Skip to main content

Companies vs. Deals: The Critical Distinction

A company is an entity that exists independently of your fund. Stripe is a company. It has founders, funding rounds, employees, a product, customers. Stripe exists whether or not your fund ever talks to them. A deal is your fund’s relationship with a company. You sourced Stripe, you met with the founders, you decided to invest (or pass), you negotiated terms, you closed the investment. The deal represents your fund’s interaction with that company. Your fund can have multiple deals with the same company. You might:
  • Pass on their seed round (Deal 1: Status = Passed)
  • Invest in their Series A two years later (Deal 2: Status = Closed)
  • Consider a follow-on in their Series B (Deal 3: Status = In Diligence)
Each deal has different properties: deal source, who at your fund is leading it, what stage it’s at, when you first talked to them, what your investment thesis was, what terms you negotiated. These properties belong to the deal, not the company. Suppose two partners are looking at the same startup through different vehicles. They share the company website, founders, and past financing rounds. One vehicle can pass because of stage while the other takes the deal to IC. If changing the first record to ‘passed’ takes the company off both partners’ pipelines, you’ve put the decision on the wrong object. Meanwhile, company properties (employee count, funding history, product description) are shared across all deals. If you update Stripe’s employee count, that change applies regardless of which deal you’re looking at.
When you query deals, you join to companies to get company information. When you’re analyzing companies (building market maps, tracking sectors), you query companies directly without caring about deals.

Other entities to model

People - Founders, executives, employees. Core attributes: name, email, LinkedIn URL. People connect to companies through roles, not direct relationships. Roles - The connection between people and companies capturing who works where, in what capacity, and when. Model as a separate entity to track full history: a founder worked at Google (2018-2020), started Company X (2020-present), while advising Company Y (2021-present). Education - Where people went to school, what they studied, when. Model separately to capture multiple degrees and overlapping education. Enables analysis like “which universities produce the most founders in our focus areas?” Funding Rounds - Capital raised by companies. Attributes: round type, amount, valuation, dates, investors. Expect messiness: unreported SAFE notes, incomplete data from vendors, unannounced rounds. Your data model needs to handle uncertainty (covered in “Dealing with Messy Reality” below). LPs (Limited Partners) - Individuals and institutions who invest in your fund. Required for fund operations tools or LP portals. LPs can be individuals (HNWIs with personal info) or institutions (pension funds, endowments with organizational info). Model with a type field and nullable columns, or separate tables if managed differently. The list above isn’t exhaustive. Depending on what you’re building, you might also model: board seats, advisors, customers (for B2B portfolio companies), competitors, acquisitions, exits, fund entities (if you manage multiple funds), and more. Start with the entities your application actually needs, not a comprehensive schema you think you might need someday.

Investors and funds

Other VC firms, angels, corporate investors who participate in rounds. You care about investors for co-investment patterns, warm intro paths, and understanding who’s active in your sectors. One firm can manage several funds with different mandates. For your own investments, you need to know which fund or SPV invested: ownership, reserves, and returns belong to that vehicle. Keep decisions such as a fund passing on a deal in your deal records, not in a table of completed round participations. For external market data, you may only know that a firm participated. That’s still useful for co-investment analysis and warm introductions. An angel may invest directly without a fund at all. Don’t invent a placeholder fund to make incomplete data fit your schema. Model each participation with an investor and an optional fund. When you know the vehicle, link it. When you don’t, leave it unknown. For example:
The SQL below is simplified to illustrate entity relationships - not production-ready code.
The composite foreign key prevents you from attaching one firm’s fund to another investor’s participation. If two vehicles from the same firm invest in a round, record each participation separately. Don’t also count the firm-level vendor record as another investment. Preserve the source and when you observed it so you can reconcile later updates. Suppose your CRM names the firm, one data provider names Fund II, and another just repeats the firm name. All three observations may concern one investment. Keep the source records so you can reconcile them when the vehicle is confirmed. Summing the three observations would inflate both participation and investment totals.

Dealing with Messy Reality

Don’t assume perfect funding round data Early-stage companies raise money in ways that aren’t always captured in data sources. SAFE notes, convertible notes, rolling closes, stealth rounds. Crunchbase and PitchBook miss things, especially for pre-seed and seed companies. Your data model should allow for:
  • Unknown or estimated funding amounts
  • Approximate dates (Q2 2024, not a specific day)
  • Rounds that might not exist (rumored but unconfirmed)
  • Multiple rounds of the same type (Seed, Seed Extension, Seed II)
Don’t build validation rules that enforce “Series B must come after Series A” or “funding amounts must be known.” Real companies break these rules constantly and your data sources, even more so. People have overlapping experiences and education Founders often work multiple jobs simultaneously: running their startup while advising other companies, sitting on boards, teaching part-time. They pursue education while working: part-time MBAs, executive programs, online courses. This is why roles and education must be separate tables with start and end dates, not single fields in the people table. You need to capture overlapping time periods.
With this structure, you can query “show me all founders who worked at Google” (join people → roles → companies where company = Google and role_type = Employee). You can analyze “which schools produce the most founders in fintech” (join people → roles → companies → education, filter for fintech sector and founder roles). If you tried to store current role and education as fields in the people table, you’d lose history and couldn’t answer these questions. For an introduction, you may need to ask whether two people actually overlapped. If one left an employer in 2018 and the other joined in 2021, a shared company ID gives you a candidate connection, but the employment dates explain why the two may never have met. Keep an unknown end date distinguishable from a role you know has ended.

Temporal Data and History

Companies change constantly. Employee count grows. Funding rounds happen. Valuations change. Products pivot. You need to decide what history to keep and how to model it. Start with append-only tables Append a new row with a timestamp when you get new information about a company. This preserves the history of changes without overwriting earlier values.
Every time you fetch company data from PitchBook (daily, weekly, or on-demand), you insert a new row. You never update existing rows. This preserves full history: you can see what PitchBook reported about a company on any date. Use dbt to transform to latest values Append-only tables are great for history but inconvenient to query. You usually want the latest information, not every historical value. Use dbt (covered in Data Warehousing) to transform append-only staging tables into intermediate tables with current values:
Now you have both: pitchbook_company_changes (full history) and pitchbook_companies (latest values). Applications usually query the latest values table. When you need to analyze “how did this company change over time?” you go back to the append-only table or create a custom table for those types of queries. For standard company datasets, storage cost is rarely the main reason to discard useful history. Keeping snapshots in staging tables can be simpler than predicting which changes will matter later. Size the data and estimate storage costs rather than assuming every source is cheap to retain. The practical considerations are query performance and data warehouse costs (if you’re using Snowflake or BigQuery where you pay per query). But even there, you’re usually querying the transformed and materialized “latest values” tables, not the full historical append-only tables. Retention and deletion are covered in Security and Compliance. A snapshot captured in June tells you what the vendor reported in June. If a round closed in March and the vendor first picked it up in May, those are two different dates. Use the closing date to analyze investment activity. Use the observation date if a partner asks what you could have known at the time.

Reconstructing what a partner could see

Suppose you’re rebuilding a market map for an IC meeting held on April 1. Filtering the latest company table to firms founded before April gets you today’s descriptions and employee counts for those firms. It also includes companies the feed picked up in May. To recreate the information available for the meeting, filter the observations by when you received them, then take the last observation per company. The following query uses the fields in the history table above. The cutoff is the start of April 1 in the same time zone used for ingestion. Here the question is what the team knew before that date.
A missing company here means you have no qualifying observation for it. That’s different from observing a company and receiving a null employee count. It’s also different from having a source that backfills historical estimates today. If you promise to reconstruct what the team actually saw, those distinctions matter. As you add more data sources, follow dbt’s layered approach (staging → intermediate → marts) to stay organized. The Data Warehousing chapter covers this methodology in detail.