Week 2: Lessons on Correlated Signals and Fallbacks
Design lessons on overlapping signals, fallbacks, and what a model should see. A lessons post, not a results report.
Simulated paper trading, not live results; not financial advice. No real money was traded. This post makes no weekly performance claims. A later review of the ledger and decision logs we kept found that the simulated run (Feb 26 to Apr 8, 2026) ended net negative; the figures and their caveats are in the retired-project case study. The bot has since been retired; see why we retired it.
Lesson: Confidence Per Signal Is Not Portfolio Risk
A common failure mode in signal-driven bots is that one strategy produces several overlapping signals on the same asset. Each signal can look well-supported on its own, yet together they are one larger, undiversified bet. The lesson is that per-signal confidence does not measure correlation, and a bot needs a separate check on overlapping exposure.
Typical mitigations include blocking a new signal on an asset that already has an open position or a queued signal, capping total exposure per asset, and adding a cooldown after a closed trade. We describe this class of control as a design principle; we are not reporting that a specific incident occurred in our run.
Lesson: Test Fallbacks Under Load
If a bot depends on an LLM provider, plan for slow or failed calls with a fallback chain, and test that chain under realistic load rather than only in isolation.
Lesson: Sanitize What the Model Sees
Contract titles and other market text can contain probabilities or other numbers that anchor a model's own estimate. Stripping that data before the model forms its estimate is a sensible design rule.
What We Are Not Claiming
We make no claims here about specific weekly results. For the whole run, the ledger and logs show a net loss in simulated paper trading; see the retired-project case study for the numbers and caveats.
See Risk Management Rules We Hard-Coded and the critic agent post for the design.
Related Reading
The trading bot has been retired. These posts cover the design and what we concluded.