WorkflowInvestment Research & Analytics

Financial Data Extraction

Financial data extraction into the analyst's own model, the day the company files, with every number linked to its page

See one case, screen by screen ↓

Built forResearch associatesEquity research analystsSenior analystsHeads of research dataCompliance officers

The Problem Today

Associates retype each filing into the analyst's model by hand, so a wrong unit or a recast year slips in unseen.

2 h
to update one company's model by hand, each quarter our estimate

Typed and checked by hand

  • Filing
  • Release
  • Deck
  • Model
  • Source? Reopen the PDFA manager's question means finding the page again
  • Vendor data stops shortSmall caps, private and local filers left out

The Solution

  1. Each filing caught at onceSEC, exchange and investor sites watched; each linked to its own model
  2. Read into your line namesEach table read in its own language, scans too, mapped to your lines
  3. Footed, never pluggedSegments added up; a gap traced to its reconciling line on the page
  4. Checked, then publishedA different senior analyst checks; the analyst publishes the version

Built-in controls

  • A reason code on each override
  • Restricted list checked at publish
  • Audit trail kept seven years, with no edits

What You Get

Models are current before the morning meeting, small caps and private issuers included, and associates spend their hours on the few numbers that are wrong.

92 h
of analyst time saved a season on a 46-name list our estimate
  • Private and local names too144A, scanned and local-language filings read
  • Recast history keptOld and new values side by side for a senior analyst
  • Learns your line namesMapping improves with each decision the associate makes
In short

What is Financial Data Extraction?

Financial data extraction, also called financial statement data extraction, is pulling reported numbers (revenue, segment results, KPIs, cash flow, debt) out of filings such as the 10-Q, 10-K, earnings releases and decks and into a model; KPI extraction is the same work for operating metrics such as orders, backlog and book-to-bill. Financial Data Extraction does it into the analyst's own Excel model: seven AI agents read each page, map each figure to the analyst's own line names, foot each subtotal and link each cell to its page. A research associate accepts each new column, a different senior analyst checks it and the analyst publishes.

AI in asset management: a buyer's guide to agentic AI for investment managers →

Updated

The problem

Financial data extraction by hand: about two hours a ticker, each quarter

When a covered company files, a research associate opens the 10-Q, the press release and the earnings deck and types thirty or more numbers into the analyst's own Excel model, under the analyst's own line names. Then come the checks: that segments add up to consolidated revenue, that thousands did not become millions, that year-to-date cash flow was not entered as the quarter, that last year's figures were not quietly recast. For a large research team, most of a quarter's filings land inside a few weeks, after 4 PM and before the morning meeting.

Data vendors and their financial data extraction tools cover the largest listed names, but in their own standardized line items, and they stop at small caps, private and 144A issuers and local-language filers, so for those issuers their numbers are typed by hand. The errors are silent: a corporate and eliminations line missing from the model, a relabeled KPI that breaks a series, recast history overwritten. And when a portfolio manager asks where a number came from, someone reopens the PDF.

typical40daysafter quarter end for a large accelerated filer's 10-Q, 45 for non-accelerated filers; then each covered model needs its new column
estimated3,000–6,000companiesin a source-linked data vendor's published coverage of listed names; private issuers and many small-cap and local-language filers are not in it
estimated15–45mina vendor says it saves per model update, and only for the names it covers
Where a ticker's update hours go, each quarterestimated
By hand2 hours
With the solution0.1 hours
  • Pulling the 10-Q, the release and the deck0.3 → 0.0 h
  • Typing values under the analyst's own line names0.8 → 0.0 h
  • Footing segments, units and year-to-date periods0.5 → 0.1 h
  • Recast history and relabeled KPIs0.3 → 0.0 h
  • Tracing a number back to its page for a PM0.3 → 0.0 h

Estimated split of about two hours by hand per ticker per quarter, and about eight minutes of review with the solution.

How it works

Financial data extraction stage by stage: read, map, foot, check, publish

What comes in
Filings inany issuer · 10-Q, 10-K, releases, decks, local filings, private packs
Agents at work
Filing Watcherlinks · clock
Then
Table Readerpages · scans
Then
Mapperyour line names
Units & Periodsscale · FX · YTD
Then
Footingfinds, never plugs
Restatementhistory kept
Then
Model Writerevery cell linked
A person decides
Associate, then senior analystaccepts · checks
What comes out
New column in the analyst's own model
Every cell's lineage to its page
Checked version published
One case, step by step

Financial data extraction overnight: a 10-Q modeled by the morning meeting

A large accelerated filer with three segments (Flow Control, Motion Systems and Aerospace Components) files its third-quarter 10-Q at 4:12 PM on a Tuesday. The analyst who covers it has a morning meeting at 9:30. Here is how the people work with the agents, screen by screen, in the working solution.

A worked example on sample data. Names, IDs and figures on the screens and in this section come from the demo.

  1. 014:12 PM, Tuesday

    Seven agents fill the new column in 40 seconds

    The agents

    The Filing Watcher picks up the 10-Q from EDGAR, matches CIK and ISIN to the issuer and links the 4:05 PM release (8-K Exhibit 99.1) and the deck. The Table Reader reads 84 pages and 63 tables. Units & Periods turns thousands into millions and the deck's “$2.31 billion” into 2,310.0, and derives the quarter's cash flow from nine months: 301.7 − 182.8 = 118.9. The Mapper fills 31 line names, Footing runs 46 checks, Restatement compares history with the last version, and the Model Writer cross-checks the estimates and writes the column.

    “Seven agents, 40 seconds, each step on the record. They propose; people decide.”

  2. 028:06 AM, Wednesday

    Good morning: what needs the associate, and what the agents found

    Research associate

    Nine of the analyst's 46 names filed since yesterday; eight models are refreshed and one is being read now. Eight decisions wait, the nearest due at 9:30 AM. Overnight findings each carry their page: the industrial's $1.7M segment gap on p.22, a German filer's two units on p.4 and p.9, a Japanese filer's rounded English summary. Asked what could miss its deadline, the Filing Watcher names two models.

    Filing Watcher: “Two models: the industrial (3 decisions) and the German drives maker (1 mapping). If they are not accepted by 9:00 AM I escalate to the senior analyst, per the SLA rule.”

  3. 03In the analyst's model

    The new quarter, in the analyst's own grid and line names

    Research associate

    The new column sits in the analyst's own workbook, beside four quarters of history, the estimate and the difference. Revenue is $814.1M, 1.1% above the $805.0M estimate; Adj. EBITDA ex-SBC is 177.0 against 171.5. A colored corner on each cell says whether it footed, needs the associate, was derived or converted, or is restated history.

    “Two decisions need you before 9:30 AM, then you accept the column. Start with the footing: it changes total revenue.”

  4. 04One click

    Each cell carries its lineage back to the page

    Research associate

    The associate clicks Adj. EBITDA ex-SBC. The lineage shows the filing (8-K Exhibit 99.1, filed 4:05 PM), page 9 with the reconciliation of non-GAAP measures, the table row printed 177.0, no conversion, and the mapping rule learned from 214 of the analyst's decisions. The check ties 136.9 + 27.1 + 9.8 + 3.2 = 177.0. The vendor dataset shows 173.8 because it excludes restructuring; the analyst's definition is kept.

    Lineage: filing → page → table row → units and period → mapping → checks → cell.

  5. 05The amber cell

    Segments don't foot, so the agent finds the line instead of plugging

    Footing

    The three segments add to $812.4M (318.6 + 276.9 + 216.9); consolidated net sales are $814.1M. The Footing agent searched all 48 pages and found “Corporate, other and eliminations” of $1,711 thousand in Note 14, “Segment information”, on page 22, a line missing from the analyst's model. The four prior quarters were short by 1.2, 1.9, 1.4 and 1.6 for the same reason. It proposes a new row, back-filled for four quarters.

    Footing: “I found the $1.7M … I will not plug it.” One click opens page 22 with the line highlighted.

  6. 068:09 to 8:14 AM

    A relabel would break the series, so the Mapper asks first

    Research associate

    The associate adds “Corp & elims rev” and total revenue foots in all five columns. Then the Mapper's question, at 0.91: the deck's “Aftermarket revenue” is now “Aftermarket & services revenue” of $289.0M, including $11.2M of field services previously in equipment revenue (deck p.6, footnote 2). Taking $289.0M as reported would break the analyst's series. The associate keeps it comparable at 277.8 with a new “Field services rev” row, then accepts the column: 33 values, 46 of 46 checks footing, the associate recorded as the maker.

    The two choices on screen: “Keep comparable · 277.8, plus a new row ‘Field services rev’ 11.2” or “Take as reported · 289.0, mark a series break”.

  7. 071:00 AM, in German

    A German quarterly beside its English values, each pinned to its page

    Units & Periods and Mapper

    A German drives maker's Quartalsmitteilung reports in Tsd. € on page 4 and Mio. € on page 9, with comma decimals; both are converted and revenue ties at €412.4M. “EBIT bereinigt” is a new label: its definition on page 11 excludes restructuring and purchase price allocation, matching the analyst's Adj. EBIT at €42.6M. At 0.89, below the 0.92 threshold, the Mapper asks instead of mapping. For a Japanese precision maker's kessan tanshin, Q2 is derived from the half: ¥48,612M − ¥23,904M = ¥24,708M.

    “The English summary rounds revenue to ¥48.6 billion; the agents used the Japanese original, ¥48,612M.”

  8. 08All morning

    What the agents could not settle, each with a proposal and its page

    Research associate

    The analyst's names show eight exceptions, five still open, and 143 are open across all teams: unmapped lines, failed footings, restatements, units and currency, low scan confidence, period derivations. A small-cap pump maker outside vendor datasets shows a new $1,384K tariff surcharge line the Mapper proposes to keep in revenue with a memo row. A private issuer's scanned 144A noteholder report read at 0.83, 0.61 and 0.82, and is held for a person with the crops.

    Table Reader: “I don't guess a digit: type what you see and the Footing agent re-checks it against the covenant table on p.9.”

  9. 098:31 AM

    A different person checks the recast and the footing

    Senior analyst

    A bearings business moved from Motion Systems to Flow Control, and an earlier 8-K recast four quarters: last year's third-quarter Flow Control revenue goes from 287.4 to 300.2 and Motion Systems from 262.0 to 249.2. The senior analyst confirms the 16 cells match Exhibit 99.2, that the new row is the reconciling line on p.22 and no plug, that the Aftermarket decision keeps the series comparable, and that there were no overrides. The associate could not have checked their own column.

    “You accepted this column, so you cannot check it. A different person checks: the senior analyst.”

  10. 108:42 AM · published

    The new version published, 48 minutes before the morning meeting

    Equity research analyst

    Before the analyst can publish, the restricted list is checked (the issuer is clear in the compliance control room at 8:40 AM), with entitlements and maker-checker: the associate made it, the senior analyst checked it, the analyst publishes. Version 14 goes to the research management system and the workbook, and the portfolio managers who follow the name are notified.

    “Publishing writes the new version to the research management system and the workbook. Subscribed portfolio managers are notified.”

Who it’s for

Five seats around one 10-Q: maker, checker, publisher, rules and records

The same 10-Q, seen by the five people who carry it, in the old earnings season and the new one.

Research associate
Accepts each column as the maker
Before
Types thirty-odd numbers per filing into the analyst's models and checks them by hand, six to ten filings a night in season, when the real job is finding the three that are wrong.
Now
Opens each model with the column already filled, footed and linked to its pages, decides the few things the agents could not settle and accepts.
Equity research analyst
Owns the model and publishes
Before
Owns the models and the estimates; when a portfolio manager asks where a number came from, someone reopens the PDF.
Now
Gets their own model, in their own line names, refreshed the day each name files, and publishes a checked, numbered version with each cell's page one click away.
Senior analyst
Checks recasts and footing
Before
Restated history could be overwritten without a trace, and a model could reach portfolio managers before anyone checked it.
Now
Checks each restated cell and footing decision before a version is shared, with old and new values side by side and the 8-K exhibit cited.
Head of research data
Owns the rules
Before
Owns the line-name dictionary, templates, thresholds, integrations and agent versions, and wants each rule change versioned and seen by a second person.
Now
Sets thresholds, unit and period rules and roles once; each change is proposed by the head of research data, approved by a second administrator and versioned.
Research compliance officer
Reads the record
Before
Answers exam and audit requests about research records, which means showing who changed a cell, when, why and from which page.
Now
Filters each event by person, agent, type and issuer, opens the page behind it, and exports one evidence pack per model version.
The agents

Reader, Mapper, Footing and the other agents behind financial data extraction

Every value is read in its own language, mapped to your line names, footed and linked to its page; an associate accepts, a different senior analyst checks, the analyst publishes.

Filing Watcher

Watches EDGAR, exchange feeds, IR sites, bondholder sites and data rooms for each covered issuer; links each document to the issuer, the analyst's model and the earlier release, and starts the clock to the model's due time.

  • Reads only public filings and documents the covering team is entitled to
  • Never opens a private pack outside the covering team's wall
Table Reader

Reads each page and table in the original language, scans included, and records each value with its page, table, row, column and printed units.

  • Never infers a number that is not printed
  • Scan reads below 0.85 go to a person with the crop
Mapper

Maps each issuer's label to the analyst's own line name, learning per analyst from the workbook and from each decision people make.

  • Below 0.92 it asks; a changed definition always asks
  • Never adds or removes a model row without a person
Units & Periods

Handles thousands against millions, currency at the house rate, sign conventions, year-to-date to quarter and each issuer's fiscal calendar.

  • Never converts with a rate other than the house rate
  • Shows its arithmetic on each derived value
Footing

Adds each subtotal (segments to consolidated, income statement, balance sheet, cash flow, each non-GAAP reconciliation to GAAP) and, when something does not foot, searches the filing for the reconciling line.

  • Never plugs a difference
  • A cell it could not foot goes to a senior analyst at check
Restatement

Compares each prior period a filing re-reports against the published model, detects recasts and restatements, and proposes the update with the old value kept.

  • Never overwrites published history
  • Recasts always go to a senior analyst
Model Writer

Cross-checks against the analyst's estimates and a vendor dataset where one exists, writes the accepted column into the workbook with a link and a note on each cell, and prepares the version diff and the evidence.

  • Never accepts, checks or publishes
  • Writes only what a person accepted
Associate and senior analyst

Accept and check before the analyst publishes.

Ask in plain words

Questions on any filing, cell or model version, answered with the page

Ask in plain words, or tell the agents what to do. Each answer cites its page, and anything that changes the model still needs a person. These answers use the sample data.

What did the agents find in the industrial's 10-Q?

31 values mapped in 40 s; 45 of 46 checks foot. Segments add to $812.4M against $814.1M: the $1.7M is “Corporate, other and eliminations” on p.22. Aftermarket's definition widened (deck p.6), and four quarters were recast (8-K).

Treat restructuring as recurring for this name

Understood. “Adj. EBITDA ex-SBC” would stop adding back restructuring: this quarter becomes 173.8 (177.0 − 3.2) and the four prior quarters change too. This changes the analyst's definition, so it needs the analyst's approval.

Why not use the Japanese filer's English summary?

The English summary rounds to ¥0.1 billion (“¥48.6 billion”), which would move Q2 revenue by up to ¥50M. The Japanese original gives ¥48,612M on p.1.

Is the industrial on the restricted list?

No. The compliance control room shows it clear as of 8:40 AM. It is checked again at the moment the analyst publishes; if it were restricted, publish would stop and Research Compliance would be told.

The screens

The model grid, cell lineage and the footing check from a working run

From the working solution, on its sample data. Pick one to see it large.

Today, in earnings seasonThe associate's morning: this week's reporting, decisions waiting, models refreshed, filing-to-model time, findings with their pages, and what could miss its deadline.
Agents at work on a 10-QKPI extraction on one 10-Q, its release and deck: pickup from EDGAR, pages and tables read, units and periods, line names, footing checks, the recast and the hand-off, each step timed and sourced.
The analyst's own Excel modelThe new column in the analyst's own Excel grid and line names, beside the estimate and the difference, with the agents' notes and the three items waiting for the associate.
Cell lineage to the pageOne click on a cell: the filing, page 9's non-GAAP reconciliation, the table row, the conversion, the mapping and the checks behind Adj. EBITDA ex-SBC.
Segments that don't footFinancial statement data extraction with a footing check: the segments short of consolidated sales, and the corporate and eliminations line on page 22 cited instead of a plug.
A relabeled KPI, decided by a personThe Mapper asks before mapping a relabeled line: keep Aftermarket revenue comparable with a new field services row, or take it as reported and mark a series break.
German filing beside English valuesExtracting financial data from a German Quartalsmitteilung: Tsd. € and Mio. € converted, a new adjusted EBIT label held for a person, each figure pinned to its page in the analyst's line names.
Exceptions the agents could not settleUnmapped lines, failed footings, restatements, units, scans and period derivations, each with the agent's proposal, confidence, page and due time, decided with a reason code.
The senior analyst's checkA different person checks the restated history against the 8-K exhibit, the footing row and the relabel decision before any model version is shared.
Publish a checked model versionRestricted list, entitlements and maker-checker confirmed before the analyst publishes the new version to the research management system and the workbook.
Audit trail for one model versionEach agent step and decision on one model version, through to the publish, with who, what, when, why and the page; nothing can be edited.
Reports for the research leaderFiling to refreshed model, hours saved per ticker (estimated), values accepted without an edit and issuers outside vendor datasets, by team and analyst.
Thresholds and footing checksMapping and scan thresholds, column-accept suggestion, footing tolerance, estimate highlight and the reason codes for overrides, each change approved by a second administrator.
Governance

Financial data extraction controls: page citations, maker-checker, audit trail

Each value cites its pageEach cell carries its filing, page, table row, the as-reported value, its unit conversion and mapping, and each agent note links to the page it rests on. The page opens with the line highlighted.
Maker, checker, publisherA research associate accepts the column, a different senior analyst checks it and the analyst publishes a numbered version. The checker and the publisher can never be the person who accepted.
No plugs, no overwritten historyThe agents do not plug a difference, overwrite published history, change a model's rows or overwrite the analyst's formula cells. Recasts and unfooted cells always go to a senior analyst.
Restricted list and entitlements at publishPublishing checks the compliance control room's restricted list and the model's entitlements. Private and 144A packs are walled to the covering teams and Research Compliance.
A reason code on each overrideEach edit, reject or return needs one of ten reason codes, such as “Wrong units or scale” or “Definition differs from the analyst's”. The agents learn from the code; a person approves what they learn.
An audit trail kept seven yearsEach agent step and decision is logged with who, what, when, why and the evidence, and nothing can be edited. Seven years is the firm's policy, set by Research Compliance, longer than the five years SEC Rule 204-2 generally requires. An evidence pack per model version (lineage, page images, decisions, the diff and agent versions) is sealed with a SHA-256 hash.
Configuration

Mapping thresholds, footing tolerance and refresh deadlines as settings

Thresholds, unit rules and service levels are settings. Each change is proposed by one administrator, approved by a second and versioned.

SettingDefaultChoose from
Mapping asks a person below0.92Slider from 0.85 to 0.99, showing how many cells a day would come to people before saving
Scan reads go to a person below0.85House default
Suggested for column accept at≥ 0.97 and footedA person still accepts every column
Footing tolerance$0.1M or 0.05%, whichever is smallerChanged by the Head of Research Data, approved by a second administrator
Highlight the difference from the analyst's estimate above3%Set per analyst
Year-to-date to quarterOnDerive the quarter from YTD less the prior YTD and show the arithmetic · on or off
CurrencyHouse rateMarket data feed, 4:00 PM New York close on the period end, labeled with the rate and date
Model refreshed by9:30 AM local for filings after 4:00 PMEscalates to the senior analyst 30 minutes before; a daily digest at 7:30 AM
Connections

EDGAR, local exchange feeds, data rooms and the analyst's Excel model

SEC filingsForm 10-Q, 10-K and 8-K releases from the EDGAR feed, with exhibits
Local filings and IR sitesexchange and disclosure feeds in each issuer's home market; decks and releases from IR sites
Private and 144A issuersbondholder reports and data-room packs, scans included, from document management
The analyst's Excel modelan Excel add-in writes accepted columns with a link and a note on each cell
Research management systempublishes checked model versions and notifies the portfolio managers who follow them
Compliance, market data and vendor datarestricted list at publish, house FX rates, and a vendor fundamentals feed for cross-checks only
What it changes

Models current the day each name files, small caps included

Each figure carries its label: a target the solution is built to, an estimate, a typical published result, or a proven one.

target
Same day
every covered name's model refreshed the day it files, including small caps, private and local-language issuers
Every reported name refreshed
filing to model: 6 min median
target
Every
number in a published model tied to its page, footed and checked by a second person
46 of 46 checks foot
a $1.7M gap found on p.22, not plugged
estimated
~2hours
saved per ticker per quarter, across every name on the coverage list
By handabout 2 h
With agentsabout 8 min

“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in the live solution · “estimated” = our estimate (where a ticker's update hours go, by hand and with the solution; the 92 hours a season is about two hours saved per ticker across a 46-name list; a source-linked data vendor's published coverage of about 3,000–6,000 listed companies and its claimed 15–45 minutes saved per model, vendor figures, directional) · “typical” = published figures (SEC Form 10-Q general instructions: 40 days after quarter end for large accelerated filers, 45 for non-accelerated filers). Names on the screens are sample data.

Questions

Before a trial of financial data extraction: what research teams check

What is financial data extraction?

Financial data extraction, also called financial statement data extraction, is pulling reported numbers (revenue, segment results, KPIs, cash flow, debt) out of filings such as the 10-Q, 10-K, earnings releases and decks and into a model; KPI extraction is the same work for operating metrics such as orders, backlog and book-to-bill. Financial Data Extraction does it into the analyst's own Excel model: seven AI agents read each page, map each figure to the analyst's own line names, foot each subtotal and link each cell to its page. A research associate accepts each new column, a different senior analyst checks it and the analyst publishes.

Can it extract financial data from PDF filings, scans and local-language reports?

Yes. The Table Reader reads each page and table in the original language, such as a German quarterly statement or a Japanese kessan tanshin, and records each value with its page, row and printed units. Units & Periods converts Tsd. €, Mio. € or ¥ millions and derives quarters from year-to-date figures, showing the arithmetic. Scanned values read below 0.85 confidence go to a person with the crop; the agents never guess a digit.

Does it update our own Excel models or a vendor's template?

Your own. This is automated financial modeling for the model each analyst already keeps: the Mapper learns each analyst's line names from the workbook and from their decisions, and maps each issuer's as-reported financials to them; when a label changes meaning, it asks instead of breaking the series. A vendor dataset is used only to cross-check, and the analyst's formula cells are left as they are.

What happens when segments don't foot or a company recasts prior quarters?

The Footing agent adds each subtotal and, when segments do not reach consolidated revenue, searches the filing for the reconciling line and proposes a row instead of plugging a difference. The Restatement agent compares re-reported periods with the published model and proposes the recast with the old values kept in each cell's history. Both always go to a senior analyst at check.

Can it cover small caps, private and 144A issuers outside vendor datasets?

Yes. The agents watch exchange feeds, IR sites, bondholder sites and data rooms for each covered issuer, and an analyst can add an issuer and its sources. In the working solution's sample data, the issuers outside vendor datasets run from a small-cap 10-Q filer to a private 144A issuer's scanned noteholder report.

Is financial data extraction the same as AI financial modeling?

They meet in the model. Financial Data Extraction does not build a new model from a template: its financial modeling automation keeps the analyst's existing one current, with their line names, their adjusted definitions and their estimates. Each new quarter is mapped, footed, checked against restated history and linked to its page, and the analyst still owns the estimates and the view.

Can a number reach a published model without a person accepting it?

No. The agents propose and do not accept, check or publish. A research associate accepts each column, a different senior analyst checks recasts and footing decisions, and the analyst publishes only after the restricted list and entitlements are checked. Each override needs a reason code, and each step is on an audit trail kept seven years.

What do we connect first, and how is it proven on our names?

Your analysts' workbooks, line names, coverage list and filing sources. The Agentic Solution Engine builds and deploys it from those requirements and documents, and it goes live once each quality gate has passed. First-time setup connects a workbook, reviews the mapping, picks the coverage, sets the checks and runs a trial on last quarter against what your team typed before anything is written.

Test financial data extraction
against last quarter's typed models.

We'll run Financial Data Extraction on last quarter's filings for your own covered names, and compare it with the models your team typed.