AI Stock Screening: From Reproducible Screen to Watchlist
A practical method for combining numeric filters and document questions, then turning each surviving stock into a sourced, monitored watchlist row.
Published August 20, 2026 · Updated August 30, 2026
In this article
AI improves stock screening when it answers a defined question across a defined investable universe and returns a source for every surviving name. It does not turn a generic prompt into differentiated ideas. Use numeric filters to enforce liquidity, size, geography, and valuation constraints; use document questions for changes that lack a clean database field; then promote each candidate into a watchlist row with a reason, disconfirming condition, owner, and next event.
This is a public-source workflow guide, not a measured comparison of screening platforms and not an investment recommendation. We build AllMind and sell a grid-based research product, so the AllMind capabilities described below are our own claims; test them on your own universe before relying on them.
Freeze the universe and the cutoff
A screen cannot be reproduced if “the market” is its universe. Save the membership rules and the data cutoff before adding factors or document questions. A usable run record looks like this illustrative specification:
| Specification field | Filled example |
|---|---|
| Screen name and version | Pricing-power change, version 1.2 |
| Eligible securities | Primary common shares listed on the NYSE or Nasdaq |
| Geography rule | U.S.-domiciled operating companies |
| Size and liquidity | Market capitalization above $2 billion and trailing 30-day average daily value traded above $10 million |
| Sector exclusions | Banks and insurers, whose financial statements require a separate screen |
| Market and fundamental cutoff | Values available at the recorded run timestamp |
| Filing cutoff | Accepted EDGAR filings available before the same run timestamp |
| Corporate actions | Point-in-time shares, tickers, and adjustment factors retained with the run |
| Delisted securities | Kept in historical results when they were eligible on the original date |
| Normalization | Currency conversion uses cutoff-date rates; fiscal comparisons use issuer periods |
The last four fields prevent common look-ahead and survivorship errors. A company that disappeared after a failed thesis should not vanish from a historical screen. A restated value should not silently replace the number that was available on the original run date.
The SEC’s EDGAR submissions and XBRL APIs illustrate the issue. Company Facts returns standardized concepts reported by one issuer. Frames aggregates a concept across entities aligned to a calendar period, but the SEC warns that fiscal calendars can begin and end on different dates. Store the filing date, report period, unit, taxonomy concept, and accession with every fact used in the screen.
For macro inputs, ALFRED preserves historical vintages of FRED series, including values as originally released and later revised. That is the model to follow for every screen input: retain what the analyst could have known at the cutoff.
Write the screen as a specification
The specification separates eligibility, evidence, and ranking. It also gives the reviewer something more useful than a natural-language prompt to approve.
| Layer | Purpose | Example rule | Failure to record |
|---|---|---|---|
| Universe | Define what could be owned | Primary listing, liquidity floor, mandate geography | Names appear that the strategy cannot trade |
| Numeric eligibility | Remove obvious non-fits cheaply | Market cap, leverage, margin, valuation band | AI reads thousands of irrelevant issuers |
| Document test | Ask what structured fields do not capture | New risk-factor language, pricing commentary, customer concentration | Generic factor list repeats the terminal screen |
| Relationship test | Extend beyond the issuer | Customer, supplier, competitor, or regulator event | Read-throughs arrive after the market moves |
| Evidence gate | Require support for a survivor | Direct passage, period, filing date, confidence | Plausible answer passes without proof |
| Ranking | Order human review | Score dimensions and weights | Opaque model confidence becomes conviction |
| Promotion | Create a maintained watchlist item | Reason, trigger, kill condition, owner | The output remains a disposable ticker list |
Each rule needs an explicit treatment for missing data. “Missing equals fail” can remove early-stage or foreign issuers. “Missing equals pass” can fill the list with low-quality records. Keep missing as its own state and decide whether it routes to manual review.
Combine numeric and document questions
Numeric filters are fast, explainable, and useful. They also tend to return the same names to every user with the same data. Document questions can add a differentiated research angle, provided the question is specific enough to falsify.
A weak screen asks which companies have “strong pricing power.” It leaves the evidence, period, and meaning of pricing power undefined.
A reviewable alternative asks whether management described a realized price increase in the latest reported quarter. For every issuer, the result must include the exact passage, product or segment, reported period, and any volume or mix offset in the same source. When the evidence is absent, the row must say “not found” rather than infer an answer.
The second question still needs human interpretation. It prevents the model from treating an aspirational pricing statement as a realized result and forces it to carry offsets that may reverse the conclusion.
Other document questions that fit a universe pass:
- Did the latest 10-Q add a risk factor or materially expand an existing one? Show the changed passage and prior-period comparator.
- Which issuers disclosed that one customer represented a material share of revenue? Return the exact threshold, period, and whether the customer was named.
- Which management teams changed formal guidance and which only changed descriptive language?
- Which companies reported a restructuring charge and also disclosed the expected cash component?
- Which issuers changed a segment definition, KPI definition, or non-GAAP reconciliation?
The SEC Full-Text Search and filing search tools are useful for designing and checking these questions. A production screen may use licensed data or a research platform, but the public filing remains the source of record for a filing-derived claim.
Make ranking math visible
Do not rank on model confidence alone. Confidence usually describes the system’s answer, not the investment attractiveness of the company.
A review score can use separate dimensions. One illustrative allocation is:
| Dimension | Weight |
|---|---|
| Thesis relevance | 30% |
| Evidence quality | 25% |
| Expected model impact | 20% |
| Novelty versus current team knowledge | 15% |
| Time sensitivity | 10% |
The weights above are illustrative. Publish the actual weights with the run. Keep the underlying dimension values beside the total and log any manual override. A PM may promote a low-confidence, high-impact signal precisely because uncertainty is the reason to investigate.
CFA Institute Standard V(A) explicitly applies its diligence and reasonable-basis requirements to computer-generated screening and ranking. Its guidance says users should understand model parameters and limitations and make reasonable efforts to test output before using it in analysis or recommendations. A ranking that cannot be reconstructed fails that practical test.
Promote survivors into a watchlist record
The promotion step is where idea generation becomes an operating process. Copy this schema into a database, spreadsheet, or research system:
| Field | Required content |
|---|---|
| Security and issuer ID | Ticker plus durable identifier such as CIK or vendor entity ID |
| Screen run | Name, version, cutoff, universe, and rank |
| Why it survived | One cited sentence tied to the screen criteria |
| Evidence status | Observed, calculated, estimated, inferred, or missing |
| Initial question | The specific diligence question that remains open |
| Thesis candidate | A falsifiable claim, clearly labeled as analyst work |
| Trigger | Filing, event, threshold, or date that merits another review |
| Kill condition | Evidence that removes the name from the watchlist |
| Expected next event | Source and expected window |
| Owner and review date | Named analyst and deadline |
| Decision history | Promote, retain, pause, reject, or archive, with reason |
A watchlist row without a kill condition accumulates indefinitely. A row without a source cannot explain why it entered the list. A row without a next event cannot become monitored coverage.
Test the screen before trusting it
Use three evaluation sets.
Known-positive set
Choose historical filings where the target disclosure definitely appears. Check whether the system retrieves the correct passage, issuer, period, and meaning. Include tables, footnotes, amendments, and unusual fiscal calendars.
Known-negative set
Include documents containing similar words without the target fact. A risk factor mentioning possible price increases should not pass a screen for realized pricing. A supplier named in a general policy should not pass as a material commercial relationship.
Live blind set
Run the screen on a current, frozen universe. Analysts review the top and bottom samples without seeing the model score, then compare judgments. Track precision at the review capacity the team actually has, as well as the rate of unsupported answers and material misses.
Record retrieval failures separately from “not found.” If a filing was inaccessible, parsing failed, or a table was omitted, the system has not established absence.
Control for idea crowding and confirmation
A screen is a prioritization mechanism. It does not establish that an observation is mispriced. After promotion, ask:
- Is this information public, and since when?
- Which expectation or valuation input does it differ from?
- Is the evidence new, or has only the wording changed?
- What alternative explanation fits the same facts?
- What would make the opportunity disappear before the next event?
Keep the screen author away from the first blind review when possible. A model that explains why each result matches the prompt can reinforce the researcher’s initial framing. Give a reviewer the source passages before the generated rationale.
Evaluate tools by the layer they own
Numeric screeners and market-data terminals are appropriate for eligibility filters and standardized fields. Filing search products can answer document questions. Relationship datasets can add supplier, customer, and ownership paths. General assistants can help refine a prompt or review a supplied result set, but they need an explicit universe and controlled sources before they can run a reproducible screen.
AllMind combines Data Viewer fundamentals, estimates, live market data, comps, and supply-chain relationships with 6,800+ premium data sources from 100+ providers and partners. Grids runs a question across a company list with cited answers per row, and the ontology resolves issuer and relationship data. Screening is therefore not limited to user-uploaded documents or narrative extraction. Those capabilities rest on our own licensed data and partnerships, and they are our claims to prove, not neutral findings. AllMind is quote-priced and requires onboarding for internal notes and positions. A buyer should trial the screen specification above, preserve every failed retrieval, and compare the output with its known-positive and known-negative sets.
For member firms, FINRA Regulatory Notice 24-09 notes that existing supervisory obligations continue when generative AI is used. The right approval, retention, and monitoring design depends on the firm’s business and jurisdiction.
The final output should be small enough to review. A screen that returns 400 “high-conviction ideas” has skipped the ranking and promotion work. A good run leaves a reproducible audit package, a handful of sourced candidates, and clear reasons for every rejection.
Sources and methodology
This guide uses official SEC, Federal Reserve Bank of St. Louis, CFA Institute, and FINRA sources accessed on August 30, 2026. No screening platform was run and no security results were generated. The score weights and schema are editable evaluation artifacts, not performance claims. Data coverage, product limits, and supervision policies should be verified again before a live screen.