Research Tracks
Non-embedding methods for joint multi-item forecasting of bank holding company income statements and balance sheets — FR Y-9C panel with Yahoo smoke. Covers empirical-Bayes panels, chained GBM, MIDAS, MinT reconciliation.
Research data path for bank statement forecasting: FFIEC NIC FR Y-9C bulk downloads and Yahoo Finance quarterly smoke. Panel assembly scripts, alignment notes, and known quirks.
Literature exploration at the intersection of dynamic factor models (ragged-edge nowcasting) and Sinkhorn / entropic optimal transport as an observation or imputation loss. Thin but real intersection; anchored by Spelta et al. (JRSS-A 2025).
Pre-registered forecasts of FR Y-9C trading revenue: per-bank baselines with W₂-barycenter combination and Wasserstein autoregression of the cross-section. One authorized out-of-sample run: primary FAIL on both hypotheses; run incomplete (stopped in a sensitivity).
Burn-in median next-month W₂² 0.0064 < half-sample noise floor 0.0269: month-to-month change in the cross-company shape of complaints is below sampling noise, so the pre-registered rule killed the object. No out-of-sample run.
Next-year change in the within-state spread of branch deposits was below the half-sample noise floor in all 9 burn-in years (median W₂² 0.0078 vs 0.0196); planning power 0.07 with 22 test years. No out-of-sample run.
Programs
Which forecasts of commercial paper issuance and maturity mix are best, whether any helps spot funding stress early, and whether optimal transport earns a place against strong baselines. Fed CP maturity mix × OT: passed burn-in, pinned, awaiting Jared's approval. Nothing held out has been scored.
Learning guides
- Optimal transport on kernels: learning guide
- CFPB complaints: what the forecast was going to test (learning page)
- FDIC branch deposits: what the forecast was going to test (learning page)
- Fed CP maturity mix: what the burn-in gate found (learning page)
- Optimal transport for operations research, risk & planning: classical methods
About this Lab
Scope
This lab collects independent research on quantitative methods for bank fundamentals and macroeconomic nowcasting. The primary application is the FR Y-9C BHC consolidated reporting panel (FFIEC NIC bulk downloads), with Yahoo Finance quarterly data as a cheap smoke-test path.
What is not here
- Production Portfolios or ETF pipeline work (see
usa-etf-feature-pipeline). - Raw panel CSVs, HF checkpoints, or secrets.
- Gated data or paywalled assets.
- Forma-release or proforma-20q production outputs.
How to navigate
Each track has an overview page, one or more teaching pages (method intuition, worked sketches), and lit memos (structured literature synthesis). Use the top nav to jump between tracks.
Adding a new track
See the README for the one-page checklist.