Economic research involves reading releases, following revisions, connecting related documents and keeping notes. Language models can assist with that work. The research record still needs a clear boundary between source evidence, an interpretation and a calculated result.
Assist with the reading
We are exploring AI assistance for identifying the relevant parts of economic releases, organising documents by topic and drafting concise notes for review. A draft should retain references to the source material behind its factual statements.
A useful assistant can surface a change in wording or an apparent disagreement between documents. That is a prompt for a researcher to inspect the evidence, not a reason to accept the interpretation automatically.
Keep numerical transformations explicit
Calculations such as percentage changes, standardisation, lag construction and forecast evaluation belong to recorded numerical procedures. The inputs, formula and output should be recoverable without relying on a language model to repeat an explanation.
Where a note includes a number, the workflow should identify whether it came directly from a source, from a defined calculation or from an estimate. These categories should not blur together in a polished summary.
Separate a forecast from a narrative
An economic narrative can explain why a model result deserves attention, but it is not itself a validated forecast. Our research design keeps model outputs, uncertainty and statistical evaluation separate from generated commentary.
If an independent model forecast is compared with published consensus, the comparison should preserve both sources and their timestamps. Consensus should not silently become an input to something described as an independent forecast.
Build review into the workflow
Before a research note is used, verify the cited publication, observation period, units and revision status. Check whether the note refers to a released value or a forecast and whether its interpretation is consistent with the document.
The goal is to reduce the time spent collecting and arranging material while leaving meaningful judgement with the researcher. A source that cannot be verified should remain unresolved in the record.
Keep trading rules accountable
Generated prose should not quietly change a trading rule. A new rule or model revision needs a versioned specification and an evaluation process. The operational system should be able to explain which rule and configuration produced a decision.
This boundary also makes errors easier to diagnose. A problematic source interpretation, a calculation error and an execution failure require different corrections.
Our current development direction
Beyond Markets is researching and developing fundamentals bots and trading systems using eighteen years of historical data. We plan to explore language-model assistance within that broader research process.
The examples on this website explain the intended workflow. They do not demonstrate a deployed AI integration, a live trading service or an independently validated performance record.
Automation is useful when it makes the evidence easier to inspect and the research easier to reproduce.
This note describes research principles and development objectives. It is not a report of a validated model or live trading performance.