Non-embedding Forecasting of Financial Statements

Valuation, credit, and stress testing need joint paths for many statement lines over multi-quarter horizons — not only next-quarter EPS. The application panel is bank holding companies on FR Y-9C (preferred: FFIEC NIC BHCF bulk zips), with cheap Yahoo Finance quarterly smoke for tickers in the reporting hierarchy.

Forma-class transformers are a useful contrast (complete statements, identities, 1–20Q), but this note prioritizes transparent, testable, non-embedding methods that can move Yahoo → Y-9C without gated fine-tuned weights.

• Web Most academic work forecasts one item (earnings) cross-sectionally (Hou–van Dijk–Zhang; Li–Mohanram) or with trees (Chen–Cho–Dou–Lev).
• Web Full-statement non-transformer systems exist mainly as chained GBM (Geertsema–Lu–Ma) and multitask nets (Alberg–Lipton).
• Web Bank-native gold: Liu–Moon–Schorfheide empirical-Bayes dynamic panels on Y-9C PPNR under macro scenarios.
• Web Mixed-frequency (Ball–Ghysels MIDAS; Babii et al. sg-LASSO MIDAS) connects Yahoo daily smoke to quarterly statements.
• Web Accounting identities are linear constraints — MinT / Girolimetto–Di Fonzo reconciliation is the coherence layer.
• NIM Recommended stack: EB panel → chained GBM → reconciliation, with MIDAS for smoke and multitask FFNN as neural baseline; park Forma FT.

A. Empirical-Bayes dynamic panel

Each BHC \(i\) has coefficients \(\theta_i\) in a short dynamic regression (lags + macros). Pooling ignores heterogeneity; firm-by-firm OLS is noisy. EB / Tweedie maps the noisy \(\hat\theta_i\) toward the cross-sectional prior:

Forecasts integrate posterior predictive draws — natural for stress scenarios on Y-9C PPNR (Liu–Moon–Schorfheide).

B. Chained multi-output GBM

Predict line items in an order that mirrors articulation (e.g. revenue → expenses → NI → RE → equity → assets). Upstream predictions become features for downstream models. Approximates identities softly; pair with hard reconciliation.

C. MIDAS / sg-LASSO MIDAS

Quarterly target \(y_{t+h}\) on high-frequency \(x\) with parsimonious lag weights \(w(k;\beta)\). sg-LASSO selects groups of mixed-freq predictors across a panel of firms — Yahoo daily → Y-9C quarterly.

D. Forecast reconciliation

Base forecasts \(\tilde{y}\) projected onto \(\{y : Cy = 0\}\) (or \(Cy = c\)):

MinT chooses \(W\) from forecast-error covariances; Girolimetto–Di Fonzo covers general linear constraints and probabilistic versions. Use predictive variances — equal-weight projection can wreck small lines.

E. Quantiles / multitask nets

Quantile regressions build earnings / PPNR distributions (Konstantinidi–Pope). Small multitask FFNNs share strength across MDRMs without foundation-model fine-tuning (Alberg–Lipton).

NIM Top 6 Yahoo smoke → FFIEC Y-9C bulk:

1. Liu–Moon–Schorfheide EB panel on BHCF PPNR (and NII / noninterest splits).
2. Geertsema–Lu–Ma chained GBM on a curated MDRM set (y9c/core_items.yaml).
3. MinT + Girolimetto–Di Fonzo post-layer for identities.
4. Babii sg-LASSO MIDAS and Ball–Ghysels MIDAS for Yahoo daily → quarterly smoke.
5. Alberg–Lipton multitask FFNN as classical neural baseline.

Data paths: y9c/processed/y9c_panel_long.parquet, yahoo_bank_quarterly/processed/yahoo_bank_quarterly_long.parquet. Preferred bulk: NIC BHCF zips via scripts/download_y9c.py.

• Point: rolling-origin RMSE / MAE on change-space (or asinh-scaled) PPNR and core MDRMs; Diebold–Mariano across calendar quarters.
• Probabilistic: CRPS, PIT, interval coverage (nominal 50/90).
• Coherence: median |identity residual| / gross scale pre vs post reconciliation; never celebrate equal-weight projection alone.
• Smoke gate: MIDAS Yahoo→PPNR must beat seasonal random walk before bulk EB/GBM spend.
• No transfer claims: do not quote Forma Compustat R² as Y-9C performance.

1. Liu, L., Moon, H. R., & Schorfheide, F. — Forecasting with dynamic panel data models (NBER w25102).
2. Geertsema, P., Lu, H., & Ma, G. — Projecting financial statements with AI (SSRN 5039433).
3. Hou, K., van Dijk, M., & Zhang, Y. (2012). JAE — ICC via cross-section earnings.
4. Li, K., & Mohanram, P. (2014). RAST — EP/RI vs HVZ.
5. Ball, R., & Ghysels, E. (2018). Management Science — automated MIDAS earnings.
6. Babii, A., Ball, R., Ghysels, E., & Striaukas, J. — Panel MIDAS sg-LASSO (arXiv 2307.02673).
7. Wickramasuriya, S., Athanasopoulos, G., & Hyndman, R. (2019). JASA — MinT.
8. Girolimetto, D., & Di Fonzo, T. (2024). SMA — general linear-constraint reconciliation.
9. Konstantinidi, T., & Pope, P. (2016). CAR — forecasting risk in earnings.
10. Alberg, J., & Lipton, Z.; Chauhan et al. (ICML 2020) — multitask fundamentals.
11. Divo et al. (TMLR 2025) — Forecasting Company Fundamentals benchmark.
12. Johnson et al. — Forma / ProForma-20Q (arXiv 2608.11327) — contrast only.