The negative result
TradeBot
A trading system measured until it proved itself wrong. The negative result became the rule that nothing ships without a measured edge.
- Problem
- A large trading stack could not connect to a broker; its tests passed only because every test used a mock.
- Constraint
- Measure before automating: costs, spread, look-ahead and a frozen holdout before any claim.
- Decision
- Rebuild around a committed measurement artifact: as-of slicing, real per-bar costs, limit entries required to be touched, and a risk envelope that cannot be widened at runtime.
- Result
- The committed run covers 1,074 positions at 38.64% win rate, -0.4868R average, profit factor 0.4183, with a 217-position holdout. Negative in every fold.
Limits
The 3,271-position aggregate quoted in the README is not committed as a raw log; only the 1,074-position run in state/measurement.json can be reproduced today.
- Reproduce
- python scripts/measure.py, then compare state/measurement.json.