Greederboard doesn't just count what happened. It projects what's next, checks itself against what actually happens, and publishes the error. Everything on this page is reproducible from the open data.
House stock-trade reports: where they're heading
House members filed 410 periodic transaction reports in 2026 through last month. At the current pace the year ends near 529, and 2027 lands around 499 if no ban becomes law. Volume fell from 832 reports in 2018 to 445 in 2024, then climbed back to 518 in 2025.
Monthly reports filed with the House Clerk, last 5 years, with a 15-month forecast (dashed) and an 80% band. Model: Holt-Winters with yearly seasonality (α=0.1, β=0.02, γ=0.15), parameters picked by rolling-origin backtest. On the last 22 months it missed by 6.4 reports per month at a 3-month horizon, against 8.7 for the naive "same month last year" guess: a 27% improvement. Monthly counts are noisy, so treat single months loosely.
The research queue, ordered by a model
346 members still need a full score. Researching one takes real time, so the order matters. A ridge regression trained on the 193 scored members predicts which unscored members most likely carry the highest scores, from public signals alone: trading volume, late filings, spouse trades, campaign money, attendance, seniority, committee posts.
Party and ideology are deliberately left out of this model. When we tried including them, ideology became the largest weight, which would have steered research toward one party. The queue decides who gets looked at next, so it has to be blind to party. And the model never produces a public score for anyone: a score needs receipts.
What predicts a high score
Standardized ridge weights (λ=10). Orange pushes a predicted score up, green pulls it down.
How good is it? Honestly: modest
Leave-one-out error 7.8 points vs 8.8 for guessing the average; R² = 0.20. Public signals only catch part of self-dealing, which is why the research step exists.
Self-learning: the model's track record
Every build saves the model's predictions for the unscored members. When one of them is later scored by hand, the prediction gets checked against the real result and the error is logged here. Then the model is refit with the new member included. That loop runs on every update.
Build
Trained on
CV error
R²
Predictions later checked
Real-world error
Bias
2026-10-08
193 members
7.8
0.201
pending
—
—
2026-10-08
169 members
7.68
0.308
24
17.5 pts
+17.5
The first check went badly, and that's the useful part. The 24 members researched first were the ones the model ranked highest, and it overshot them by 17.5 points on average (real-world error 17.5 vs 7.68 in cross-validation). That's the winner's curse: picking the top predictions selects for the ones that happen to be too high. Heavy trading turned out to predict a high stock sub-score but not a high overall score. The model has been refit with those 24 members included, and the queue order is what it's for; its numbers aren't published for anyone.
Next on the forecast list: a pledge tracker for the incoming 120th Congress, and estimates of trading activity under each version of the ban.