A running list of writing ideas

Here are some ideas I have that probably aren’t worthy of a full research project (or I don’t have the expertise to do them) but I think are interesting.

Markets

• Calibration of prediction markets
  • Prediction markets have become very popular in the past couple years. Are they actually calibrated? At what time horizons? How do they compare to general prediction tools?
• How to evaluate investment portfolio strategies
  • Is there a way to establish statistical significance of A vs B strategy?
  • What is the null distribution of returns? How can we construct a null distribution of strategies?
  • Should we use the integral between A and B as a metric? Should we only use a single point? What time horizons should we use?

Personal/Random

• Analyzing songs through embedding APIs and/or SAEs
  • What themes are most important in different song genres?
• Analyze my own bike trips
  • I have lots of Strava data

Social Choice and Economcs

• How can we aggregate human preferences?
  • How can we align ML systems if we can’t aggregate human preferences properly?

Biology and Health

• What is the entropy of the human genome?
• Do people actually make better decisions when given calibrated advice? Do doctors? Should we only use calibrated models in clinical decision making?
• What is the full distribution of SNP effect sizes? Are there ways to collect and interpret SNP effect sizes without using linear models?
• How can we assess if polygenic risk scores are actually useful? If we tell people their PRS and high-risk individuals don’t develop disease X (e.g., a cardiovascular disease), is it because the PRS was wrong or because they positively changed their behavior?
• For a long time, bulk RNA-seq was the gold standard for measuring gene expression. After that came single-cell RNA-seq. Now we have access to spatial transcriptomics. If we could design a new method for measuring gene expression (without any constraints), what measurements should we prioritize?