Three stories, one mechanism
Last week was not about a new charting feature. It was about how automated systems share a view of the same public objects — a Fed speech, a futures strip, a research prior — and then move together.
This post is educational. It is not financial advice and it does not promise returns.
Fed odds are a traded price, not a poll
CME FedWatch is a translation of 30-day Fed funds futures into implied probabilities for the next FOMC. When Fed Governor Christopher Waller pushed back on a September hike, those implied odds dropped on the order of 12 percentage points within minutes — coverage puts the move from around two-thirds toward the mid-50s. Later prints (labor data) pushed the path back toward the high 50s into 7 September.
The useful point for an algo trader is not which number to believe. It is that “hike odds” are a live order book. A language change on inflation is an input; futures inventory is the output. If your system treats a probability headline as a slow, survey-like state variable, you are late relative to anyone trading the contract itself.
The September FOMC is 15–16 September. CPI/PPI still sit in front of that meeting. Expect another fast repricing. Size for the gap, not for a narrative.
Safeguards built for glitches, not for systems that work
On 3 September, CFTC international-affairs director Mel Gunewardena spoke at George Washington University. His own remarks (not necessarily the Commission's) described markets moving beyond electronic and algorithmic trading into continuous, programmable, composable, increasingly autonomous structure.
The line that matters for operators: existing circuit breakers and price limits were designed for computer-era failures — bad ticks, runaway loops, fat fingers. A future dislocation, he argued, may come from autonomous systems functioning as designed, possibly starting in a thin overnight or weekend venue and only then printing into the regulated open.
If you run session-aware automation, that is an ops constraint, not a slogan. Kill switches that assume a human is at the desk between Friday close and Sunday futures are the same class of mismatch: the code is doing what you asked; the calendar is not the one the control was written for.
Better models can still crowd the same trade
A 3 September arXiv paper, _Why Better Models Can Create Riskier Systems_, tests LLM traders in an agent-based market. The result is a capability paradox: more capable general-purpose models can become more correlated with each other. When the shared reasoning is right, adding agents can dampen market-level risk. When they share a misinformation environment, the same correlation becomes a non-diversifiable floor.
You do not need to run LLMs in production for this to bite. Any stack that consumes the same headlines, the same vol surface, and the same risk-off playbook has a crowding problem. “My signal is original” is an empirical claim. Correlation in the book during a Waller-style minute is the test.
Durable lesson
Treat public macro as a price process, treat other algos as potential copies of you, and treat safeguards as session-limited. Concretely:
- If a policy probability can move 10+ points on a speech, your max position into that speech is a risk parameter, not a conviction parameter
- Overnight / weekend / 24-7 venues can gap into your Monday open; assume they might
- Diversification across models that share data and training priors is weaker than it looks
None of that requires a particular terminal. It requires a view of the market as a system of coupled automata, some of which you do not control.
Soft next step
If you want a dry check on copy-path sizing after a vol spike, the free copy ratio and lot size calculators are there. They do not replace exchange specs, firm rules, or a kill switch.