Research files

Forecasts, outcomes, and code for checking the published results. The trade-model archive preserves the inputs and rules behind the October release.

2026 forecast results

Scored October 18, 2026, using regular-season outcomes from July 16 through September 27.

ShapeCast: 284 of 387 pitchers met the required 75 batters faced. Forecast error was 1.3% lower than the results-only baseline; the 95% interval ranges from 1.9% worse to 4.1% better.

Miguel Vargas: 28 later pitches fell in the registered high-away four-seam region. The required sample was 75, so the registered tests remain unevaluated.

Read the scoring report.

The bundle includes outcome counts, the frozen files, and a Python replay script. It reproduces the scoring without refitting the forecasting model. ShapeCast’s protocol was saved after some target games had occurred, before their outcomes were imported into the workspace; the report preserves that timing limitation.

Trade-model archive

Model 2026-10-17.1, captured October 17, 2026. The underlying player and playoff inputs are dated July 31, 2026. The capture date does not make those inputs current.

Unpack the archive and follow its README to verify the files and replay a player’s value. It includes the original settings and model code. A later revision is saved separately rather than replacing this copy.

Player case studies explain how salaries, option paths, and distant projection years affect the estimates. The transaction ledger keeps the dates and terms of historical moves separate from these later calculations.

Source data

The scoring outcomes were downloaded from Baseball Savant. The manifest records the requested date windows and each downloaded file’s checksum. The trade database uses the sources listed in Methods & data.

New trade-data downloads go to a separate review directory. A candidate must pass checks for missing players, projection years, and contract paths before it can replace the published July inputs.