Research & Data

Terminal chat can combine market research, chart context, and supported datasets. Dataset selection narrows which broad data families the assistant should use; it does not guarantee that every field exists for every company, asset, date, or plan.

Dataset selector

The chat composer exposes four dataset families, selected by default:

The in-product dataset information view highlights current use cases and coverage such as prices and technicals, fundamentals and SEC filings, insider and U.S. Senate activity, analyst targets and ratings, institutional ownership and 13F filings, economic and macro data, and news.

Ask better research questions

Specify the symbol or universe, date range, comparison basis, source family, and desired format. Ask for citations or the data timestamp when recency matters. Separate a request for factual research from a request to create or modify a trading strategy.

Examples:

• “Compare revenue growth and margin changes for these three companies over the last eight quarters.”
• “Show the latest analyst rating changes for AAPL and state the effective dates.”
• “Find recent insider purchases, then chart the publication dates against price.”
• “Build a rule from this event only after we inspect the event timing and fields.”

Timing and coverage

Alternative and event records can be published after the underlying real-world event. Backtests must use the time the information became available to avoid look-ahead bias. Revisions, delayed filings, symbol changes, and incomplete historical coverage can also affect a result.

Research before strategy generation

Use a two-stage conversation for consequential ideas. First ask for facts, dates, definitions, and uncertainty. Then decide which information can be represented as a testable rule. This makes it easier to distinguish a missing data field from a strategy-generation mistake.

For example, “find companies with improving margins” is a research request. “Enter on the first daily bar after a qualifying filing becomes available” adds timing, assets, and an action rule. The second request still needs an inspectable definition of “improving” and a publication timestamp.

Source-aware questions

Ask Astral to state:

• the effective or publication date of each observation;
• the period the value describes;
• whether a figure is reported, estimated, derived, or summarized;
• the units, currency, and adjustment basis;
• which requested symbols or periods have no result;
• whether the answer combines providers with different update schedules.

A well-formed answer should preserve missingness. Converting “not returned” into zero can reverse a ranking or create a false strategy signal.

Common research traps

Look-ahead bias

Using an earnings value, filing, analyst change, or transaction before it was published makes a historical test unrealistically informed. Base the rule on the availability timestamp, not only the period the record describes.

Survivorship and universe drift

A current symbol list can exclude delisted or renamed instruments. Testing today’s universe over an earlier period can make results look cleaner than the decisions available at the time.

Revision and restatement

Macro series, financial statements, estimates, and classifications can change. A current historical series may include revisions that were not known on the original date.

Correlation presented as explanation

A chart alignment does not prove a dataset caused a price move. Use research to formulate hypotheses, then test clearly defined behavior with appropriate controls.

Turning findings into reusable work

Save the strategy only after the data rule, timestamps, missing-data behavior, and fallback logic are clear. Record the research window and assumptions in the description so a later revision can be compared honestly.