Delphi Labs Ventures / GetEven

Predict the direction of congressional trades, before disclosure.

GetEven scores tickers for the probability that a specific (member, ticker, sector) triple is about to become an interesting STOCK Act disclosure — using only legally observable congressional signal: committee assignments, hearings, lobbying filings, earmarks, IG reports, contract recompetes, and ticker news.

Status · In development Domain · Public-record predictive intelligence Surface · Tiered signals + validation

The pitch — stated honestly

Not "copy Pelosi". A scored forecast on legally observable context.

The pitch is not a copy-trading product. The pitch is: given everything legally observable about Congress's recent activity around a ticker, output a tiered prediction that a specific (member, ticker, sector) triple is about to become an interesting disclosed trade.

Every prediction carries a conviction score in [0, 100] from a transparent six-factor scorer, a tier, a stable prediction_unit_hash, and a 66-day validation window that the system later resolves against the public record as matched, missed, or pending.

Tier	Meaning
`background`	Observed, scored, archived. Not surfaced.
`watch`	Above-baseline conviction. Tracked for follow-through.
`active_signal`	Crosses the live-signal threshold. Shipped to subscribers.
`fired_even_signal`	Highest conviction tier. The bar is intentionally hard.

Why this problem is hard

Five ways this domain quietly destroys naive backtests.

01 · Look-ahead bias

The STOCK Act delay is a trap.

Members can wait up to 45 days between trade and disclosure. Any backtest anchored to tradeDate reports alpha a public subscriber could never have captured. GetEven reports two returns side by side and refuses to pretend they are the same number.

02 · Scoring ≠ predicting

Calibration before claims.

Scoring already-disclosed trades is signal scoring calibration, not prediction. The current surface is labeled as such, and the pre-disclosure predictive track ships as a distinct, separately evaluated product.

03 · Survivorship

Point-in-time committee state.

Scoring a 2018 trade with 2026 committee assignments is leakage. GetEven resolves committeesAsOf(member, date) from historical membership across the 115th–119th Congresses so every score sees only what was knowable then.

04 · Cluster contamination

Boosts must look backward only.

Cluster boosts — up to 1.40× — are the largest single multiplier in the model. Naive grouping retroactively credits the earliest trade in a 14-day window with future trades. Fix: per-trade cluster context anchored at (disclosureDate ?? tradeDate), looking back 14 days only.

05 · Multiple testing

Deflated Sharpe, not p-hacking.

Every weight profile, amplifier, and gate variant is another trial. Without correction, significance is inevitable and meaningless. GetEven evaluates against the full trial ledger and reports the Deflated Sharpe Ratio.

Honest evaluation surface

Two returns. One framing. No marketing arithmetic.

The single biggest gap between marketed alpha and shippable alpha in this domain is the STOCK Act disclosure delay. GetEven's evaluation surface emits both returns and labels them honestly:

Field	Entry anchor	Question it answers
`memberReturn90d`	`tradeDate`	Was the member right?
`actionableReturn90d`	`disclosureDate + 1 trading day`	Could a public subscriber capture it?

Every prediction is later resolved against the congress_trades_raw scan inside the 66-day validation window, producing matched, missed, or pending labels. Resolved labels feed PSR, DSR, Brier, and IC — and a prereg-lock verdict of pass, fail, or drift.

The domain in one picture

The legal observability timeline of one trade.

  ┌─────────────────────────────────────────────────────────────────┐
  │ The legal observability timeline of one congressional trade     │
  └─────────────────────────────────────────────────────────────────┘

    ▼ committee hearings, bills, IG, earmarks, news...
  ─────────────────────────────────────────▶── time
                                            ▲ tradeDate (private)
                                                   ▲ disclosureDate (≤ 45d later)
                                                   │
     GetEven sees only what is left of           public wire
     the timeline at the score date

  ┌───────────────────────────────────────────────────────────┐
  │  What the engine claims vs what it can prove              │
  └───────────────────────────────────────────────────────────┘

    Six-factor score (0..100) ──▶ tier ──▶ prediction row
                                            │
                                66-day validation window
                                            │
                                            ▼
                       congress_trades_raw scan ──▶ matched | missed | pending
                                            │
                                            ▼
                    PSR / DSR / Brier / IC on resolved labels
                                            │
                                            ▼
                         prereg-lock verdict (pass | fail | drift)

Early access to GetEven.

Operators, funds, and researchers working in the public-record predictive intelligence space can request a closed-beta walkthrough of the signal surface, the validation ledger, and the dual-return evaluation.

Request early access Back to Ventures