What an AI paper trading test should show
A useful test records what the workflow received, which rule fired, and what happened next. A changing virtual balance alone does not explain whether the automation followed your instructions. Start with a small workflow whose decisions you can inspect, then add more assets or model steps after the basic path is understandable.
Paper trading uses current market inputs and simulated execution. Historical simulation asks a different question: how a rule would have behaved on past data. See paper trading versus backtesting for that distinction. Neither establishes how a future live account will behave.
Walkthrough: a model-consensus paper workflow
The existing Multi-LLM Consensus Trader page shows its node graph and decision stages. Use it as a specification to inspect and adapt. The walkthrough below describes its configuration and the evidence to collect; it is not a report of measured trading results.
- Open the strategy. Read the input, model-review, and order stages before starting. Continue with the strategy context, then ask Nick to create an inactive draft that uses the paper exchange. Review the draft before running it.
- Check every account destination. The portfolio input and order node should both use your intended virtual account. A workflow can contain multiple connections, so inspect each one rather than relying on its name.
- Inspect the inputs. Confirm the symbols, lookback, timestamps, and missing-data behavior. If a feed is unavailable, the workflow should record that state and skip the decision rather than substitute an invented price.
- Make agreement explicit. Inspect the model prompts, required output fields, and consensus condition. Define what happens on disagreement, invalid output, or a timeout. Review any sizing and order checks in the workflow itself.
- Run manually and inspect the log. Check the data output, each model response, the conditional branch, and any simulated order receipt. The execution guide explains where to inspect node results.
- Schedule only after review. Choose a supported cadence and confirm the timezone. Estimate credit use from observed runs, especially when the workflow contains several model calls.
A review record you can copy
For each run, record: run ID, timestamp, symbols, source timestamps, model decisions, gate outcome, account mode, order status, and notification status. These are review fields, not a claim that every template emits one combined object.
An illustrative no-trade record could read: source status: complete; model agreement: no; order branch: skipped; account mode: paper; notification: delivered. Replace those example values with the actual node outputs. A skipped order can be the correct result.
Cover the branches, not just the calendar
Review an agreement case, a disagreement case, a missing-input case, and a rejected-order case. Keep a record of configuration changes so a later run can be compared with the version that produced an earlier result. A quiet week with no triggered condition does not test the order path.
Paper environments make different assumptions about fills and liquidity. For example, Alpaca documents simulation limitations including market impact, latency-related slippage, and queue position. NickAI's paper exchange and an Alpaca paper account are separate environments; test the actual destination you intend to use.
For an equities example, follow the Alpaca stock workflow walkthrough. For the model gate, see how to configure and evaluate consensus.