What Matters
An AI trading journal should record more than entry, exit and P&L. To audit an autonomous trade, you need to know which model/version acted, what authority it had, which input prices it saw, which quote it expected, what actually executed, every trading cost and whether a human override or safety gate changed the result.
DN Evidence Block
Last reviewed: 15 November 2026.
Basis: DN’s existing AI trading bot live-results audit framework, execution-quality work and agent permission research.
Boundary: this article proposes an operational record format. It is not a claim that every exchange or bot exposes every field automatically.
The 12 Fields
| # | Field | Why it matters |
|---|---|---|
| 1 | Agent / model version | Separates performance before and after a strategy/model change. |
| 2 | Strategy ID | Prevents unrelated tactics from being averaged together. |
| 3 | Decision timestamp | Anchors the inputs the agent could actually see. |
| 4 | Signal / rationale code | Records what triggered the action. |
| 5 | Authority state | Shows permitted symbols, leverage, notional and whether human approval was required. |
| 6 | Reference/oracle input | Lets you test whether the decision was based on stale or abnormal data. |
| 7 | Expected quote | Creates the benchmark for slippage and rejection analysis. |
| 8 | Actual execution | Price, size, fill time, venue and order ID. |
| 9 | Fees and funding | Prevents gross P&L from masquerading as net performance. |
| 10 | Exit reason | Target, stop, liquidation, timeout, kill switch or human override. |
| 11 | Net realised P&L | What remained after all trading costs. |
| 12 | Incident / override note | Captures unusual behavior, intervention and control failures. |
Why Versioning Matters
If a bot changes prompts, model provider, risk gates or execution code, the performance series has changed. A journal that only says “AI bot” cannot tell you whether an apparent improvement came from the strategy, a new model, different leverage or simply a better market regime.
Gross Returns Are Not Enough
Record:
Net P&L = realised trading P&L − commissions − spread/slippage − funding − borrow cost − other execution charges.
DN’s prior live-results work uses this same principle: reconstruct the economics from the actual trade record rather than trusting a dashboard headline.
Record Rejections Too
A rejected order is part of the strategy’s real performance. If the agent repeatedly chooses trades the venue will not accept, that is a production failure even when no realised loss appears in the account.
Maintain a separate rejection log with reason codes such as insufficient margin, invalid size, price protection, stale nonce, rate limit or venue downtime.
DN AI Trading Journal Completeness Score
Enter how many of the 12 required fields your current journal captures for every trade.
This is an editorial completeness score, not a performance score.
Review Cadence
Daily: control failures, rejected orders, unexpected leverage, broken stop behavior.
Weekly: slippage, fee drag, win/loss distribution, strategy drift and human interventions.
Monthly: net performance versus benchmark, maximum drawdown, capital efficiency, model/version comparison and whether the agent still deserves its current authority.
What a Good Journal Lets You Answer
- Did the model make a bad decision or did execution fail?
- Did costs erase the strategy edge?
- Did a new model improve net results or just change risk?
- How often did humans override the agent?
- How many proposed trades were rejected?
- Which input data was present when the trade was initiated?
- Which control ended the position?
Frequently Asked Questions
Should I log every AI-agent trade?
Yes if you want an auditable performance history. Sampling can hide execution failures and rejected orders.
Do I need the model version?
Yes. Otherwise performance before and after model or strategy changes gets mixed together.
Why log expected quote and actual fill?
The difference reveals slippage and execution quality rather than attributing all outcome variance to the strategy.
Should paper trades and live trades share one journal?
They can share a schema, but they should be clearly separated because fills, latency, liquidity and costs differ.
Final Verdict
A trading journal turns “the bot made money” into a testable statement. Without version, authority, input, execution and cost records, an AI trading history is mostly a sequence of outcomes with no reliable explanation for why they happened.
Related DN Research
