evidence / Calibration & market
The Devig Stack -- market math from scratch, used as the yardstick I grade against
I implemented four de-vig methods from scratch -- including the Shin (1992) insider-trading model via a numerically stable bisection solver -- wired them into production, and then used the devigged closing line as the BENCHMARK MY OWN MODEL IS GRADED AGAINST. This is not a marketing prop ("look, I know Shin"): it is the honest yardstick that let me prove the market is efficient and that my model only MATCHES it. The single truth-source for any figure below is docs/JOB_EVIDENCE_PACKET.md; edge_claimed = False throughout and no dollar, ROI, or edge figure appears anywhere on this page.
strongest single receipt
the claim
A vigged sportsbook price is not a probability -- it carries the book's margin (the overround). To grade a forecaster against "what the market really thinks" you first have to strip that vig honestly, and the naive retail method (proportional / additive) is biased on favourite-longshot lines. So src/prediction/devig.py implements four methods from scratch and dispatches between them:
- proportional / additive -- symmetric power-sum normalization, the biased retail default. - multiplicative -- power-renormalization, solved by bisection for the exponent k. - power -- the closed-form n-th-root approximation. - shin -- the Shin (1992) insider-trading model, solved by a numerically stable bisection on z (the inferred informed-money fraction). Shin loads the vig asymmetrically -- more onto the longshot -- so on heavy favourites it returns a higher favourite probability than proportional does. It is the closest thing to the sharp fair price, which is why it is the default.
All four are pure functions over probabilities (with American-odds converters), production-wired behind POST /api/devig where method defaults to shin, and covered by tests. The point is not that the code exists; it is what it grades against.
receipts
| claim | committed artifact |
|---|---|
| Four devig methods from scratch; Shin via stable bisection on `z` | src/prediction/devig.py |
| Production-wired, defaults to `shin` | api/devig_router.py (POST /api/devig, method="shin" default) |
| Devig unit tests + router tests | tests/test_devig.py, tests/test_devig_router.py |
| Devigged close as the graded yardstick (MATCH/TRAILS) | scripts/platformkit/beat_the_close_scoreboard.py |
| Consolidated candidate-REJECT scoreboard | scripts/platformkit/edge_hunt_scoreboard.py |
| Recorded MATCH/TRAILS table + full write-up | docs/MARKET_EFFICIENCY_PROOF.md |
| Leak-free gate the scoring rides on | scripts/platformkit/eval_gate/walkforward.py, scripts/platformkit/eval_gate/dm_test.py |
| Committed end-to-end proof fixture (private corpora absent on clone) | tests/fixtures/proof/ |
| Honesty truth-source for every number | docs/JOB_EVIDENCE_PACKET.md (rows 81, 82, 87) |
why this matters
Anyone can call a devig library. The hire signal is what the devig is FOR: I built the market math from first principles, chose the Shin model because it is the sharpest fair price, and then pointed it AT MY OWN FORECASTER -- making the devigged close the benchmark I have to match rather than a number I advertise beating. The result: I MATCH the efficient close and beat nothing pregame, and I copied that verdict out without softening it. The instrument that proves it -- honest de-vig plus a leak-free gate -- is the credential.
edge_claimed = False everywhere. Every number is Brier / RMSE / row-count vs a real out-of-sample corpus, never a dollar figure. Retracted measurement artifacts appear only in [JOB_EVIDENCE_PACKET.md](../JOB_EVIDENCE_PACKET.md), never on this page.
reproduce
# devig math + router behaviour (runs on any clone)
python -m pytest tests/test_devig.py tests/test_devig_router.py -q
# pregame MATCH baselines via the committed fixture (proves the pipeline end-to-end)
python -m scripts.platformkit.beat_the_close_scoreboard --corpus tests/fixtures/proof
# consolidated market-efficiency / candidate-REJECT scoreboard (recorded table; <60s)
python -m scripts.platformkit.edge_hunt_scoreboardThis site reports calibration and sharpness only, never a dollar edge, ROI, or bankroll result. An honest null is a success.