A small site about measuring things properly

The study started from a newspaper profile of an investor handing money to named AI agents. Before building any agent, we built the measuring apparatus: a replay backtester and a paper-trading harness that share one risk gate, prevent look-ahead by construction, fill orders on the day after the decision, and charge trading costs measured from live quotes rather than assumed.

Experiments were run in order, each because the previous one left a specific objection standing, and the failures are reported alongside the rest. Where a configuration was chosen, the choice was written down before the run.

• Universes: 112 US large caps and 110 small caps, September 2016 – September 2026, daily bars.
• Capital: $100,000, simulated. No real money was traded.
• Limits: the conclusions apply to price-derived signals on these two universes over this decade; the paper's §6 lists the limitations in full.

Not investment advice. Nothing on this site recommends buying, selling or holding any security or strategy, and the author is not a licensed investment adviser. Past and simulated results do not predict future returns.

No tracking: the site sets no cookies and runs no analytics. Fonts are loaded from Google Fonts.

1. Treat a backtest as a claim to audit, not a result. Look-ahead, survivorship and free trading each manufacture returns on their own; the harness had to rule out all three before any number meant anything.
2. Costs and turnover are the parts you control. Every edge in this study was smaller than, or about the size of, a realistic trading cost.
3. Compare against the right benchmark. Hand-picked lists of today's companies beat the index by 150–226 points on hindsight alone; beating an equal weight of the same names is the bar.
4. A clever model needs new information, not more arithmetic. Thirty-one transformations of the same prices behaved like about six.

These are conclusions from a research exercise on simulated money, not advice for your situation.

Corrections, questions and replications are welcome: notes@shrov.com. Messages are read; replies may take a while. Please don't send personal financial details — this site can't give individual advice.