From a document to results
Once you add a document, the platform reads it and writes a research note with idea cards; from an idea card you pick the next step: write a strategy, mine a factor or run an event study.
Add a document
"Add a document" on the "Research" page takes three kinds of input:
- File: PDF with a text layer, plain text, Markdown, HTML. With "pictures" under "Includes" on Settings → Points, figures, tables, formulas and scanned pages in a PDF are read too (by a third-party service, marked apart from the text; see the Privacy Policy), and PNG, JPEG and WebP images can be uploaded directly; tick "Text only (do not read the pictures)" if you do not want that.
- Link: a public web page or PDF. Only that one link is fetched; pages behind a sign-in, a paywall or a bot check cannot be read, so download them and upload the file instead.
- Text: paste it in, with an optional title.
Uploading the same file twice does not read it again; you get the earlier note.
Notes and idea cards
A document opens with "Document" on the left and "Research note" on the right; on a narrow screen they become two tabs, "Note" and "Document".
The note holds a summary, the method details the source states (such as the sample period, trading costs and how many settings were tried), idea cards and the data needed. Figures in the summary that cannot be found in the source are marked.
Idea cards come in four kinds: scoring rule, trading rule, event and view. Each card quotes the source; click "page N, show" to jump back to that page. Cards also state how feasible the idea is and what would disprove it.
On a card you can:
- Set aside / restore
- Scoring rule cards: write a candidate formula, or edit it
- Trading rule cards: click "Build this strategy in the chat" to build it in a conversation
- Event cards: when the platform has data for that kind of event, test it with an event study
- Similar ideas: find similar idea cards in your other documents
To read again, click "Read for ideas again" and say what to look out for if you like; each reading makes a new note, and earlier ones stay.
Mine factors from a document
Mining tests a document's candidate formulas together and picks out the ones that pass the platform's threshold. You need "mining" under "Includes", see Tiers and points.
On one or more scoring rule cards with candidate formulas, click "test this idea", choose the candidates and click "Test these N", then:
- Set the terms: against the source's terms, you can change the universe, range and holding period; trading costs are set by the platform. The range is split in two: the first part is for selecting, the second is held out to check.
- Confirm: the confirmation card states how many formulas, the threshold and the estimated time. Nothing runs until you confirm, and the terms are fixed from then on. The confirmation card lapses after its deadline.
- Read the results: rows that pass the threshold come first and are built into factors that run the four gates; other rows can also be made into factors with "make a factor". Each row's details compare the source's claims with what the platform measured; once a row's factor has run the four gates, you can also ask the Agent to interpret it.
Papers on US stocks: once mining on US stocks is open, a paper measured on stocks is mined on them by default (the card's market row can change it): the pool us_top100 or us_top500, SPY as the benchmark, 5 basis points per trade, a month counted as 21 trading sessions. Every card states the effective sample: sessions × stocks and the independent holding periods. A candidate that needs data US stocks lack (the market capitalisation, the shares outstanding, financial statements, industries) is stopped on its own with the reason; size in the platform's neutralisation is the 30-session dollar volume, said on the card.
The more you try on the same stretch of the market, the higher the threshold. The held-out second part has to keep the same direction without clearly losing predictive power to pass.
The common marks in the results table are below. A row with one of the first seven marks needs the acknowledgement ticked when its factor is promoted. The other marks are only hints:
did not reach the adjusted threshold
With this many tried together, luck could produce this result
no better than a shuffled copy of itself
The scores, moved onto other contracts, did as well on the same market
not confirmed in the hold-out
On the data not used for selecting, the direction flipped or the power clearly shrank; probably overfit
no better than a copy with dealt returns
Time-series jobs only: with the later returns dealt out of order, it did as well
no better than a random score
Older jobs only: a randomly drawn score did as well
no better than a stale copy of itself
Older jobs only: the score shifted in time did as well
hangs on its window
One step longer or shorter in the window and the result disappears or flips
measured in the opposite direction
The side the source buys did worse than the other side
ranks and portfolio disagree
The ranking is significant but a long-short portfolio in that direction loses; read both together
not tested: too little data
Too few bars to get a t value; not ranked
"not looked at" in the hold-out column means the candidate did not pass the threshold, so its hold-out is not checked.
Event study
Research → Event study (crypto events only for now): line up one kind of event (perpetual listing, perpetual delisting, spot listing, spot delisting, token unlock) at the moment it happens and look at the cumulative return of the symbols in the days around it, and the excess return over BTCUSDT over the same days. You can set how many days to look at before and after the event, only take events after a given date, and whether day 0 is when the event takes effect or when it is announced. Results are grouped by event kind × year; each group gets a mean curve, hit rate and t value, and groups with fewer than 10 events only report a count.