Dear markets and research team,

The project of mine closest to your work is The time machine problem. It builds a point-in-time database of SEC filings and tests how much AI models remember, rather than predict, when asked about the past. Three things from it matter for market data.

Today's data isn't yesterday's. 7.6% of standardized SEC numbers were later re-reported with a different value, so a backtest on current data uses information nobody had at the time.

Filings break identities more often than you'd think. Quarters fail to add up to the year in 6.6% of checks, and cash-flow statements in 4.4%.

Masks can leak. A company is often recognizable from its numbers alone, so a "masked" AI backtest can still remember. Try guessing three companies.

At UNC I led a team running a simulated $100,000 multi-asset ETF portfolio that finished first among the participating teams. Next I'm testing whether AI models read Chinese annual reports as accurately as English ones; the plan is here.

If any of this is close to what your team works on, I'd be glad to talk. I'm looking for a part-time, remote internship, 10 to 20 hours a week: adamzhao795@gmail.com.

Junming