GovGreed Home Bill Pass Index How It Works For Institutions
Methodology · v3.1

Bill Pass Index (BPI)

A calibrated, empirically-grounded probability that any US congressional bill becomes law.

Model v3.1-empirical-117+118 Published 2026-04-17 Sample 37,132 labeled bills AUC 0.7378 Status Live

The Bill Pass Index is a deterministic, explainable model that outputs an integer 0–100 representing the probability that a given bill in the US Congress becomes law. Unlike ad-hoc stage-based guesses, every weight in the model is calibrated against the actual outcomes of 37,132 bills from the 117th and 118th Congress. Across a held-out predictive backtest of 5,340 bills, the model achieves AUC 0.7378 with near-perfect calibration in the dense 5–14% probability band (predicted 12.2%, actual 12.29%).

Contents

  1. The prediction problem
  2. Model inputs
  3. The formula
  4. Empirical priors
  5. Backtest methodology
  6. Calibration results
  7. Known limitations
  8. Version history

1. The prediction problem

Approximately 10,000–20,000 bills are introduced per US Congress. The vast majority (>98%) die in committee. A small tail (<2%) become law. Competitor products either ignore bill-passage probability entirely or use hand-coded stage priors (e.g., "committee bills pass 22% of the time") that are not backed by any empirical distribution.

The v1 problem we solved Our own earlier model used a CASE statement that assigned bills in reported state a 71% pass probability. Against 118th Congress ground truth, the actual rate was 10.67%. A 5.7× overconfidence bias. v3.1 is calibrated against ground truth.

BPI reframes bill-passage prediction as a calibrated probability-ranking task: given a bill's current pipeline state, pipeline history, and structural features, output a probability such that bills scored at p% actually pass at a rate approximating p%.

2. Model inputs

The model reads six features per bill:

FeatureSourceWhat it captures
govtrack_statusGovTrack.us /api/v2/bill Current pipeline state (e.g. reported, pass_over_house, conference, enacted_signed).
is_aliveGovTrack.us Congressional life-cycle flag. False for bills Congress has declared dead.
states_visited[]bill_major_actions table Array of every pipeline state the bill has ever entered. A bill that once passed the House carries that signal forward.
related_bill_countGovTrack.us Number of companion bills in the other chamber. Empirically monotonic in passage rate.
introduced_dateGovTrack.us When Congress first received the bill. Penalizes very new (under 14 days).
latest_action_datebills table (Congress.gov) Last recorded action. Penalizes bills with no movement in 180+ days.

All six features are drawn from approved direct connections to federal data systems. No scraped sources. No inferred metadata. Every input is auditable at the source.

3. The formula

Simplified PL/pgSQL pseudocode (full function: calculate_pass_likelihood_v3):

IF govtrack_status IN ('enacted_signed','enacted_tendayrule', 'enacted_veto_override','enacted_unknown') OR 'ENACTED:*' in states_visited: RETURN 100 IF govtrack_status LIKE 'prov_kill_%' OR 'fail_%': RETURN 2 IF govtrack_status = 'vetoed_pocket': RETURN 3 IF is_alive IS FALSE AND not enacted: RETURN 3 ELSE: base = empirical_prior[highest_state_visited] base = CASE WHEN 'PASSED:BILL' visited THEN 95 WHEN 'PASS_BACK:SENATE' visited THEN 85 WHEN 'PASS_BACK:HOUSE' visited THEN 75 WHEN 'PASS_OVER:SENATE' visited THEN 43 WHEN 'PASS_OVER:HOUSE' visited THEN 30 WHEN govtrack_status = 'conference' THEN 50 WHEN govtrack_status LIKE 'pass_back_%' THEN 80 WHEN 'REPORTED' visited THEN 12 WHEN govtrack_status = 'introduced' THEN 1 ELSE 1 END + companion_bonus(related_bill_count): 0 → 0 · 1–2 → +2 · 3–4 → +3 · 5–9 → +5 · 10–19 → +10 · 20+ → +15 + staleness_adjustment(latest_action_date): >300d → −8 · >180d → −4 · <14d → +3 · else → 0 RETURN CLAMP(base + companion_bonus + staleness_adjustment, 1, 95)

Design choices

4. Empirical priors

Every state prior is derived from ground-truth outcomes. For each pipeline state S, we compute P(enacted | bill ever entered S) across all 117th + 118th Congress bills with terminal outcomes.

Pipeline state N bills Actual P(enact) v1 used v3.1 base
PASSED:BILL (both chambers cleared)28595.44%84%95
PASS_BACK:SENATE1392.31%34%85
PASS_BACK:HOUSE1275.00%34%75
PASS_OVER:SENATE29730.64%32%43
PASS_OVER:HOUSE70326.03%32%30
REPORTED (out of committee)1,82710.67%71%12
introduced (no subsequent action)16,5190.50%4%1

Some v3.1 base values (e.g. PASS_OVER:SENATE at 43) differ slightly from single-congress empirical rates. These were adjusted against combined 117+118 data where sample sizes are larger (43% combined vs 30.64% from 118th alone).

Companion-bill lift

Companion billsN billsPass rateLift vs base
09,8950.79%1.0×
17,1081.96%2.5×
21,2282.44%3.1×
3–48841.92%2.4×
5–91462.74%3.5×
10–19277.41%9.4×
20+2714.81%18.7×

5. Backtest methodology

A naive backtest is trivial: score each 118th Congress bill, compare against outcomes, report AUC. Because our model takes govtrack_status as input and terminal states map directly to outcomes, this produces AUC = 1.000 — which proves nothing.

The honest backtest is a predictive test: for each labeled bill, simulate scoring the bill at the moment it first entered a non-terminal state. At that moment, we know: the state just entered, states visited before it, companion count, introduction date. We do not know the final outcome. The model's score is then compared to the eventual terminal outcome (enacted = 1, died = 0).

Sample construction

For each bill B in 117th + 118th Congress: If B has no non-terminal action → skip If B is still alive at time of measurement → skip (no label) first_action = earliest non-terminal major_action event states_up_to = states visited on or before first_action score = score_v3(first_action.state, companion_count, states_up_to, blind_terminal=True) label = 1 if B ultimately enacted else 0 samples.append((score, label))
Why this is the right test This simulates live inference: the moment a bill becomes "interesting" (first non-trivial action), what does the model say about its eventual outcome? The blind_terminal=True flag strips all post-outcome information so the model must predict from pipeline state alone.

Result

0.7378
AUC (ROC)
5,340
predictive samples
639
enacted (positive labels)
11.97%
base rate
0.3184
log loss

6. Calibration

AUC measures ranking skill. Calibration is the stronger test: when the model says 30%, do 30% of those bills actually pass?

Predicted probability bucket N bills Avg predicted Actual pass rate Calibration
0 – 4%1,3242.4%0.00%✓ Excellent
5 – 14%3,25512.2%12.29%✓ Essentially perfect
15 – 24%10117.6%7.92%⚠ Overconfident (known)
25 – 34%63829.5%34.95%✓ Well-calibrated
35 – 54%2238.5%36.36%✓ Well-calibrated
The headline calibration fact 3,255 of our 5,340 predictive samples fall in the 5–14% bucket. We predict an average of 12.2% for these bills. They actually pass at 12.29%. The model is essentially perfectly calibrated at the modal density of the distribution.

7. Known limitations

7.1 The 15–24% overconfidence band

The 15–24% bucket shows 7.92% actual pass vs 17.6% predicted — the only bucket where the model is meaningfully miscalibrated. Root-cause analysis points to bills in REPORTED state with moderate staleness accumulating bonus points they shouldn't. Planned fix: dampen the REPORTED prior when latest_action_date exceeds 180 days.

7.2 Small cell counts at the top

Only 22 samples fall in the 35–54% bucket, and 0 samples in the 55%+ bands below terminal. Live-bill predictions in these higher ranges are therefore extrapolated from smaller calibration evidence. We address this by clamping live bills to [1, 95].

7.3 Stage data staleness

Our internal bills.bill_stage column is collection-time-stale; many bills in our database show stage = INTRODUCED even after enactment. GovTrack's live govtrack_status is the authoritative signal. A v3 pipeline feeding stale bill_stage as a feature would produce AUC near 0.06 (worse than random). This is an implementation pitfall, not a model flaw.

7.4 Terminal-state tautology

Bills whose current govtrack_status is already terminal (enacted, killed, failed) are not predictions — they are known outcomes. The model reports 100 / 2 / 3 for these cases. When citing "AUC 0.7378," we refer exclusively to the predictive backtest with blind_terminal=True.

7.5 Prospective validation

Our daily snapshot table (pass_likelihood_snapshots, started 2026-04-17) records v3.1 scores for every live 119th Congress bill. As these bills reach terminal outcomes over the next 12–24 months, we will publish prospective AUC and calibration measurements entirely independent of the 117+118 training set.

8. Version history

VersionDateChange
v1—Hardcoded CASE statement priors. 5.7× overconfident at REPORTED stage.
v22026-04-16Replaced v1 priors with GovTrack-aware logic. Still used guessed weights, not empirical.
v3.02026-04-17Priors calibrated against 118th Congress ground truth. AUC 0.7189.
v3.12026-04-17Combined 117+118 priors. PASS_OVER:SENATE 30→43, PASS_OVER:HOUSE 26→30. AUC 0.7378.
Reference Model function: calculate_pass_likelihood_v3(govtrack_status, is_alive, related_bill_count, introduced_date, latest_action_date, states_visited[])
Lookup RPC: get_pass_likelihood(bill_number, congress)
Bulk RPC: get_pass_likelihood_v3_bulk(bill_numbers[], congress)
Explainability: get_pass_likelihood_breakdown(bill_number, congress)
Whale EV: get_whale_opportunities(congress, min_pass, min_ratio, limit)

GovGreed Bill Pass Index is for informational purposes. Not financial advice. Congressional outcomes depend on many unmodeled factors (floor schedule, leadership decisions, current events). Past calibration is not a guarantee of prospective accuracy.

Owner: IPS Innovative Platform Solutions · team@mmamodel.ai · Published 2026-04-17.