Point-in-Time Company Fundamentals

MCP demonstration · verified sample

Apple (AAPL) · annual revenue · mode=as_of · knowledge cutoff December 31, 2025. This replays the annual value available by that date, with the reported period and knowledge timestamp kept alongside it.

• Reported period: fiscal year ended September 27, 2025 (12M_FY)
• Revenue: 416,161,000,000 USD ($416.161 billion; raw units, filer currency USD)
• Known at: October 31, 2025 at 10:01:26 · source: measured filing acceptance time
• Vintage: version 1 · zero_look_ahead=true · this observation is marked as a derived identity from filed data
• API request: GET /api/v1/companies/AAPL/fundamentals?metric=total_revenue&mode=as_of&as_of=2025-12-31&duration=12M_FY
• Open the interactive API documentation · Requires the fundamentals.read scope for licensed API access.

Scale and completeness · measured 2026-10-29

The figures behind the claims, so a buyer can check them rather than take them on trust. Published live at /api/coverage.

• 50,307,195 observations across 17,020 companies, covering periods from 1991 to 2027
• 3,286,802 superseded observations retained, the same figure kept up to 16 vintages deep. This is the part most data does not keep: a restatement here does not destroy what it replaced
• Knowledge time on 99.1% of observations (49,832,401 of 50,307,195), resolved from measured filing acceptance time rather than an inferred date, with its source named on every row
• 12,233,686 observations labelled derived rather than as-filed, so a quant can exclude them. A derived value is never presented as a tagged one
• More companies than the live universe: 17,020 against 13,544 currently filing, because delisted and deregistered filers are retained. That difference is what makes a survivorship-free study possible

What data it carries

Statement fundamentals with their revision history and knowledge time preserved, rather than a single current-state value that overwrites what came before.

• Three retrieval modes · latest (current view), original (as first filed) and as_of (as known on a date you choose), never collapsed into one another
• Revision vintages · each restatement kept as its own observation, so a superseded figure is still retrievable rather than destroyed by the correction
• Knowledge time · when a value actually became knowable, resolved from filing acceptance where available and reported with its source so you can see how it was determined
• Period structure · fiscal period, period start and end, and duration, with duration applied as a hard filter so a three-month figure is never returned where an annual one was asked for
• Provenance · the accession and tag behind each observation
• Earnings-announcement calendar · for every 10-K and 10-Q, the date results were separately announced in an earnings 8-K alongside the date the formal statement was filed, and the gap between them. This is what lets knowledge time fall on the announcement rather than the filing, and about 28% of reports have no separate announcement, reported as such rather than guessed

Why this is a separate product

Ordinary financial APIs answer “what is the figure for FY2023?”. That question has a different answer depending on when you ask it, and most data silently gives you the newest one.

• A backtest built on restated fundamentals is look-ahead biased: it trades on numbers that were not published until months later.
• A transaction multiple computed from today’s EBITDA is not the multiple the buyer paid: the target’s EBITDA at the time is the only correct denominator.
• This dataset answers with what was available at or before your as-of date, so a historical study can be reproduced rather than merely asserted.
• Units are returned raw, exactly as filed, and labelled, not silently rescaled to match another endpoint’s convention.

Best real-life usage

• Honest backtests: build a factor or screen on the fundamentals that were genuinely knowable at each rebalance.
• Transaction multiples: value a deal against the target’s figures as of the announcement date.
• Restatement research: measure how often, how far and in which direction a filer revises.
• Model audit: reproduce what a model saw on the day it made a call.

Combine with other datasets

Each analysis below pairs this dataset with another to answer a question neither can alone.

Over MCP with AI agents

Connect an AI agent to the CSIMarket MCP server (mcp.csimarket.com) and it can call these endpoints as tools, no glue code. Ask in plain language; the agent picks the endpoints, fetches the data, and composes the answer.

• “When did this company actually announce results, and how many days before the formal filing?”
• “What financial information was available for Apple as of June 30, 2024, as it was known then, not as later restated?”
• “Show me the originally reported figure and every subsequent revision for this company.”