Forma — pro-forma financial statement forecaster
Forma jointly forecasts the complete quarterly financial statements of U.S. public firms at horizons of 1–20 quarters, emitting a predictive distribution per statement item.
This repository holds the trained weights and the artifacts required to run or
score them. The model code is separate and Apache-2.0:
forma-lab-mccombs/forma-release.
Contents
| file | what |
|---|---|
checkpoints/forma_gaussian_seed{60..64}.ckpt | Gaussian β-NLL fleet, 5 seeds (11.4 MB each) — the squared-error / point track |
checkpoints/forma_laplace_seed{60..64}.ckpt | Laplace fleet, 5 seeds — the absolute-error track |
reference/regularization_stats.parquet | per-(account, quarter) μ / σ / k — required to encode inputs or decode outputs |
reference/account_id_map.csv | account name ↔ id; the account embedding is indexed by this, so it must match training |
The weights are inert without the regularization statistics. Forma predicts a per-account, per-quarter asinh z-score; without (μ, σ, k) you can neither build inputs nor interpret outputs. They ship here for that reason.
Training configurations live in the Apache-2.0 code repository
(forma-release), not here.
No firm-identifier map is distributed: Forma consumes no firm identity — there is
no firm embedding — so none is needed to run the model.
Model
Gaussian β-NLL, transformer-only (no constraint-correction layer, no reconciliation), ~1M parameters. Trained on Compustat quarterly filings 1971–2001, validated 2002–2009, tested 2010+. Universal forecaster: observable accounting and industry data only, no firm-specific identifiers.
The released forecasts are the 5-seed equal-weight mixture of the Gaussian fleet, not any single seed.
Intended use and limitations
Academic research on financial-statement forecasting. Not investment advice;
see LICENSE.md §9.
- Calibration. The 90% interval is close to nominal (0.905–0.939 measured); the central 50% interval is empirically over-wide (~0.58–0.69 vs 0.50).
- Reg stats stop at 2024Q4. Applying Forma to later data requires re-estimating (μ, σ, k) from your own Compustat pull — the builder is in the Apache repo. Re-estimated statistics define a different target space, and scores computed against them are not comparable to published numbers.
- Financial firms are excluded (SIC 6000–6999).
- Dollar-space skew. If you back-transform to dollars, note that a Gaussian in z-space is lognormal-tailed in dollars: use quantiles, and treat the predictive standard deviation as unusable in the tail. See the USD forecast dataset card.
Related
| Model code, configs, competitors | forma-release (Apache-2.0) |
| Benchmark package + builder | proforma-20q (Apache-2.0) |
| Forecasts in USD ($M) | forma-lab-mccombs/forma-usd-forecasts (WRDS-conditioned, gated) |
| Standardized forecasts + sample mask | forma-lab-mccombs/proforma-20q-artifacts (WRDS-conditioned, gated) |
Licence
Forma Non-Commercial Research Licence v1.0 — see LICENSE.md.
Non-commercial academic research only; 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).
Downloading the weights does not require a WRDS/Compustat licence: no Compustat records are distributed here and firm-level values cannot be reconstructed from ~1M model parameters. Rebuilding the training or evaluation data does require your own WRDS access — see the benchmark repo.
Some of these weights were previously distributed under Apache-2.0. That grant
is irrevocable as to copies already obtained; this licence governs materials
obtained here and all subsequent versions (LICENSE.md §10).
Derived from Compustat under UT Austin's academic subscription. S&P Global Market Intelligence retains all rights in the underlying data.
Citation
See CITATION.cff in forma-release.