Datasets:

forma-lab-mccombs
/

proforma-20q-artifacts

license: other
license_name: forma-nc-wrds-1.0
license_link: LICENSE.md
language:
  - en
tags:
  - finance
  - accounting
  - financial-statement-forecasting
  - probabilistic-forecasting
  - benchmark
  - forma
  - proforma-20q
pretty_name: ProForma-20Q evaluation artifacts
size_categories:
  - 100M<n<1B
extra_gated_prompt: >-
  These files are model OUTPUTS and sample-definition indices derived from
  Compustat via WRDS under The University of Texas at Austin's academic
  subscription. They are released under the Forma Non-Commercial Research
  Licence (WRDS-Conditioned) v1.0 — see LICENSE.md. Access requires that you
  hold your own current Compustat/WRDS licence, use the files for non-commercial
  academic research only, and do not redistribute them. No raw Compustat values
  are distributed here; rebuilding the benchmark data requires your own WRDS
  access, which you need in order to use these artifacts at all.
extra_gated_fields:
  Full name: text
  Institution: text
  Institutional email: text
  I hold a current Compustat/WRDS licence: checkbox
  I will use these files for non-commercial academic research only: checkbox
  I will not redistribute these files or firm-level values derived from them: checkbox
  Intended research use: text
extra_gated_auto_approve: true
extra_gated_button_content: Accept licence and download

ProForma-20Q evaluation artifacts

The artifacts needed to reproduce the paper's pooled evaluation without regenerating every model: the scored forecast files in standardized space, the sample-coverage mask that defines the common evaluation footprint, and the per-horizon calibration series.

This complements — it does not replace — the benchmark package. The pipeline, task definition, evaluation code, and one-command WRDS rebuild live in proforma-20q (Apache-2.0).

These forecasts are in the regularized (asinh z-score) target space, which is what the evaluator scores. If you want dollars, use forma-usd-forecasts instead — do not try to interpret these values as currency.

file	what	size
`forecasts/forma_fgrid__pf_full__test__predictions.parquet`	Forma Gaussian 5-seed mixture — the squared-error / point track (Panel A)	3.98 GB
`forecasts/ffnn_linear_b50__pf_full__test__predictions.parquet`	FFNN (linear) 5-seed mixture — Panel A comparator row	4.77 GB
`forecasts/ffnn_large_b50__pf_full__test__predictions.parquet`	FFNN (large) 5-seed mixture — Panel A comparator row	4.78 GB
`forecasts/forma_lap05_fgrid__pf_full__test__predictions.parquet`	Forma Laplace 5-seed mixture — the absolute-error track (Panel B)	7.99 GB
`forecasts/forma_lap05_fgrid__pf_full__test__predictions.nll.json`	density-family sidecar for the above — see the warning below	<1 KB
`mask/full_sample_mask_bits.npy`	the 327,244,429-cell Full-sample coverage bitmap (grid-aligned packbits; no firm identifiers)	69 MB
`mask/full_sample_grid_rows.parquet`	canonical row index `(grid_row, firm, origin)` the bitmap is a bitmap over	0.9 MB
`checksums/canonical_column_stats.json`	canonical per-column drift statistics (coverage, mean, sd, p05/p50/p95 per regularized column) — what `proforma20q report-drift` compares against	1.7 MB
`calibration/forma_fgrid_calibration/coverage_by_horizon.csv`	Forma Gaussian per-horizon PIT, z², CRPS and 50/80/90/95% interval coverage — the series §5.2 quotes	3 KB
`calibration/<model>_calibration/coverage_by_horizon.csv`	the same series for the Laplace, FFNN and Chronos comparators	~3 KB each

The calibration series ship for every model in the figure, including ones whose forecast files are not hosted here: they are results the paper promises (§5.2, "per-horizon series in the release"), and they are kilobytes, not gigabytes. Every model's series sits under its own <model>_calibration/ directory, the canonical Gaussian one included — it was briefly published at calibration/coverage_by_horizon.csv, so update any pin still pointing there. The bytes are unchanged (md5 a0949f70…).

⚠️ Keep the .nll.json sidecar next to its parquet. The evaluator resolves the density family from {stem}.nll.json and silently defaults to Gaussian when it is absent. A Laplace forecast scored as Gaussian produces plausible, wrong numbers with no error. If you move or rename the Laplace parquet, move the sidecar with it.

Why the mask ships as a bitmap plus a row index

The paper's headline numbers are computed on the intersection of every model's finite predictions with the truth — 327.2M cells for the Full sample. Recomputing that intersection requires every model's forecasts, which is a multi-hundred-GB download nobody wants. The bitmap encodes membership directly.

full_sample_grid_rows.parquet is what makes the bitmap portable. It is the canonical (grid_row, firm, origin) index the bits are indexed against, so the mask still applies to a rebuild from a different Compustat vintage, matched by (firm, origin) value rather than by position. It is 0.9 MB and it cannot be derived from the forecasts — the canonical grid contains rows no model ever forecast. Pass both:

proforma20q evaluate my_forecasts.parquet --sample-mask full_sample_mask_bits.npy \
                                          --grid-rows full_sample_grid_rows.parquet

What is deliberately not here

No Compustat records. The truth panel is rebuilt from your own WRDS pull with the benchmark package; that is why a WRDS licence is a precondition rather than a formality.
No dollar values. See forma-usd-forecasts.
Not every competitor forecast. The two FFNN mixtures ship here because they are the learned Panel A comparator rows the benchmark's scripts/download_artifacts.py pins for direct rescoring. For the remaining competitors — Chronos, chained GBM, naive, ElasticNet, random forest — the paper promises seeded regeneration scripts, not outputs; they are regenerated from the Apache-2.0 repositories rather than hosted here.
No regularization statistics. Those live once, with the model, in forma-lab-mccombs/forma, so there is a single source of truth for the (μ, σ, k) that define the target space.

Provenance

This repository is the canonical and sole distribution point for these files. proforma-20q's scripts/download_artifacts.py pins an md5 for every file it fetches and verifies each download against it, so a truncated or substituted transfer fails loudly rather than scoring silently against the wrong bytes.


Benchmark package, builder, evaluation code	`proforma-20q` (Apache-2.0)
Model code, configs, competitor code	`forma-release` (Apache-2.0)
Trained weights + regularization statistics	forma-lab-mccombs/forma
Forecasts in USD ($M)	forma-lab-mccombs/forma-usd-forecasts

Licence

Forma Non-Commercial Research Licence (WRDS-Conditioned) v1.0 — see LICENSE.md. Non-commercial academic research only; you must hold your own current Compustat/WRDS licence; no redistribution; attribution required.

All commercial rights are reserved to The University of Texas at Austin. For commercial licensing, contact UT Discovery to Impact (ip@discoveries.utexas.edu).

Derived from Compustat under UT Austin's academic subscription. S&P Global Market Intelligence retains all rights in the underlying data. Nothing here is investment advice.

Citation

Cite the companion paper; see CITATION.cff in forma-release.