I learn things best by going on deep dives, often through playing around with code or collecting pieces of intuition I like. These posts are not meant to be instructive but to document my deep dives.
Cambridge Part IIB dissertation. Evidence of strategic vagueness in Singaporean political communication. A signal model with rational voters predicts vagueness shields a favourable prior but not an unfavourable one. I build a semantic dispersion measure with LLMs (92.3% agreement with human raters), test it on 62,169 hedged Singapore parliamentary claims, and identify the effect off a redistricting shock plus the Nominated MP institution.
Interactive 3D view of a utility hill, indifference curves as contours, and budget geometry.
Singapore's short rates sit 100–250bp below USD rates, yet unhedged carry produces a Sharpe of −0.05. The MAS crawling band pins expectations, fast mean-reversion kills the carry window, and overwhelming reserves make the regime unbreakable.
I scraped a massive corpus of guest participants' interviews and publications to build personas, used a multi-agent workflow to simulate debate engagements dozens of times, and clustered extracted arguments to build a "debate landscape" forecast.
Working through the variance of Ridge predictions, I noticed the connection between Ridge shrinkage and principal component analysis — and why an isotropic penalty on anisotropic data produces PCA's coordinate system.
Twelve parts in one page: OLS as projection, L² and expectations, classic identities, inference, GMM, asymptotics, BLUE, bias, IV and 2SLS, efficiency, and Gauss-Markov assumptions, with figures and a left sidebar to switch parts.