What AI Software Should a Long/Short Equity Fund Use?
A decision brief for long/short equity funds choosing AI research software: the two-book jobs to cover, a sourced vendor matrix, and a pilot that tests both books.
Vanessa Voss
Published September 3, 2026
In this article
Buy by job, for two books. A long/short equity fund needs AI software for six jobs: screening and idea generation, thesis monitoring against a written long or short case, short-interest and borrow data, factor and crowding hygiene, a catalyst calendar spanning both books, and permissions with an audit trail. Start with the research system that joins filings, transcripts, estimates, ownership, expert content and news with the fund's own thesis ledger, and pilot it on one long and one short. The OMS, borrow desk and factor model stay in place.
This decision brief is built from public product pages, regulator documents, news reporting and SEC filings opened on September 3, 2026; we did not test any product. We build AllMind, which appears below as a candidate, so the conflict of interest is material. Our claims are first person and trace to our live pages; competitor claims are vendor-reported with dates; both deserve the same two-book pilot.
Why the short book changes the requirement
A short can be right on the fundamentals and wrong on the positioning: borrow gets expensive or recalled, the holder base is crowded, and a factor move squeezes every short in the sector at once. The data that decides those outcomes (short interest, borrow cost, holder concentration, factor exposure) does not live in filings, so software chosen for the long book alone never sees it.
The official short-interest series is slower than vendor pages imply. FINRA collects member short positions twice a month, at settlement on the 15th and on the last business day, and publishes the compiled data on the seventh business day after each settlement date, per FINRA's equity short interest page. Anything faster is a vendor estimate modeled from securities-lending data; S3 Partners and Ortex sell such estimates, and the matrix below carries their claims and prices.
Positioning hygiene stopped being optional this summer. Per CNBC on August 21, 2026, Goldman Sachs called July 2026 the worst month of hedge-fund underperformance against the S&P 500 in more than 20 years of its data and one of the sharpest de-grossing episodes of the past decade. Crowding had hit a record in the second quarter, and US equity long/short funds were still up 10% through mid-August. A research system will not compute factor exposure and should not try; it has to read the factor model's output and the crowding data next to each thesis, so a crowded short is flagged before the de-grossing day.
When both books report in the same week, the desk needs the pre-print brief, the guidance diff and the broker-by-broker read for both sides. The requirement is six jobs, each in a long-book and a short-book version:
- Screening and idea generation: a reproducible screen over a frozen universe, run forward for longs and inverted for shorts, following our screening and watchlist guide.
- Thesis monitoring: a written long or short case with named pillars, checked against each new filing, transcript and headline, with the claim ledger as the record.
- Short-side data: the official FINRA print and the vendor estimate side by side, both dated.
- Factor and crowding hygiene: the risk system's exposures and the crowding data read next to each name.
- A catalyst calendar across both books, with the guidance-change mechanics applied to shorts as carefully as to longs.
- Permissions and audit: who saw which short thesis, expert call and broker note, logged.
The vendor matrix, by job
The matrix lists the candidate tools we could document for each job from public pages, with the price where one is published; cells that say "our" or "ours" describe AllMind, and no product was tested.
| Job | Candidate tools | What the public page documents | Evidence status | Published price | Last checked |
|---|---|---|---|---|---|
| Screening and idea generation | Koyfin; our Grids | Koyfin Plus: 10 years of financials, estimates, transcripts, a screener; our Grids: one question across a ticker list, a cited answer per cell, Excel export | Koyfin observed; ours first-party | Koyfin Free $0, Plus $39 per month, Premium $79 per month; AllMind quote only | 2026-09-03 |
| Thesis monitoring and document reading | Hebbia (Max, Matrix); our Agent Studio | Hebbia lists FactSet, S&P Capital IQ, PitchBook, SharePoint and Box integrations; our agent checks a stated thesis against each new filing, transcript and headline | Hebbia vendor-reported; ours first-party | Quote only | 2026-09-03 |
| Short interest, borrow and crowding | FINRA (official); S3 Partners; Ortex | FINRA positions twice monthly (linked above); S3 claims "daily short interest across 65K+ global securities" and disclosure coverage across 100+ countries; Ortex Advanced adds real-time short interest and cost to borrow | FINRA observed; S3 vendor-reported; Ortex prices observed | FINRA free; Ortex Basic $49 per month, Advanced $149 per month; S3 quote only | 2026-09-03 |
| Factor and crowding hygiene | Axioma (SimCorp); MSCI Barra; Arcana Analytics | Axioma markets equity factor risk models, attribution and stress testing to hedge funds among others; MSCI's Barra pages could not be opened; Barra and Arcana appear on our data catalog as integrations by scope | Axioma vendor-reported; MSCI unverified; ours first-party | Quote only | 2026-09-03 |
| Catalyst calendar and live events | Aiera; our earnings triggers | Aiera claims "50k+ events tracked" and "15k+ global equities monitored"; our triggers hang off the earnings calendar with the lead time the desk chooses | Aiera vendor-reported; ours first-party | Quote only | 2026-09-03 |
| Channel checks and expert content | AlphaSense Expert Insights; Third Bridge; our Expert Insights | AlphaSense claims "300,000+ investor-led insights" across "29K+ companies"; Third Bridge claims "100,000+ expert transcripts" and "75,000 companies"; our catalog lists Third Bridge Expert Insights, 100,000+ transcripts, included in the seat | Both vendor-reported; ours first-party | Quote only; AlphaSense advertises "transparent pricing" without a figure | 2026-09-03 |
Only Koyfin and Ortex publish a price; every other row, ours included, is quote-based. None of the research assistants documents borrow or factor data as a native feed, so that layer stays with the risk vendor or the prime.
Decision branches by fund profile
Concentrated fundamental long/short
A desk running a few dozen longs and a smaller short book from written theses should pilot the research system first, on one long and one short with a print inside the pilot window. The prime's risk report serves as the factor layer, and borrow data is added when intraday borrow starts changing sizing decisions. The test that matters is the thesis-monitoring row of the acceptance table below, run on both names.
Pod inside a multi-manager platform
The platform owns identity, entitlements, logging and approved data paths, and the pod configures inside them; the pod stack covers that layer. The long/short-specific addition is scoping: a short thesis is the most sensitive document a pod holds, so the research system has to prove that an agent run by one pod cannot surface another pod's thesis, and that the denial is logged.
Quantamental
If the alpha model is statistical, the factor model and the data APIs come first and the research system becomes a feature factory: guidance-language changes, estimate revisions and supplier read-across as structured rows with the source passage attached. The acceptance test is an export with a cited cell that the quant team can join to its own tables without anyone retyping numbers.
Event-driven overlay
An event book adds large document sets (merger material, amended agreements, court filings) and probabilities. Hebbia's homepage lists SharePoint, Box and IntraLinks among its connectors, which is where such a set usually lives, and our catalog carries Kalshi and Polymarket implied probabilities beside the earnings calendar, so a deal or regulatory catalyst sits on the same calendar as a print. The hedge-fund pilot design already covers the event-driven and credit case.
The proof artifact: a two-book pilot acceptance table
Run the pilot on both books at once, with the same names for every vendor. A vendor passes a row when the evidence in the last column exists, with timestamps, for both tests.
| Job | Long-book test | Short-book test | Pass evidence |
|---|---|---|---|
| Screening | Re-run the fund's long screen on a frozen universe with a dated cutoff; a known-positive set of past longs survives | Invert the specification for shorts (guidance cuts, downward revisions, rising borrow cost, holder concentration); a known-negative set fails | Specification, universe, cutoff, survivors, misses |
| Thesis monitoring | A beat and a raised guide are logged as pillar confirmations without the conclusion being rewritten | A guidance cut surfaces by the desk's deadline with the exact management sentence and the segment behind it | Pillar, new fact, source passage, disposition, timestamp |
| Short-side data | Holder concentration and position changes per long, dated | The FINRA print reconciled to the vendor estimate per short, both dated, the estimate labeled | Two series side by side with dates |
| Factor and crowding | Gross and net exposure by factor from the risk system beside each thesis | Crowded shorts flagged, with the crowding source named, before any sizing change | Factor report, crowding flag, source, date |
| Catalyst calendar | A pre-print brief three days before a long reports | The same brief for a short reporting the same week, plus a broker-by-broker note on print day | Brief timestamps, sources used, blanks |
| Permissions and audit | A restricted-list name blocked in search, agent and export for a user without the right | The short thesis never surfaces outside the team, and the denial is logged | Denial record, access log entry |
What a monitor had to catch in Q2 2026 reporting
The three disclosures below are illustrations drawn from public filings of what the thesis-monitoring row must surface. They are not positions, recommendations or views on these companies, and no AllMind monitor produced them: we located the filings through our document search and opened each on EDGAR on September 3, 2026.
| Company and filing | What changed | What the monitor must surface | Source |
|---|---|---|---|
| Honeywell Aerospace (HONA), short-book test; Q2 2026 release on Form 8-K filed August 5, 2026 | FY2026 organic sales growth guidance cut to 4% to 5% from 7% to 9%; pro forma standalone adjusted EBIT guidance to $4.35B to $4.45B from $4.65B to $4.75B; Engines and Power Systems adjusted EBIT down 32% to $174M while consolidated sales grew 5% | The CEO sentence "we believe it is prudent to align our guidance to our supply chain's demonstrated capabilities at the end of the second quarter" and the segment line; shares fell about 20% intraday on August 6, per CNBC | Exhibit 99.1 on EDGAR |
| SoundThinking (SSTI), short-book test; Q2 2026 release on Form 8-K filed August 13, 2026 | FY2026 revenue guidance to $99.0M to $100.0M from $109.0M to $111.0M; adjusted EBITDA margin guidance to 8% to 9% from 16% to 18%; Q2 revenue $23.9M, down 8%, about $2.2M of it from non-renewed or delayed contracts | The named pillar (management says its targets do not depend on the Chicago procurement process it continues to monitor), the churn line, and cash of $6.4M against about $36.0M available on the credit facility | Exhibit 99.1 on EDGAR |
| AMETEK (AME), long-book test; Q2 2026 release on Form 8-K filed August 4, 2026 | Record sales of $2.04B, up 15%; adjusted EPS of $2.09, up 17%; orders up 28%; FY2026 adjusted EPS guidance raised to $8.20 to $8.30 from $7.94 to $8.14 | Pillar confirmations by segment: EIG sales of $1.32B, up 14%, at a 30.1% core margin; EMG sales of $723.2M, up 17%, with 290 basis points of core margin expansion | Exhibit 99.1 on EDGAR |
Two of the three turned on language as much as on numbers: the HONA cut is explained in one sentence about supply-chain capability that a language-change alert should lift verbatim, and the SSTI release names the pillar management would like the market to stop watching.
Compliance prerequisites specific to a long/short desk
The general control set for an AI tool at a hedge fund is in our compliance guide. Three items are specific to a desk that shorts.
Form SHO is not live. The SEC's exemptive order of December 3, 2025 (Release 34-104303) moved Rule 13f-2 compliance to January 2, 2028, with the first Form SHO, covering January 2028, due within 14 calendar days after month-end, and pushed Rule 10c-1a securities-lending reporting to September 28, 2028. A page saying the first filing was due in February 2026 is stale.
MNPI controls reach the short book first. Section 204A requires written policies on material nonpublic information at every adviser, registered or not. The SEC's April 26, 2022 risk alert cited advisers with no system for deciding when alternative-data diligence had to be re-performed, no policies for "value-add investors" such as public-company officers, and no log of expert-network calls or review of the notes. A short thesis built on an expert call or an alternative-data panel is where an examiner looks first, so the system has to show who read which transcript and which panel, and when.
AI representations are an exam item. The Division of Examinations' FY2026 priorities, released November 17, 2025, say staff "will review for accuracy registrant representations regarding their AI capabilities" and will assess whether firms have "adequate policies and procedures to monitor and/or supervise their use of AI technologies" (priorities document). If the questionnaire says the short book is monitored by AI, the exam checks that the monitor exists, runs and is supervised.
Where AllMind fits a two-book desk
For a fundamental long/short desk that runs both books from written theses, already pays for a terminal and a prime's risk report, and routes a feed or two into its own warehouse, AllMind is the strongest first pilot. The reason is the join. Our ontology holds filings, transcripts, LSEG I/B/E/S estimates and revisions, ownership and holdings, broker research, Expert Insights and news as one graph. The fund's thesis ledger and models join it: Snowflake, Databricks or S3 are queried in place, and research-management systems such as Verity RMS, FactSet RMS and BipSync connect by integration scope. When a broker revises an estimate, every comp table and thesis built on the old number is flagged, the fund's included.
Two mechanisms carry the two-book workflow. Monitoring agents in Agent Studio take the thesis once, check it against each new filing, transcript and headline, flag evidence that cuts either way, and alert when the way management talks about guidance, demand or risk changes from one call to the next. Our hedge-fund workflow page puts the short side plainly: write down why you are short, then watch those pillars, nightly or at the interval the desk sets. Grids runs one question across both books at once, with a cited answer in every cell, change alerts and an Excel export.
Permissions are enforced in the graph: restricted lists and deal walls map to roles, an agent inherits the role of whoever ran it and can never widen it, and every access is logged. That is the evidence the audit row and the MNPI paragraph both ask for. We are SOC 2 Type II certified, with ISO 27001 and GDPR certification targeted for Q1 2027. The finished artifact is a two-book thesis ledger: each long and each short with its pillars, a nightly exception report naming the pillar that moved, the passage behind it and the analyst's disposition, and a note per name on print day. When a name is ready for the committee, the memo comes out of Reports in the house template with every figure linked to its passage.
The boundary is the trading and risk stack. We are not a trading or execution terminal, so the OMS, the borrow locate and the factor model stay where they are. We do not supply short interest, borrow rates or factor exposures as native feeds; S3 Partners, MSCI Barra and Arcana Analytics connect by integration scope. A fund whose first unsolved job is factor risk modeling should license the model first; one whose first job is execution should buy the terminal seat; one that mostly needs a single expert-transcript library should buy that library first. Aftermarket broker research is included on a delay that varies by broker; live embargoed notes run on the firm's own entitlement, so a desk that lives on morning notes should check entitlements before the pilot.
What we could not verify
- No product was tested, and no AllMind monitor, Grid or agent ran on the Q2 2026 disclosures.
- Closing-day share reactions for HONA, SSTI and AME; only CNBC's midday HONA figure is cited.
- MSCI's Barra pages, FactSet's AI page, S&P Capital IQ Pro and Visible Alpha could not be opened; no claim rests on them.
- S3 Partners' daily global short interest has no published methodology on the page we opened.
- Every quote-only price, ours included.
Frequently asked questions
How do long/short funds use AI to screen for new positions?
The screen only produces candidates; the research system then reads the filings and transcripts behind each survivor and drafts the pillars a monitor will watch, so a screen hit becomes a written thesis before it becomes a position. A name that passes both the long screen and the inverted short screen means the specification is loose, so the desk tightens the filter and re-runs it. On the short side, the FINRA print and the vendor estimate stay as two dated columns; when they disagree, the desk records whether the gap is timing, because the estimate is dated after the settlement date, or the vendor's model, and never averages the two.
Does a long/short fund need a factor risk model before an AI research system?
If the mandate is market neutral or beta constrained, yes: the factor model is the risk control, and the research system's job is to read its output next to each thesis. A concentrated fundamental long/short desk can run its first pilot on the prime broker's risk report and license a model when gross exposure or the short book grows.
Sources and methodology
Every page and filing named in this public-source brief was opened on September 3, 2026; no product was tested, and no AllMind agent ran on the Q2 disclosures.
Regulators: FINRA's equity short interest page (undated), the SEC exemptive order of December 3, 2025, the FY2026 examination priorities with the SEC's November 17, 2025 release announcing them, and the MNPI risk alert of April 26, 2022. Reporting: CNBC on August 21 and August 6, 2026. Filings: the Honeywell Aerospace, SoundThinking and AMETEK Form 8-K exhibits filed August 5, August 13 and August 4, 2026. Vendor pages, all undated: Koyfin, Ortex, S3 Partners, Hebbia, Axioma, Aiera, AlphaSense Expert Insights and Third Bridge.
If the two-book test is the decision in front of the desk, bring one written long, one written short and the last two prints for each. We will run the acceptance table above on them and return the exception report with its blanks left visible.