Do not claim: Forma OOS numbers transfer to Y-9C.
All rankings and "what we'd try" synthesized via NVIDIA NIM glm-5.3-flash;
paper anchors from manual WebSearch/WebFetch/arXiv HTML.
1. Scope
In scope: Joint (or multi-item) forecasting of complete-ish statements — income, balance sheet, CF-like flows — at horizons ~1–20 quarters, on a panel of many firms/BHCs, with preference for methods that can respect accounting identities / coherence.
Primary application panel: FR Y-9C bank holding company consolidated reports
via FFIEC NIC BHCF bulk zips (preferred data path).
Smoke path: Yahoo Finance quarterly revenue/expense for BHC tickers in a
reporting hierarchy.
Method families included:
- Classical / statistical: VAR/VECM, DFM/factor, state-space/Kalman, Bayesian hierarchical / empirical-Bayes dynamic panels, MIDAS / mixed-frequency, nowcasting panels.
- Structured ML (non-LLM backbone): GBM/RF, elastic-net / sg-LASSO, GPs, quantile / distributional regression, multi-task / multi-output, small multitask nets (not foundation FT).
- Probabilistic / coherent: multivariate densities, quantile stacks, forecast reconciliation (MinT / general linear constraints).
- LLM-as-feature-guide only (prompted sentiment as exogenous) — not FT embeddings.
Excluded: Large embedding / transformer foundation models trained or heavily fine-tuned on statements (Forma-class) — contrast only. Pure price/return alpha ML.
2. Cluster map
Cluster A — Y-9C / BHC Panel Anchors
| Paper | Claim | Fit |
|---|---|---|
| Liu, Moon, Schorfheide — Forecasting with Dynamic Panel Data Models, NBER w25102 (2018) | Empirical-Bayes / Tweedie posterior means for heterogeneous dynamic-panel coefficients; forecasts BHC PPNR from Y-9C under actual vs severely adverse macros. | #1 NIM pick. Direct Y-9C PPNR application; hierarchical shrinkage for unbalanced BHC panel; posterior predictive = probabilistic. |
| Mitra et al. — Interest rate uncertainty and the predictability of bank revenues, Journal of Forecasting | Extends Liu-style Bayes panel PPNR with rate-uncertainty; NII vs noninterest split. | Stress-test relevant exogenous for Y-9C revenue blocks. |
| Medeiros-style firm VAR/VECM statement systems | Cointegrated statement systems; impulse responses. | Single-firm / small-N intuition for identities; not primary for large BHC panel. |
Cluster B — Full / Multi-item Statement ML
| Paper | Claim | Fit |
|---|---|---|
| Geertsema, Lu, Ma — Projecting Financial Statements with AI (SSRN 5039433) | Multi-target chained GBM for interdependent IS+BS items; articulation via chain order. | #2 NIM pick. Closest non-transformer full-statement baseline. |
| Alberg & Lipton (2017/18); Chauhan, Alberg, Lipton (ICML 2020) | Joint multi-fundamental forecasts with uncertainty via multitask FFNN/LSTM. | #6 NIM pick. Classical multitask net benchmark (not foundation FT). |
| Divo et al. — Forecasting Company Fundamentals, TMLR 2025 | Benchmark of ~24 models on selected fundamentals; open code (GitHub). | Protocol / baseline zoo for Yahoo smoke before Y-9C scale. |
| Chen, Cho, Dou, Lev (JAR 2022) — RF/GBM on detailed financials | Earnings-change direction from rich XBRL/financial features. | Item-level tree baselines on Y-9C MDRMs. |
Cluster C — Cross-section Earnings Classics
| Paper | Claim | Fit |
|---|---|---|
| Hou, van Dijk, Zhang (JAE 2012) | Cross-sectional earnings model for ICC; coverage vs analysts. | Foundational earnings forecast anchor; map to BHCK income lines. |
| Li & Mohanram (RAST 2014) | EP/RI cross-section vs HVZ for ICC. | Parsimonious cross-section competitors. |
| Fairfield, Sweeney, Yohn / Nissim–Penman | Disaggregation & ratio mean-reversion. | Feature design for Y-9C ratios. |
| Fama–French profitability forecasting (JB 2000; JF 2006) | Cross-section profitability predictability. | Mean-reversion baselines. |
Cluster D — Mixed-frequency / Nowcasting
| Paper | Claim | Fit |
|---|---|---|
| Ball & Ghysels (Management Science 2018) — MIDAS earnings | MIDAS earnings; combine with analysts; daily→quarterly. | #5 NIM pick. Yahoo daily → quarterly Y-9C smoke. |
| Babii, Ball, Ghysels, Striaukas — Panel MIDAS sg-LASSO (arXiv 2307.02673) | Sparse-group LASSO MIDAS panel nowcasts P/E & earnings with news+macro. | #4 NIM pick. Structured high-dim mixed-freq for many BHCs. |
Cluster E — Coherence / Reconciliation
| Paper | Claim | Fit |
|---|---|---|
| Wickramasuriya, Athanasopoulos, Hyndman — MinT (JASA 2019) | Trace-minimization reconciliation for hierarchical/grouped series. | Post-process any base Y-9C forecasts to satisfy additivity. FoReco / hts R packages. |
| Girolimetto & Di Fonzo (Stat Methods Appl 2024; arXiv 2305.05330) | Point & probabilistic reconciliation for general linear constraints. | Accounting identities as signed linear constraints — native for Y-9C hierarchies. |
| Forma (Johnson et al., arXiv 2608.11327) — variance-weighted identity projection | Reconciliation cost on statement forecasts (contrast). | Proof that variance-weighted projection matters; do not adopt Forma FT path. |
Cluster F — Distributional / Quantile
| Paper | Claim | Fit |
|---|---|---|
| Konstantinidi & Pope (CAR 2016) | Quantile regressions → earnings distribution / risk metrics. | Tail risk on Y-9C PPNR / NI; feed CRPS evaluation. |
Cluster H — Parked / Deprioritized
| Item | Why parked |
|---|---|
| Forma tuple-set transformer + ProForma-20Q | Specialist FT on Compustat statements; gated weights; not Y-9C MDRM schema; non-embedding path requested first. |
| Chronos / TimesFM / frontier LLM zero-shot statement prompts | Weak change-space R² on Forma's own table; expensive; not identity-native. |
| Large FT embedding models without call-report FT track record | Explicitly out of scope until FT workflow proven. |
3. Recommended stack (Yahoo smoke → Y-9C bulk)
- Liu–Moon–Schorfheide EB dynamic panel (PPNR) — Hierarchical empirical-Bayes shrinkage across BHCs; needs BHCF quarterly PPNR + macros.
- Geertsema–Lu–Ma chained GBM — Multi-target chained boosting across IS+BS; needs full BHCF item set per BHC-quarter.
- MinT + Girolimetto–Di Fonzo reconciliation — Probabilistic post-layer enforcing Y-9C additivity; needs hierarchy metadata from BHCF.
- Babii et al. sg-LASSO MIDAS — Sparse mixed-frequency map from Yahoo daily → quarterly Y-9C; smoke + bulk dual use.
- Ball–Ghysels MIDAS — Classic daily→quarterly baseline for Yahoo smoke before scaling.
- Alberg–Lipton multitask FFNN — Joint multi-line neural benchmark (classical multitask, not foundation FT).
- (Honorable) Hou–van Dijk–Zhang / Li–Mohanram cross-section earnings — fast earnings-only baselines on BHCK income.
- (Honorable) Konstantinidi–Pope quantiles — distributional layer / CRPS targets on PPNR & NI.
Preferred data path:
FFIEC NIC ReturnBHCFZipFiles?zipfilename=BHCF{YYYYMMDD}.zip → y9c/ scripts.
Yahoo under yahoo_bank_quarterly/ is smoke only.
4. Not recommended now
- Forma-class / large statement transformers and embedding FT without a validated call-report FT track record.
- Zero-shot frontier LLMs as primary statement engines (cost, non-determinism, weak change-space skill).
- Equal-weighted identity projection without predictive variances (Forma Table 3 warning).
- Claiming Compustat/Forma OOS metrics transfer to Y-9C.
5. Open questions / next experiments
E1 — Yahoo smoke (Ball–Ghysels + sg-LASSO MIDAS). Forecast quarterly PPNR for ~20 large BHCs from Yahoo daily (yields, VIX, spreads); score vs BHCF actuals with rolling-origin RMSE/CRPS.
E2 — Bulk EB panel (Liu–Moon–Schorfheide). EB dynamic panel on all BHCF filers' PPNR; coverage & small-BHC shrinkage vs pooled OLS.
E3 — Articulation + reconciliation (chained GBM → MinT). Fit chained GBM on IS+BS MDRMs; apply MinT/Girolimetto–Di Fonzo; measure identity-violation reduction pre/post.
Additional: Map ProForma-style identity list → Y-9C MDRM linear constraints. Point-in-time / revision awareness for Y-9C restatements. Scenario conditioning (unemployment, FF rate, spreads) à la Liu stress path — without Forma pinned-future FT.
6. Citation anchors
- Liu, Moon, Schorfheide — NBER w25102 / dynamic panel EB forecasting (Y-9C PPNR app).
- Mitra et al. — J. Forecasting, bank revenue + rate uncertainty.
- Geertsema, Lu, Ma — SSRN 5039433, chained GBM statements.
- Hou, van Dijk, Zhang — J. Accounting & Economics 2012.
- Li & Mohanram — Review of Accounting Studies 2014.
- Ball & Ghysels — Management Science 2018.
- Babii, Ball, Ghysels, Striaukas — arXiv 2307.02673.
- Wickramasuriya, Athanasopoulos, Hyndman — MinT, JASA 2019.
- Girolimetto & Di Fonzo — Statistical Methods & Applications 2024.
- Konstantinidi & Pope — Contemporary Accounting Research 2016.
- Alberg & Lipton; Chauhan et al. ICML 2020.
- Divo et al. — TMLR 2025.
- Chen, Cho, Dou, Lev — JAR 2022.
- Johnson et al. — Forma / ProForma-20Q, arXiv 2608.11327 (contrast only).
7. Operational pointers
- Y-9C download/build:
python scripts/download_y9c.py;python scripts/build_y9c_panel.py - Yahoo bank quarterly:
python scripts/yahoo_bank_quarterly.py - Processed:
y9c/processed/y9c_panel_long.parquet,yahoo_bank_quarterly/processed/yahoo_bank_quarterly_long.parquet