This is the third note in a series on quantamental stock rankings. The
first note
(Ranking Before Prediction) and the second one was
(Why
Learning Factor Weights Is an Ill-Posed Inverse Problem)
In quantitative equity research, we rank stocks.
In large language models, we rank tokens.
At first glance, these activities seem unrelated. One deals with financial
markets, the other with text. One operates on quarterly or monthly
horizons, the other on milliseconds. And yet, once you strip away
domain-specific language, the core operation in both systems is the
same:
Given a large discrete universe and incomplete
information, construct an ordering and act on the top of
it.
Prediction is often presented as the central task. In practice, selection
is.
In equities, we rarely need a precise return forecast for every stock.
What matters is which stocks end up near the top of the list. In
language models, the system does not need to know the “correct” next
word. It produces a ranked distribution and selects from it.
This blog explores why ranking-based systems arise naturally across
domains, why early language models looked surprisingly similar to
factor models, and what modern language models changed—not in the
objective, but in where structure lives ...
Please read the full note in pdf
