Use this process to review an AI trading agent in paper mode: define the version, collect simulated entries and skips, inspect risk behavior, score the journal, and change only one rule at a time.
Trading Boy does not execute live trades, hold funds, or provide financial advice. This process is for simulated AI agent review only and should not be interpreted as live bot deployment, signal delivery, or a recommendation to trade.
Run the review in the same order every time. The process should make weak evidence visible before a paper result can influence the next prompt change.
| Step | What to inspect | Review artifact | Decision rule |
|---|---|---|---|
| 1. Freeze the version | Prompt, persona, market universe, output format, and risk limits. | A version label and review window. | No mid-sample edits without starting a new version. |
| 2. Collect decisions | Entries, exits, skips, missed trades, and excluded records. | A complete paper journal. | Missing skips count as weak evidence. |
| 3. Check risk | Paper size, drawdown, exposure, frequency, and pause rules. | A risk-behavior row for each decision. | Rule-breaking wins do not pass review. |
| 4. Score output | Thesis, invalidation, setup name, rationale, and next question. | Structured output records. | Vague output triggers an output-format change. |
| 5. Review sample | Market regime, sample size, outliers, and version comparability. | Evaluation checklist and notes. | One next action, not a broad rewrite. |
| 6. Change carefully | The single behavior that needs improvement. | Prompt-versioning record. | Change one rule or keep collecting paper evidence. |
Version: A simulated AI paper agent runs version 1.3 for a four-week crypto momentum test. The prompt, watchlist, output format, and paper risk cap are frozen before the review window begins.
Evidence: The sample includes 25 entries, 18 skips, 8 exits from prior entries, and 4 excluded records caused by unclear market data. The journal links each decision to a setup name, invalidation, and risk note.
Finding: The agent followed the entry rule, but skipped-trade notes are too vague. Several skips say context unclear without naming the failed condition. Paper PnL is positive, but the review focuses on the missing evidence.
Decision: Keep version 1.3 archived. Start version 1.4 with one change: every skip must name the blocked condition from the written rule. Do not change the watchlist, risk cap, or setup definition yet.
This process prepares the evidence that the AI paper trading agent evaluation checklist scores. If the review process cannot produce complete entries and skips, the checklist should fail the sample instead of filling gaps with assumptions.
After the process identifies a repeated issue, record the change with AI trading agent prompt versioning. That keeps the next sample comparable and prevents hidden prompt drift.
The review should not rank an agent by paper PnL alone, rewrite the prompt after one winner or loser, hide skipped trades, or treat confidence as a reason to override paper risk. Those shortcuts make the agent look active, but they reduce the quality of the evidence.
The review should also avoid live-execution language. A clean paper review can justify another simulated test, a narrower output format, or a stricter risk rule. It cannot prove that a live bot would fill correctly, handle slippage, obey emotions, avoid outages, or produce future returns.
Choose the review cadence before the sample starts. A weekly review may work for a high-frequency paper workflow, while a slower swing-trade workflow may need a fixed number of decisions before the evidence is useful. The important rule is consistency: do not shorten the window because the sample looks good or extend it only because the first few outcomes look bad.
Each review note should include the version, start date, end date, market regime, number of entries, number of skips, excluded records, and the single decision made after review. If the agent changes output format, risk rule, or watchlist, start a new version before the next sample. That gives future reviews a clean comparison point instead of a mixed record.
It is a repeatable paper-trading workflow for reviewing AI agent decisions, skips, risk behavior, journal evidence, and version changes without live execution.
Review cadence depends on the sample, but the process should use a fixed window and avoid changing rules after every isolated paper result.
No. The process organizes simulated practice and review. It does not execute live trades, provide financial advice, or create buy or sell signals.