How to Build Trading Comps with AI

To build trading comps with AI, let the AI propose peers and draft the sheet, and make sure every input is pulled from a filing and every multiple is computed in code. That’s how Kepler does it: it gathers each peer’s LTM figures from the filings, runs NTM on your own estimates, builds the enterprise value bridge from the latest balance sheet, applies each filer’s own adjusted EBITDA or your firm’s definition, calendarizes fiscal years and returns an Excel sheet where every cell is a formula or a cited input. We build Kepler. This guide walks the six steps an analyst follows, shows where general-purpose AI tends to break, and ends with a checklist to run before a comps page goes in front of a PM or committee.

The six steps of a trading comps sheet

Trading comps (comparable company analysis) price a business off what the market pays for similar public companies. The mechanics are simple and the judgment is everywhere:

1. Pick the peer set.
2. Pull LTM and NTM financials on a consistent basis.
3. Build each company’s enterprise value bridge.
4. Decide whose adjusted EBITDA you’re dividing by.
5. Calendarize fiscal years so every peer covers the same months.
6. Deal with outliers and present the range.

Each step below says what to do by hand, what AI can safely take over, and what to check.

Step 1: Pick the peer set

Start with a long list, then cut. Screen on business model, end market, geography, size, growth and margin profile. Industry codes get you a first draft; the 10-K’s business section (Item 1) and the competitors the company names in its own filings get you closer. A company’s proxy (DEF 14A) also lists the peer group its board uses for pay, which is a useful cross-check on who management thinks it competes with.

AI is good at this step when it can read every candidate’s filings. Ask for companies that “sell endpoint and cloud security software primarily by subscription, with more than half of revenue from the US,” and a tool that has read the business sections of every filer returns a long list with a reason for each name. A tool that searched the web returns the ten most-written-about names. In Kepler, a saved peer group lives in your firm’s ontology with the reason for each name beside it, so the next analyst inherits the logic along with the list. Write down why each peer is in, and why each obvious candidate is out.

Step 2: Pull LTM and NTM figures

LTM (last twelve months) = latest fiscal year + current year-to-date − prior year-to-date. If a company reported fiscal 2025 revenue of $4,000M, nine-month 2026 revenue of $3,300M and nine-month 2025 revenue of $2,900M, LTM revenue is $4,400M. Pull the year-to-date columns from the latest 10-Q and the fiscal year from the 10-K, and check that neither has been recast since (our guide to restatements and reclassifications shows how).

NTM (next twelve months) blends two fiscal years of forecasts by the months remaining. NTM comes from consensus or your own estimates, so record which one and its date. Kepler reads your own model as a source, so NTM can run on the house numbers rather than a figure someone pasted in last month.

The rule that matters: every peer on the same basis. One company on LTM through September, another on its last fiscal year ended June and a third on calendar 2025 will produce a median that means nothing.

Step 3: Build the enterprise value bridge

Item	Where it lives	Treatment
Diluted shares	10-Q cover page (basic), equity and stock-comp notes (options, RSUs, convertibles)	Treasury stock method for options; if-converted or net-share settlement for convertibles
Share price	Market data, as of one date for all peers	Same close for every company
Debt	Balance sheet and debt footnote	Principal of each instrument, short and long term; finance leases included
Preferred stock	Balance sheet, equity note	Add at liquidation preference
Noncontrolling interests	Balance sheet	Add, so EV matches consolidated EBITDA
Cash and investments	Balance sheet	Subtract cash and short-term investments; decide on long-term securities and be consistent
Equity-method investments	Balance sheet, investments note	Subtract if their income sits below EBITDA
Operating leases	Balance sheet, lease note	Exclude under US GAAP when EBITDA is after rent; include under IFRS 16

A worked bridge for a hypothetical peer: 320M diluted shares at $50 gives $16,000M of equity value; add $3,000M of debt and $200M of noncontrolling interests, subtract $700M of cash, and enterprise value is $18,500M.

The debt line deserves care. A note’s carrying amount on the balance sheet is net of discounts and issuance costs; the principal is what a buyer would have to repay. Our research note The quieter hallucination walks through LinkedIn’s convertible notes before Microsoft’s offer: holders could make LinkedIn buy the notes back at 100% of principal in a takeover, so pricing the deal off the balance-sheet figure instead moves enterprise value by $184 million, with a citation that looks perfect. Pick the basis on purpose, and make the sheet say which one it used.

Step 4: Decide whose adjusted EBITDA

Most companies report an adjusted EBITDA, and each one adjusts differently. Under Regulation G and Item 10(e) of Regulation S-K, any non-GAAP measure in an earnings release has to come with a reconciliation to the nearest GAAP figure, and that reconciliation table (usually in Exhibit 99.1 to the 8-K, and again in the 10-Q or 10-K when the measure appears there) is the place to see what was added back.

Common add-backs: stock-based compensation, restructuring, acquisition and integration costs, litigation, impairments, and “other.” Two companies can both report adjusted EBITDA margins near 25% while one adds back stock comp worth 15% of revenue and the other adds back almost nothing.

You have three options, and the best sheets show at least two:

• Each filer’s own adjusted EBITDA. What management and the sell side quote. Easy to defend, hard to compare.
• A standard definition applied to every peer. For example, GAAP operating income plus D&A, with only restructuring added back. Comparable across the set.
• Your firm’s definition. The one your IC or PM uses everywhere else. This is the one that should drive the decision.

Continuing the example: if our hypothetical peer’s $1,000M of adjusted EBITDA includes a $180M stock comp add-back, then on a definition that treats stock comp as a cost, EBITDA is $820M and EV/EBITDA moves from 18.5x to 22.6x, four turns on the same company and the same share price. Kepler shows both columns side by side, each add-back cited to its reconciliation line. Our guide to getting AI to use your firm’s definitions covers how to make that definition stick.

Step 5: Calendarize fiscal years

Peers rarely share a year-end. In US cybersecurity, CrowdStrike, Okta and SentinelOne end their fiscal years on January 31, Palo Alto Networks and Zscaler on July 31, and Fortinet on December 31. Comparing each company’s “fiscal 2026” compares different stretches of time.

Weight the two fiscal years that overlap the calendar year by months:

CY2025 = (7/12 × FY2025) + (5/12 × FY2026) for a July year-end.

With FY2025 EBITDA of $900M and FY2026 EBITDA of $1,140M, calendar 2025 is $525M + $475M = $1,000M. January year-ends are usually treated as the prior calendar year with no weighting. When quarterly data are available, sum the four quarters closest to the calendar year instead; it’s more precise and easier to audit. Kepler does the quarter sum in code and shows which four quarters it used for each peer.

Hypothetical peer	Year-end	EV ($M)	CY2025 EBITDA ($M)	EV/EBITDA
Company A	December	24,000	1,600	15.0x
Company B	July	18,500	1,000	18.5x
Company C	January	9,800	350	28.0x
Median				18.5x

Step 6: Handle outliers and present the range

Company C above is the kind of name that distorts a set: a margin depressed by a one-off charge, a recent acquisition not yet in LTM, or a growth rate twice the group’s. Options, in order of preference: explain it in a footnote, show the multiple on an NTM basis where the distortion washes out, or exclude it with a stated reason. Show the median and the interquartile range, never only the mean, and say “NM” for negative or meaningless multiples rather than printing them.

Where general-purpose AI breaks on comps

A frontier model on its own can lay out a comps template and explain every step above. The failures show up in the numbers:

• It types the multiple. The model writes 18.5x as text. If an input drifted, nothing catches it, and the citation points to a source without proving the figure came from it.
• It mixes periods. One peer on LTM, one on fiscal year, one on a calendar year it calendarized in its head.
• It misses the reconciliation. Adjusted EBITDA comes from a press summary rather than the Exhibit 99.1 table, so you can’t see what was added back.
• It uses the wrong share count. Basic shares from the cover page, dated weeks after the balance sheet, with no options or RSUs.
• It can’t reproduce. Ask twice and the median moves, because the model chose different inputs.

Check any AI-built comps with the ten tests in how to check AI-generated numbers.

How the tools on the market approach comps

• S&P Capital IQ Pro builds comps on its standardized financials. Its October 6, 2026 update added point-in-time trading comparables, adjustable total enterprise value calculations and ChatIQ Spreadsheets, an Excel modeling assistant that works with refreshable Capital IQ Pro formulas.
• FactSet offers Screening for Dealmakers, with an AI-powered Screen On tool that searches public and private companies and automates early screening and comparable analysis.
• Claude for Financial Services added Agent Skills in its October 2025 update, including comparable company analysis with multiples and operating metrics that can be refreshed as data updates.
• ChatGPT for Financial Services, launched September 10, 2026, includes data from Daloopa, PitchBook, LSEG News and Crunchbase, with citations that trace figures to specific tables and passages.
• Rogo lists a comparable companies analysis among the pre-built agents on its agents page, built for banking deliverables.
• Daloopa supplies source-linked fundamentals for 6,000+ public companies through an Excel add-in and an MCP server, so other tools can build comps on its data. See our Daloopa alternatives page for more.

The questions to ask any of them: whose adjusted EBITDA each multiple uses, what a click on a cell opens, whether the same request returns the same sheet, and whether it can apply your firm’s definition to every peer.

How Kepler does this

Kepler builds comps the way a careful associate would, and keeps the receipts for every cell.

Your peer sets and definitions live in your firm’s ontology. Our forward-deployed engineers build a living model of your firm: your coverage list, your saved peer groups, and your definition of adjusted EBITDA, written once and applied to every company. Kepler also ships with an ontology of public companies built from SEC filings, so subsidiaries, tickers and former names resolve to one record.

Every filing is read in full before you ask. Kepler holds 26M+ SEC filings across 20+ form types and 40M+ IR documents and transcripts for 14,000+ companies. Extraction models we train read each earnings release at ingest, so every line of the non-GAAP reconciliation is kept with its add-back, its period and its page. That’s how a peer set can be screened on business descriptions across every filer, and how each company’s adjustments sit beside your standard definition.

Code computes every number. The model never writes a multiple. LTM, calendarization, the treasury stock method, the EV bridge and each multiple run as formulas over cited inputs, as filed or as restated. Ask for the same comps twice and you get the same sheet with the same sources.

Every cell opens to its source. The output is an Excel workbook with live formulas. Click a multiple to see the formula and each input; click an input to land on the filing line it came from. When the release and the 10-Q disagree, Kepler shows both and flags it.

Comps stay current on their own. Save your comps template as a playbook, and agents refresh it when any peer files, with a note on what moved. The same sheet is available in PowerPoint, inside Claude and ChatGPT, and through our API. Every refresh is logged and permissions mirror your firm’s. Our engineers built this layer before at Palantir, and they set yours up on site.

Try it with a question like “Trading comps for US cybersecurity, EV/EBITDA on each filer’s own adjusted EBITDA and on ours, calendarized to 2025.” For more on the tools, see the best AI tools for equity research.

Checklist before comps go in a deck

1. Every peer has a written reason for being in the set.
2. All figures share one basis: LTM through the same quarter, or the same calendar year.
3. Share price is the same date for every company, and the share count is diluted.
4. Debt is at principal, and the sheet says so.
5. Leases are treated consistently with the EBITDA you divide by.
6. Each adjusted EBITDA ties to the company’s reconciliation table, and your own definition sits beside it.
7. No restated or recast period is mixed with an as-filed one.
8. Outliers are footnoted, and negative multiples show as NM.
9. Ten random cells click through to the right line.
10. Running it again gives the same median.

Build your first comps sheet on app.kepler.ai, or book a demo to see it on your peer sets and your definitions.

Frequently asked questions

Can AI build trading comps?

Yes. Kepler, an enterprise intelligence platform for investment firms, builds a comps sheet in Excel in minutes. It pulls every historical input from the filings and market data, runs forward years on your own estimates, computes each multiple in code, applies each filer's own adjusted EBITDA or your firm's definition, and calendarizes fiscal years. Every cell is either a formula or an input cited to its line, so the sheet reproduces and any number can be checked in one click.

What is the difference between LTM and NTM multiples?

An LTM multiple divides enterprise value by the last twelve months of reported results, usually the latest fiscal year plus year-to-date results minus the prior year's year-to-date. An NTM multiple uses the next twelve months of forecasts, typically blended from two fiscal years of consensus estimates. LTM is auditable; NTM reflects where the business is heading. Most comps sheets show both.

Should comps use the company's adjusted EBITDA or a standard definition?

Show both if you can. Each company's own adjusted EBITDA, taken from its non-GAAP reconciliation, is what management and the sell side quote. A standard definition applied to every peer, such as treating stock-based compensation as a cost, makes the multiples comparable. The gap between the two is often the most useful line on the sheet.

How do you calendarize fiscal years for comps?

Weight the two fiscal years that overlap the calendar year by the months each contributes. For a company with a July year-end, calendar 2025 is seven twelfths of fiscal 2025 plus five twelfths of fiscal 2026. Where quarterly data exist, summing the four quarters closest to the calendar year is more precise. Fiscal years ending in January are usually treated as the prior calendar year.

Are operating leases included in enterprise value?

It depends on the EBITDA you divide by. Under US GAAP, operating lease cost sits inside operating expenses, so EBITDA is already after rent, and operating lease liabilities usually stay out of enterprise value. Finance leases are debt and go in. Under IFRS 16, lease cost moves below EBITDA, so lease liabilities belong in enterprise value. Keep the numerator and denominator on the same basis.

How many companies should a comps set include?

Usually five to fifteen. Fewer than five and one outlier moves the median; more than fifteen and the set drifts away from the business you're pricing. Start wide on business model, end market, size, growth and margin, then cut to the names a PM or committee would accept as true peers, and write down why each one is in.