Hundred Years of Data
Harry Markowitz formalised the second stage of portfolio choice in 1952. The first stage was deciding what to believe about the future.
Where the Proof Runs Out
The same local solver can report success at $0, $100 and $400 on one small continuous problem. Only a valid bound can say which answer is global.
Optimising Against a Guess
The optimum of a model is the optimum of your estimates. Those are not the same thing.
When the Answer Is Yes or No
Half a warehouse is not a warehouse. That innocent fact changes both the answer and the proof.
Rows, Columns, and What They Cost
Stigler's diet has 77 choices and nearly half a trillion candidate column sets. It still solves in milliseconds.
The Last Place "Optimal" Means Anything
OPTIMIZE is the motto of the age. In one discipline the word still means something you can prove, and in 1947 that proof cost 120 man-days.
The Swiss Commodity Trading World, on Two Axes
One of the oldest industries there is, and one of the least documented. An interactive map that puts every kind of actor on a grid of what they do against what they move, so you can see where a firm like Trafigura actually sits.
New anchor essays go up roughly every two weeks, and the RSS feed carries them as they land.