AI-assisted investing · paper trading
AI-assisted investing lab: what a six-role paper trading desk can test
Six AI roles discussing a trade can sound like a small investment team. More roles, however, do not make the data accurate or cap losses automatically. The useful part of a Paper Desk is a repeatable record: the hypothesis, opposing evidence, risk budget and outcome can all be examined before real money is at stake.
The short version
- AI can organize candidates and news leads. A person still needs to verify prices, dates, filings and sources.
- Six roles are a review workflow, not six independent forecasting models or six layers of guaranteed safety.
- Alpaca Paper is a simulation. Paper fills and live fills may differ, and a paper return cannot predict a live return.
- Start with ideas only, then test paper orders and keep a full log. Evaluate any real-money decision separately.
What is the experiment?
Yi Hang's Short introduces a Grok AI trading team. The accompanying Paper Desk note on Substack describes six sequential roles: Scanner, News, Entry/Exit, Risk, Skeptic and Desk Lead. Its stated scope is long-only US equities in an Alpaca paper account. This article explains the research method. We have not independently audited the program described in the note or verified that its code enforces every stated limit.
Scanner proposes a candidate. News checks the event and its timing. Entry/Exit writes the hypothetical entry, invalidation and exit conditions. Risk calculates possible loss and concentration. Skeptic searches for missing evidence and counterarguments. Desk Lead records whether to wait, reject or test the idea in simulation. If one model plays all six roles, shared blind spots remain possible. The skeptical role needs checkable evidence.
Five checks before a paper order
- Define the question: What asset, time horizon, purpose and no-trade condition are being tested? Keep living reserves separate with the three-bucket framework.
- Verify the facts: Trace a headline back to a company filing, regulator or exchange. Record when it was published. Missing or stale evidence should produce WAIT.
- Write the risk: Use the hypothetical entry, invalidation price and share count to estimate loss. Check portfolio concentration. A percentage cap is an experiment setting, not a universal safe level.
- Ask for disconfirming evidence: What would break the thesis? What about liquidity, spreads and execution? The spread explainer shows why displayed prices need care.
- Keep the full record: Save rejected ideas, source snapshots, paper orders and later outcomes. A log containing winners alone makes review misleading.
What do the proposed limits mean?
The Substack note describes hard-coded experimental limits: risk per idea at or below 0.5% of paper equity, no single position above 10%, and no averaging down. These are Yi Hang's chosen test parameters. They cannot guarantee a maximum actual loss: gaps, poor liquidity, bad data, software faults and market halts can invalidate an assumed exit. Even in paper trading, examine the logs, inputs and rejection rules instead of looking only at an equity curve.
At US$20,000 of hypothetical paper equity, 0.5% is US$100 and 10% is a US$2,000 single-name position cap. The first number is a planned risk budget; the second is a position-size limit. Neither guarantees that a real-world loss stops at US$100.
What paper trading cannot show
Alpaca says its paper environment simulates orders but does not fully capture market impact, latency-related slippage, queue position, some fees or dividends. Paper and live accounts use different credentials and endpoints. Alpaca identifies paper-api.alpaca.markets as its paper endpoint. A dashboard label alone does not prove which environment the running program has connected to; inspect the actual configuration and credentials. See Alpaca's paper trading and authentication documentation.
The original note proposes an ideas-only first week, followed by paper-order testing and a reviewable simulation log. This can reduce operational mistakes but even a month of paper results is not live validation. AI API costs change with model, volume and current pricing, so check the provider's present billing and limits rather than treating the note's monthly estimate as a quote.
Three human responsibilities
- Check primary evidence: An AI summary can omit dates, conditions or contrary facts. The SEC's investor alert warns against relying solely on AI-generated investment information.
- Separate money by purpose: A paper account, long-term capital and living cash have different jobs. Start with the cash-flow checklist.
- Review exits and failure modes: If the system cannot explain why it entered, what would invalidate the idea or how it stops after an error, the appropriate output is no trade.
Here, “AI-assisted investing” means a tool for research, challenge and record-keeping. It does not promise a return or assume responsibility for the reader's trades. Be skeptical of claims that AI can deliver high returns with little risk; the joint SEC/FINRA/NASAA investor alert explains why.
Watch, read and verify
Watch Yi Hang's six-role Short · Read the original Paper Desk note
Sources checked 5 Nov 2026: Alpaca Paper Trading · Alpaca Authentication · Investor.gov AI and Investment Fraud. The six-role design and numerical test limits are the author's description in the Substack note.
General education and technical discussion only. This is not personalized investment, financial, legal or tax advice. No security or real-money trading configuration is recommended. Paper results do not predict live results; live trading can lose principal. External links are not affiliate links.