Developer documentation API v1

Research data and feature contracts

Join information by identity, economic meaning and decision-time availability.

Data required by the question

Input family	Required semantics	Research use
Prices and corporate actions	Listing, session calendar, raw/adjusted convention, splits, distributions and missing bars.	Comparable returns and position accounting.
Filings and fundamentals	Issuer, share class where relevant, taxonomy, fiscal period, acceptance time and amendment.	Valuation and changes in company economics.
Events and expectations	Release timestamp, original actual, pre-release consensus and revision.	Earnings or macroeconomic surprise.
Social and news	Original source, duplicate family, claimed entity, publication/availability time and licensed text coverage.	Event discovery and conflicting evidence.
Quotes and liquidity	Venue, side, size, quote age, fee inclusion and available depth.	Tradable cost and capacity.
Portfolio and financing	Balances, commitments, settlement state, exposure, borrow/funding and instrument terms.	Sizing and risk under actual constraints.
Derivatives and token exposure	Contract multiplier, expiry, mark/index distinction, margin, finality and underlying relationship.	Basis, financing and nonlinear scenario analysis.

Discover coverage before requesting a calculation

Instrument discovery and tool manifests identify available operations; evidence results retain the actual window, sources and missing reasons. An options surface, borrow rate, order book or earnings consensus is required only for a method that consumes it. A feed without that data cannot silently replace it with a guessed constant. The capability returns UNSUPPORTED_CAPABILITY or INSUFFICIENT_EVIDENCE where its preconditions fail.

Freeze the information set

POST /v1/evidence-snapshots accepts exact evidence revision references, a cutoff and either workflow_replay or publication_reconstruction. Workflow replay enforces available_at ≤ cutoff throughout the lineage. Publication reconstruction uses eligible publication vintages and is labeled separately; it cannot prove the platform observed those records historically.

Universe membership, delistings, corporate actions, entity maps and transformation inputs belong to the snapshot. Selecting only today’s surviving listings introduces survivorship bias. Backfilled history keeps its true acquisition time.

Create a workflow snapshot

POST /v1/evidence-snapshots. The server rejects ineligible references with 422 and does not silently omit them.

json

{
  "cutoff": "2026-10-01T20:05:00Z",
  "mode": "workflow_replay",
  "evidence_refs": [
    {
      "id": "ev_release_example",
      "revision": "1",
      "source_ref": "source:release:example"
    }
  ]
}

Feature registry and transformation closure

features.compute accepts an evidence snapshot and registered feature revisions. Each feature declares input kinds, lookback, lag, normalization, output units, missing-data policy and parameters. The artifact retains the full dependency closure and fit snapshot. Cross-sectional normalization is fitted on the declared eligible universe; time-series normalization uses only earlier eligible observations. Feature selection is part of the fitted pipeline.

Incremental predictive information

Rank IC compares a feature with subsequent returns across the same eligible assets at time t. Ties use average ranks; fewer than three valid pairs or constant ranks give an undefined statistic. The loss comparison uses identical held-out observations. Positive ΔL measures incremental predictive fit on that sample, not a causal effect or net trading return.

\operatorname{IC}_t=\operatorname{corr}_{\mathrm{rank}}(x_{i,t},r_{i,t\rightarrow t+h}),\qquad\Delta L=\overline L_{\mathrm{baseline}}-\overline L_{\mathrm{augmented}}

Agent research records

A research brief stores its question, horizon, evidence references, numerical artifact IDs, alternative explanations, unresolved dependencies and conditions that invalidate the thesis. Model-generated statements retain their category as interpretation. Multiple analyst roles share structured references instead of repeatedly rewriting source facts. A consensus among agents is not additional independent evidence.

Source and method references

SEC EDGAR APIsFiler identity, submission history and XBRL company facts.ALFREDOriginal releases and later revisions of economic series.TradingAgentsResearch on specialized analyst roles and structured collaboration; its backtest is not Seer performance.

Fit snapshot identity

fit_snapshot_id is an evidence-snapshot ID containing the training observations for a fitted transformation. The feature revision specifies how it is fitted and which inputs it consumes. Stateless transforms use null; a fitted transform without its required snapshot returns 422 INSUFFICIENT_EVIDENCE. The fit snapshot’s eligible training range must precede every evaluation decision that consumes it.