ProForma-20Q: reference model forecasts and the Full-sample evaluation mask

Evaluation artifacts for the ProForma-20Q benchmark: long-horizon forecasting of complete quarterly financial statements (78 line items, horizons 1-20 quarters) for U.S. non-financial firms. This record contains model outputs and a coverage bitmap only — no WRDS/Compustat-derived firm-level values. The benchmark's data environment is rebuilt from the user's own Compustat licence via the accompanying software, which provides scripts/download_artifacts.py to fetch and md5-verify these files. Contents forma_fgrid__pf_full__test__predictions.parquet — canonical Forma 5-seed Gaussian mixture forecast, squared-error track (Table 1 Panel A). md5 1820fcc90e71989af558f9d103d6fc31, 3.98 GB. ffnn_linear_b50__pf_full__test__predictions.parquet — FFNN (linear) 5-seed mixture, Panel A comparator. md5 e419c8330ff6c9c6396a7d2e04f05c3e, 4.77 GB. ffnn_large_b50__pf_full__test__predictions.parquet — FFNN (large) 5-seed mixture, Panel A comparator. md5 915779a3ff79b6e344d45910ac5e4026, 4.78 GB. forma_lap05_fgrid__pf_full__test__predictions.parquet — canonical Forma Laplace mixture, absolute-error track (Panel B). md5 1e8b0415905eeac7cf46b052f5c1cbf5, 7.99 GB. forma_lap05_fgrid__pf_full__test__predictions.nll.json — density-family sidecar for the Laplace forecast. Keep it in the same directory as its parquet: the evaluator resolves the density family from {stem}.nll.json and defaults to Gaussian without it. md5 a3d8659a201a2081dd693a8f0de051c3, 33 bytes. full_sample_mask_bits.npy — the 327,244,429-cell Full-sample evaluation mask (grid-aligned packbits; contains no firm identifiers). md5 a36008d8dbfeb56992f1049fd543d781, 68.83 MB. full_sample_grid_rows.parquet — the canonical row index the mask is a bitmap over. Pass it alongside the mask as evaluate --grid-rows so the mask applies to a rebuild from a later Compustat vintage. md5 adbc2ae6eef7f23b3576af525cfbeeec, 1.87 MB. Forecasts are keyed by gvkey, fiscal quarter, target item, and horizon, with predictions in the benchmark's regularized space. Note that quarter-end timestamps in the forecast files carry microsecond precision while a canonical truth build uses nanoseconds; compare origins as calendar quarters rather than as raw timestamps.