RESEARCH
Research
What we measured, how, and where it is not yet good enough. Every number has a date range and a source.
- What is tax-loss harvesting worth? About 2,000 backtests across four benchmarks 2026-11-13 · Expected tax alpha against the S&P 500, Russell 1000, S&P 400 and S&P 600, by start date and account size, and how it compares with AQR, Vanguard and others.
- How accurate is our risk model? 2026-11-13 · A short read: 406 month-ends from 1993 to 2026, five alternative models, and why optimized portfolios are the test that matters.
- Building an equity risk model for AI agents 2026-11-06 · A point-in-time factor risk model, scored on 406 month-ends from 1993 to 2026 with bias statistics, tracking error and optimized-portfolio tests. Why most risk-model failures are evaluation failures, the bugs the evaluation found, and whether an LLM could tune it.
- Why LLMs cannot write feasible trade lists Coming next · Frontier models on real accounts: cash that does not add up, fractional shares, missed wash sales, and what an exact solver does instead.
- Text to portfolio: training a model on a million exact-solver decisions Coming next · Distilling an exact mixed-integer solver into a model that proposes plans in tens of milliseconds, and how we check it never leaks the label.
- Point-in-time or nothing: thirty years of survivorship, delistings and filing dates Coming next · The data discipline behind every backtest: delisting returns, as-of fundamentals, effective-dated index membership.