LEARN / TRACK 2
Rules, models and agents: who actually decides?
"Does it use AI?" is the wrong question. Ask who decides, from what information, under what limits, and how you’d know it helped.
Start with a rule you can inspect
A deterministic rule gives the same output for the same inputs and version. For example: if the current price is above an average and the paper budget is free, queue an order; otherwise wait. You can replay each decision. A rule can still be badly chosen, fed bad data or expensive to execute. The SEC uses “algorithmic trading” for a broad range of automated processes, not only artificial intelligence (AI). SEC ↗
A learned model estimates; it does not know
A statistical model fits patterns in past data, then produces a forecast, score or classification. If its designer tries many settings and keeps only the best-looking one, it can overfit noise. When market behavior changes, a past relationship may fail. Backtest-overfitting research documents this selection problem. SSRN ↗
An AI-managed paper portfolio would let a model propose allocations under a fixed budget and risk policy. CouchChange has no such running portfolio today. This describes a possible separate experiment, not current performance.
An agent adds a work loop
An agent can choose a sequence of steps, call tools, inspect outputs and revise a plan. More tool access means more ways for a mistake or hostile text to reach an action. NIST documents confabulation and information-security risks for generative AI. A system should treat a news article or token description as data, not as permission to trade. NIST ↗
In a future paper experiment, hard limits would live outside the language model: maximum size, permitted markets, fresh-data checks, a kill switch and human review for material changes. CFTC discussions of automated-trading controls support keeping risk checks independent of a strategy's prose. CFTC ↗
Compare the decision chain
| Question | Fixed rule | Learned model | Tool-using agent |
|---|---|---|---|
| Who decides? | Written if/then logic | Trained scoring function within a policy | Model plans steps; external policy accepts or rejects actions |
| What does it see? | Specified current inputs | Specified features, then new inputs | Tool outputs, prompts and possibly untrusted text |
| How to test? | Replay checks; freeze and paper-test | Separate training and later evidence | Audit tool calls, policy decisions and later evidence |
| How can it fail? | Bad rule or data | Overfit or changed conditions | Those, plus invented claims or prompt injection |
These are design patterns for teaching, not a ranking of trading performance. The CFTC warns that AI cannot predict the future or guarantee returns. CFTC ↗ Investor.gov ↗
Sources
Examples marked made-up numbers are invented. Nothing in this lesson is a recorded CouchChange result or a recommendation.
Try it
A model says “buy” but a fixed position limit says “too large.” Which instruction controls the order?
Think it through, then open an answer
The independent position limit rejects it. A persuasive model explanation is not authority to change the limit.