How we build trustworthy financial data

Survivorship bias is the single most expensive hidden flaw in a backtest. Most datasets only keep the companies that are still trading today — so your strategy is silently tested against a universe that already knows which companies survived. The Enrons, the Lehmans, the RadioShacks vanish, and historical returns inflate by a percentage point or two that evaporates the moment you trade live.

Valuein retains every entity that ever filed XBRL financial statements — delisted, bankrupt, acquired, merged — with its complete filing history through its final SEC filing. Roughly half of the 19,000+ entities in the universe are no longer actively trading. They stay in the dataset; they are simply not present on dates after they stopped filing, which is exactly how they would have appeared in real time.

Point-in-time and survivorship-free are two of the trust guarantees we make. See the trust & security overview for provenance, zero-retention, and reliability — this page covers the data construction underneath them.

The most common look-ahead bias in financial data isn't malicious — it's using the wrong date column. Three timestamps live on every fact, and they mean different things.

Every PIT-safe MCP tool and SDK method accepts an as_of_date parameter. Internally, that filters on accepted_at <= as_of_date — the queryable equivalent of “what did the market know on this date?”

A figure is only trustworthy if you can prove where it came from. Every fact in the dataset carries a deterministic fact_id — a SHA-256 hash of the entity, the SEC accession, the concept, the period end, and the unit:

# Deterministic in the pipeline, recomputable in the SDK.fact_id = sha256(    f"{entity_id}|{accession_id}|{concept}|{period_end}|{unit}").hexdigest() # The SDK exposes the same function — same inputs, same id, anywhere.from valuein_sdk import compute_fact_idcompute_fact_id(entity_id, accession_id, concept, period_end, unit)

Because the hash is deterministic, the fact_id is identical in the Parquet files, the Python SDK, and the MCP server — there is no separate provenance database to drift out of sync; the identity travels with the value. Pass any fact_id to verify_fact_lineage and it resolves back to its source XBRL tag, accession ID, and filing URL for one-click verification against EDGAR.

The envelope on every fact

This envelope is the foundation for everything below: the model never has to be trusted with a number, because the number arrives already attributable, gradeable, and verifiable.

“What was in the S&P 500 on March 1, 2014?” is a survivorship trap in disguise: screen the index by its current members and you have quietly excluded everyone who was dropped. Index membership is therefore tracked the same way facts are — historically, with effective and removal dates.

The index_membership table records membership spells for S&P 500 and Russell 1000 / 2000 / 3000 with an effective_date and a removal_date per spell, using half-open [effective, removal) interval semantics. A company that left and rejoined gets two spells, not a merged one — so a point-in-time universe on any date reconstructs the index exactly as it stood.

SELECT r.symbol, r.nameFROM index_membership imJOIN references r ON r.cik = im.cikWHERE im.index_name = 'SP500'  AND '2014-03-01' >= im.effective_date  AND ('2014-03-01' < im.removal_date OR im.removal_date IS NULL);

There is no is_sp500 flag — a single boolean can only describe one index at one moment, which is precisely the snapshot bias we avoid. Membership is always a JOIN on cik.

Foreign private issuers don't file 10-Ks. They file 20-F (and Canadian issuers file 40-F), often under IFRS rather than US GAAP, with their own concept names. Those filings flow through the same standardization pipeline: concepts map into the same canonical standard_concept vocabulary, and an is_foreign flag on the entity lets you include or isolate them. The result is that a US filer and a foreign issuer answer the same query the same way.

The defining failure mode of an LLM over financial data is a confidently-stated figure that no filing supports. Valuein eliminates the failure at its root: the model is never the source of a number, and never recomputes one. Figures are born in the data layer, each already carrying its fact_id and source filing, and the agent is instructed to use them exactly.

The MCP server's provenance rules are explicit and binding on every tool response:

• Use the returned value exactly — never round, restate, recompute, or estimate it.
• Never do arithmetic on returned figures. A derived metric (growth, ratio, margin, valuation) must be requested from the tool that returns it pre-computed with its own input provenance.
• Cite the source filing or fact_id for every figure stated; verify_fact_lineage resolves it back to the filing.
• If a figure's availability is not_reported, not_mapped, suppressed, or error, state that it is unavailable — never substitute a value from prior knowledge.
• Distinguish a genuine reported zero from missing data.

The result is that a hallucinated figure is structurally impossible to pass off as data: anything the model states either carries a verifiable fact_id or is explicitly labelled unavailable. There is no third category where a plausible-sounding number can hide.

An agentic data service has two attack surfaces a plain API doesn't: untrusted filing text that could carry injected instructions, and a request path that has to authorize before it touches data. Both are handled at the boundary.

Untrusted-text fencing

SEC filing prose is attacker-controllable — a 10-K can contain text engineered to read like an instruction. Before any filing text is handed to a model, it is passed through a wrapUntrusted() fence that marks it as data, not directives, and known injection-style patterns are stripped or neutralized by a set of regex filters. The model reads the filing as evidence, never as a command.

Layered request hardening

Tiering is enforced here too: a tool a caller's plan doesn't cover returns a structured featureNotAvailable envelope with an upgrade path — never a partial or silently downgraded result.

Every claim on this page is testable from the sample tier — no token, no signup. Pick any S&P500 ticker and inspect the lineage of any fact via verify_fact_lineage:

curl -X POST https://mcp.valuein.biz/mcp \  -H "Content-Type: application/json" \  -d '{    "jsonrpc": "2.0",    "id": 1,    "method": "tools/call",    "params": {      "name": "verify_fact_lineage",      "arguments": {        "ticker": "AAPL",        "concept": "TotalRevenue",        "period_end": "2024-12-31"      }    }  }'

The response chains the standardized value back to its source XBRL tag, the SEC accession ID, and the filing URL. If we changed it, you can see why.