Congressional Corruption Prediction Engine

01 / The Problem

The data exists.
Nobody connects it.

The STOCK Act requires members of Congress to disclose stock trades within 45 days. The data is public — but useless in isolation. A trade disclosure tells you a senator bought $50K of Pfizer stock. It does not tell you anything else.

It does not tell you: that senator chairs the Senate Health Committee, which healthcare bills are pending before that committee, that Pfizer lobbied that same office last quarter, and that if a specific bill passes, Pfizer's revenue impact is approximately 30% of market cap. GovGreed connects all those facts and scores them together.

What the raw data shows

Sen. X bought $50,000 of Pfizer (PFE) stock on March 15, 2025

What GovGreed adds

+Sen. X chairs Senate Health Committee (voting_power = 1.0)
+Senate Bill S.1234 (healthcare pricing) is before that committee
+Pfizer lobbied Sen. X's office Q1 2025
+If S.1234 passes: estimated Pfizer revenue impact = ~30% of market cap
→opportunity_score = 847.3 — RED alert, triple signal active

02 / The Data We Collected

Eleven federal data systems.
Zero paywalled sources.

Every data source in GovGreed is free, public, and API-accessible. No scrapers that violate ToS. No paid subscriptions. No data brokers. The challenge was not obtaining the data — it was connecting eleven data systems that use completely different identifier schemes.

Dataset	Count	Source	What it covers
Politicians	538	Congress.gov API	All House + Senate members, 119th Congress
Bills	62,815	Congress.gov + detail API	All tracked bills, fully dated + status; 42,143 scored by the ML model
STOCK Act trades	103,342	QuiverQuant (no key)	Deduplicated member stock disclosures (191,238 raw rows), ticker + date + amount
Committees + assignments	230 / 3,908	congress-legislators GitHub YAML	Who sits on what committee, with role (Chair/Member)
Companies	581	Financial Modeling Prep	Market cap, sector, exchange for all tickers traded
Campaign contributions	11,850	FEC API	Top donations per politician by company/PAC
Lobbying filings	570	Senate LDA API	Quarterly filings, sector-matched to members
Federal contracts	152	USASpending.gov	Government contracts to public companies
bill_impacts (generated)	924,012	IPS-built cross-reference	Bill→ticker impact mappings (no official source for this)

🔗

Cross-source ID normalization

Congress.gov uses bioguide IDs. FEC uses candidate IDs. The YAML file uses THOMAS IDs. Lobbying filings use registrant names. We built the matching logic to unify them all into a single member record.

📊

Ticker-to-bill matching (IPS-built)

No official source maps companies to bills. We built bill_impacts: 924,012 generated bill→ticker impact mappings linking tickers to legislation by sector keyword matching. This is the table that makes everything else computable.

🧠

Sector taxonomy unification

FEC uses one sector classification. FMP uses GICS sectors. Lobbying filings use issue codes. Congress bills use CRS subject codes. We built the translation layer that maps all four systems to a common taxonomy.

03 / The Scoring Engine

One formula.
13,052 scored predictions.

Every prediction is produced by a single deterministic scoring formula that combines committee power, bill financial impact, stock trading behavior, and campaign finance patterns into a single conflict-of-interest score.

      -- GovGreed core scoring formula

      opportunity_score =

          voting_power

          × MIN(impact_ratio, 0.50)

          × LOG(consortium_count)

          × multiplier           -- 2000x if STOCK Act trade exists, 150x if not

          × triple_bonus        -- 1.5x if all three signals active

voting_power — Committee role determines access

Chair

1.0

Ranking Mbr

0.8

Member

0.6

No seat

0.1

The multiplier — the trade is the tell

2000x

STOCK Act trade exists for affected ticker

When a committee member has filed a STOCK Act disclosure for a stock that stands to benefit from their own legislation, the score multiplies by 2,000. The trade is the smoking gun.

150x

No trade on record

Committee seat + affected bill + campaign contributions still score, but at 150x. There is a conflict of interest even without the trade disclosure.

Score distribution — 13,052 active predictions classified

RED

2.6%

Score ≥ 200

ORANGE

14.7%

Score ≥ 50

YELLOW

62.9%

Score ≥ 10

GREEN

19.7%

Score < 10

04 / The Triple Signal

When all three overlap.
752 predictions fire.

The Triple Signal fires when a single politician shows all three simultaneous indicators pointing at the same bill. This is the highest-confidence conflict-of-interest alert in the system.

1

Committee Seat

Politician sits on the committee with direct jurisdiction over the bill. Chair = 1.0 voting power. The committee controls whether the bill reaches a floor vote at all.

2

STOCK Act Trade

A STOCK Act disclosure exists for a position in the company directly affected by this bill. The filing is dated within the period while the bill is active. This is the 2,000x multiplier trigger.

3

Campaign Contribution

The politician received a campaign contribution from the company or its PAC that is affected by this same bill. FEC-sourced. This is the 1.5x triple_bonus trigger when combined with signals 1 and 2.

Top Triple Signal detections — 752 total across all 13,052 active ML predictions

Politician	Triple Signals	Company / Ticker	Signal
Markwayne Mullin	114	ConocoPhillips (COP)	RED
Gary Peters	73	Multiple sectors	RED
Angus King	58	Comcast (CMCSA)	RED

05 / The ML System

Machine learning.
Entirely in PostgreSQL.

No Python service. No model files to deploy. No infrastructure to maintain. Four PostgreSQL stored functions score the full prediction matrix in approximately 30 seconds using pure set-based SQL.

-- 4 PostgreSQL RPC functions

Function	Output	Description
extract_bill_features()	21 features × 42,143 bills	Feature extraction from all source tables
learn_investability_weights()	Weight vector	Z-score analysis of bill financial characteristics
compute_investability()	Score 0–100 per bill	Scores all 42,143 bills by financial opportunity
compute_predictions()	13,052 active rows	Cross-product: all politicians × all bills, scored

Key ML findings

▶

Politicians disproportionately invest on bills affecting many small-cap companies across multiple sectors — not single large-cap bets.

▶

Healthcare and VA legislation scores highest for cross-signal conflict of interest. Defense and energy second.

▶

Committee chairs account for 67% of RED-tier predictions despite being a small fraction of total members. Power concentration is measurable.

▶

Impact ratio is capped at 0.50 to prevent single-company bills from dominating. Small-cap + niche legislation is the real signal — not obvious big-name bets.

      -- compute_predictions() simplified: 10 CTEs, pure set-based SQL, no loops

      -- Full prediction matrix scored in ~30 seconds inside PostgreSQL.

      WITH committee_power AS (

        SELECT member_id, bill_id,

          CASE role WHEN 'Chair' THEN 1.0 WHEN 'Ranking Member' THEN 0.8 ELSE 0.6 END AS voting_power

        FROM committee_assignments JOIN bill_committee_refs USING (committee_id)

      ), trade_signal AS (

        SELECT member_id, ticker,
          CASE WHEN trade_date IS NOT NULL THEN 2000 ELSE 150 END AS multiplier

        FROM stock_trades

      ), final_score AS (

        SELECT member_id, bill_id,

          voting_power * LEAST(impact_ratio, 0.5) * LN(consortium_count) * multiplier * triple_bonus

          AS opportunity_score

        FROM committee_power JOIN trade_signal JOIN bill_impacts JOIN triple_check ...

      )

      INSERT INTO predictions SELECT * FROM final_score;

06 / Tech Stack

A full intelligence platform.
At $0/month.

Supabase free tier for the database. Vercel hobby tier for the frontend. No ML server because the ML runs in SQL. No data pipeline infrastructure because the RPCs handle it. The entire stack is free.

Layer	Technology	Cost
Frontend	Static HTML + Alpine.js + GSAP	Vercel hobby ($0)
Database	Supabase PostgreSQL (ConTrack project)	Free tier ($0)
ML / Scoring	Pure PostgreSQL stored functions (4 RPCs)	No server needed ($0)
Auth	Supabase Auth (email/password)	Included ($0)
Data sources	Congress.gov, FEC, Senate LDA, QuiverQuant, USASpending.gov, FMP	All free ($0)
Total monthly	$0/month — full production intelligence platform

HTML5Alpine.jsGSAP SupabasePostgreSQLSQL RPCs Supabase AuthVercelCongress.gov API FEC APISenate LDA APIQuiverQuant USASpending.govFinancial Modeling Prep

07 / What Made It Hard

Five problems
nobody had solved yet.

The STOCK Act data has been public for years. So has the congressional record. So have campaign finance filings. The data is not the problem. The problem is everything required to connect eleven incompatible data systems and score them in 30 seconds.

01

Cross-source normalization — six incompatible ID schemes

Six datasets use completely different identifier schemes: bioguide IDs, FEC candidate IDs, THOMAS IDs, OpenSecrets IDs, registrant name strings, and company tickers. We built the matching logic to unify them into a single member record. No two datasets agreed on how to identify the same person.

02

Ticker-to-bill matching — no official source exists

No official dataset maps public companies to legislative bills. We built bill_impacts from scratch: 38,955 generated pairings linking tickers to legislation through sector taxonomy matching between GICS codes and CRS subject headings. This intermediate table is what makes the entire scoring engine computable. Without it, there is nothing to score.

03

ML in SQL — four PostgreSQL RPCs, zero external infrastructure

Most ML systems require a Python service, a model file, a deployment pipeline, and monitoring. Ours is four stored SQL functions. Extract features, learn weights, score bills, generate predictions — all in database. When Supabase is up, the ML is up. No separate service to fail.

04

Full prediction matrix in 30 seconds — pure set-based SQL

Generating the full prediction matrix is a cross-product of politicians × bills, joined against six tables, scored with a multi-factor formula, and filtered into risk tiers. We did this in 10 CTEs using pure set-based SQL with no application-layer loops. The database engine does the work. No iteration. No row-by-row processing.

05

$0/month — Supabase free + Vercel hobby

A full congressional intelligence platform with authenticated access, 13,052 active ML predictions, live ML scoring, and a public-facing dashboard — all running at zero infrastructure cost. Supabase free tier for the database. Vercel hobby for the frontend. The value is the data connections and the scoring logic. Not the servers.

Congressional Corruption
Prediction Engine.

The data exists.
Nobody connects it.

Eleven federal data systems.
Zero paywalled sources.

Cross-source ID normalization

Ticker-to-bill matching (IPS-built)

Sector taxonomy unification

One formula.
13,052 scored predictions.

voting_power — Committee role determines access

The multiplier — the trade is the tell

Score distribution — 13,052 active predictions classified

When all three overlap.
752 predictions fire.

Committee Seat

STOCK Act Trade

Campaign Contribution

Top Triple Signal detections — 752 total across all 13,052 active ML predictions

Machine learning.
Entirely in PostgreSQL.

Key ML findings

A full intelligence platform.
At $0/month.

Five problems
nobody had solved yet.

Ready to build something
like this?

Congressional CorruptionPrediction Engine.

The data exists.Nobody connects it.

Eleven federal data systems.Zero paywalled sources.

Cross-source ID normalization

Ticker-to-bill matching (IPS-built)

Sector taxonomy unification

One formula.13,052 scored predictions.

voting_power — Committee role determines access

The multiplier — the trade is the tell

Score distribution — 13,052 active predictions classified

When all three overlap.752 predictions fire.

Committee Seat

STOCK Act Trade

Campaign Contribution

Top Triple Signal detections — 752 total across all 13,052 active ML predictions

Machine learning.Entirely in PostgreSQL.

Key ML findings

A full intelligence platform.At $0/month.

Five problemsnobody had solved yet.

Ready to build somethinglike this?

Congressional Corruption
Prediction Engine.

The data exists.
Nobody connects it.

Eleven federal data systems.
Zero paywalled sources.

One formula.
13,052 scored predictions.

When all three overlap.
752 predictions fire.

Machine learning.
Entirely in PostgreSQL.

A full intelligence platform.
At $0/month.

Five problems
nobody had solved yet.

Ready to build something
like this?