These files are model OUTPUTS 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 ground truth requires your own WRDS access.

Forma ProForma-20Q Forecasts in USD ($M)

Probabilistic forecasts of complete quarterly financial statements for U.S. public firms, denominated in millions of USD, at horizons of 1–20 quarters.

This is the dollar-space companion to the standardized-space release. Forma internally predicts a per-account, per-quarter asinh z-score; that is the right space for scoring and the wrong space for almost everything else. This dataset inverts the full normalization chain so the forecasts can be merged against Compustat, analyst forecasts, or prices without reimplementing the pipeline.


Rows	460,773,192 (0 null)
Firms × accounts × horizons	~1.17M firm-quarters × 78 accounts × 20 horizons
Target quarters	2010Q1 – 2024Q4 — the test period
Origin quarters	2008Q1 – 2024Q3
Size	12.09 GB, Hive-partitioned by `target` (78 partitions, ~165 MB each)
Model	Forma, Gaussian β-NLL, transformer-only — 5-seed equal-weight mixture

Quick start

import pandas as pd

# one account only -- 165 MB, not 12.09 GB
fc = pd.read_parquet("hf://datasets/forma-lab-mccombs/forma-usd-forecasts/target=revtq/")
fc = fc[fc.forecast_horizon == 4]
fc[["firm_id", "target_quarter", "pred_p05_musd", "pred_p50_musd", "pred_p95_musd"]].head()

Schema

column	meaning
`firm_id`	gvkey, zero-padded to 6
`quarter`	forecast origin (q0), calendar quarter end
`target_quarter`	quarter being forecast = `quarter` + `forecast_horizon`
`target`	Compustat-style account name — the partition key: a path segment, not a stored column (see below)
`forecast_horizon`	1–20 quarters
`pred_p05/p25/p50/p75/p95_musd`	predictive quantiles, $M
`pred_mean_musd`	predictive mean, $M

Use pred_p50_musd as the point forecast. The mean is exact but is not the median — the gap is the skew, and it is large in the tail (see below).

target is in the path, not in the file. Reading one target=…/ directory gives you every column above except target — there is nothing to restore it from. This matters when you merge: joining such a frame on a key list that includes "target" raises KeyError: 'target'. For a single account the column is a constant, so just put it back:

fc = pd.read_parquet("hf://datasets/forma-lab-mccombs/forma-usd-forecasts/target=revtq/")
fc["target"] = "revtq"

To pull several accounts in one read, filter from the repo root — there target is reconstructed from the paths. Note the ignore_prefixes: this repo's root holds README.md and LICENSE.md beside the partitions, and pyarrow's dataset discovery would otherwise try to open them as parquet and fail with ArrowInvalid: ... Is this a 'parquet' file?. (_meta/ needs no help — the _ prefix already excludes it.)

fc = pd.read_parquet("hf://datasets/forma-lab-mccombs/forma-usd-forecasts",
                     filters=[("target", "in", ["revtq", "niq"])],
                     ignore_prefixes=["_", ".", "README", "LICENSE"])

ignore_prefixes is a pyarrow argument; fastparquet rejects it with TypeError — and does not need it, since it globs for *.parquet rather than walking every entry, so under that engine the root read works unmodified.

Why quantiles and not a standard deviation

The back-transform V = sinh(a + b·Z)·S/k is strictly monotone, so quantiles map through it exactly and interval coverage is identical in z-space and dollars.

Moments also have closed forms, but the second moment carries exp(2b²) with b = σ_reg·σ_model, whose p99 is 3.11 and max 5.69 — so exp(2b²) reaches 1e28. A Gaussian in z-space is lognormal-tailed in dollars; a standard deviation would be meaningless for the top few percent of cells and would silently dominate any pooled statistic. It is deliberately not shipped.

What sample this is

Every forecast here has a target quarter in the test period (2010Q1–2024Q4). Nothing in the file was fit on: the model trained on 1971–2001 and was validated on 2002–2009, and forecasts of validation-period targets are not released.

Origins run earlier than targets do, from 2008Q1, because a forecast made in 2008Q3 at horizon 6 lands in 2010Q1. That is a legitimate out-of-sample forecast and it is kept. The consequence is that horizon coverage is truncated at both ends, which matters if you compute anything per-horizon:


early origins	lose their SHORT horizons — 2008Q1 keeps h=8..20, since h=1..7 would land before 2010Q1
late origins	lose their LONG horizons — 2023Q4 keeps h=1..4, 2024Q3 keeps only h=1

The late-end truncation is not a filter: the model is never asked to forecast past the end of the data, so no target after 2024Q4 exists at any horizon.

Take horizon-balanced subsets from the interior if you need equal coverage, and do not read a thinning tail as a change in model behaviour.

Relation to the paper. The scored sample in the paper is smaller than this file — it additionally requires the ORIGIN to be in the test period, and takes the intersection across every model in the pool. To reproduce paper numbers, use the standardized-space forecasts and the published evaluation mask in proforma-20q-artifacts; this file is a superset and will not reproduce them on its own.

Three traps when merging

Join on target_quarter, not quarter. quarter is the origin.
26 of 78 targets are not Compustat columns, or do not mean the same thing under them. 20 are quarterly flows derived by differencing YTD items (oancfq←oancfy, …); 6 are composites. Merging the YTD column directly is silently wrong — Q4 median oancfq/oancfy = 0.308. wcapq is worse: a native Compustat column of that name exists and merges cleanly, but these forecasts are of actq − lctq. Use build_usd_truth.py, which rebuilds the truth panel from your own Compustat pull. It lives in proforma-20q, is public and ungated, and imports the de-cumulation, the formula table and the target universe from proforma20q itself, so it cannot drift from the benchmark. It also refuses to emit a target it could not actually derive, rather than passing a native column of the same name through as truth.

It is the only supported truth builder. Earlier versions of this dataset shipped a standalone copy as _meta/build_usd_truth.py; that copy restated the definitions locally and had no derivability check, and it has been withdrawn. If you have it, discard it and use the proforma-20q one.
Filter forecast_horizon or your panel multiplies ~10× (median 10, max 20 origins forecast the same target quarter).

Full recipe, including I/B/E/S information-set alignment (rdq lags the fiscal quarter end by a median of 37 days) — see _meta/FORECAST_USD_RELEASE.md, which ships with this dataset alongside the build and partition reports.

Calibration

Inherited from the model and unchanged by the transform: 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). A monotone transform cannot alter this — it is a property of Forma, not of the conversion.

Related releases


Model code (no weights)	`forma-lab-mccombs/forma-release` (Apache-2.0)
Benchmark + builder	`forma-lab-mccombs/proforma-20q` (Apache-2.0; hosts `scripts/build_usd_truth.py`, ungated)
Standardized-space forecasts + sample mask	forma-lab-mccombs/proforma-20q-artifacts (same licence, gated)
Trained weights + regularization stats	forma-lab-mccombs/forma (non-commercial, no WRDS condition)

Licence and data provenance

Released under the Forma Non-Commercial Research Licence (WRDS-Conditioned) v1.0 (LICENSE.md) — not a CC licence. Non-commercial academic research only; you must hold your own current Compustat/WRDS licence; no redistribution; attribution required.

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

Citation

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

Downloads last month: 46

Total file size:

12.1 GB