peerset.ai

Methodology

One rule governs everything below: a wrong number is worse than a missing one. Where the data cannot support an answer, peerset says so instead of estimating.

Sources — as filed, nothing else

All figures come from official public filings and registers: SEC EDGAR (XBRL, US filers), EU/UK ESEF filings, FFIEC bank call reports, USAspending federal award records, GLEIF LEI records and IRS nonprofit filings. We store what was filed — no estimates, no analyst adjustments, no forecasts. When an issuer restates, the old value is superseded and kept, never overwritten: the correction history stays visible.

The model never computes a number

Questions are translated into parametrized database queries. Every aggregate — medians, percentiles (percentile_cont), ratios, growth rates — is computed in PostgreSQL over the full population, never by the AI model. A numeric post-check compares every figure in the written answer against the database results; a figure that matches nothing is rejected and the answer falls back to a deterministic, database-only rendering. Same question, same answer — temperature 0, SQL underneath.

Definitions we commit to

• Peer groups follow the SEC's SIC industry classification. The population size (n) is stated in every answer and the group can be steered ("SIC 3827 only").
• EBITDA is computed as-filed: operating income plus depreciation & amortization from the same fiscal-year filings — never "adjusted EBITDA". It is labeled as computed wherever it appears.
• Ratios divide figures from the same fiscal year; balance-sheet instants are matched to the fiscal-year end (±7 days). Implausible scale artifacts (|ratio| > 10×) are excluded and the exclusion is part of the query, not an editorial choice.
• Federal awards are counted as obligations in the federal fiscal year of the award action, and only complete fiscal years are reported. Group consolidation follows evidenced identity links; SAM child registrations without their own public anchor are shown separately — a visible gap rather than a silent misattribution.

Every figure carries its proof

Each input figure links to a permanent fact URL — SEC accession number, reporting period, retrieval hash. Each computed figure links to a live recompute page showing the result, every input row and the exact SQL. You are never asked to trust; you are given the means to check.

Known limits

No share prices, multiples or valuations (trading multiples on licensed end-of-day data are on the roadmap). No private companies — benchmarks run against the listed universe. No forecasts, ever. US GAAP is the deepest coverage; EU/UK ESEF is available for single-company lookups.

← Back to peerset.ai  ·  Data license