Teaching: DFM × Sinkhorn Nowcasting

1. Why this paper cluster

Two mature literatures sit next to each other:

Dynamic factor models (DFMs) for large panels and ragged-edge nowcasting (data arrive at different lags).
Sinkhorn / entropic optimal transport as a differentiable loss between probability measures (batch distributions, residual clouds, feature maps).

The natural research question: can a Sinkhorn divergence replace or augment the usual Gaussian / MSE observation criterion inside a factor state-space model, so forecasts respect distributional structure while still exploiting low-rank co-movement?

Honest map: On arXiv-first retrieval, the true intersection is thin. Most hits are adjacent (deep TS + Sinkhorn, Wasserstein tensors, COT-GANs, martingale OT in derivatives). The clearest intersection paper found in this test is journal-side: Spelta, Pagnottoni & Pecora (JRSS-A 2025).

2. What original work found

Classical DFM forecasting

Stock & Watson: few latent factors summarize panels; forecast via estimated factors.
Banbura, Giannone & Reichlin: nowcasting with mixed-frequency / ragged-edge data via state-space DFMs and Kalman filtering/smoothing; EM for parameters. Observation equation \(y_t = \Lambda f_t + e_t\), factor dynamics (VAR), missing entries handled by Kalman with time-varying observation selection; typically Gaussian / MSE loss.

Sinkhorn divergence technology

Cuturi (2013): entropic regularization makes OT differentiable/scalable via Sinkhorn–Knopp matrix scaling.
Genevay / Feydy-style Sinkhorn divergence: debiased Sinkhorn loss \(S_\varepsilon(\mu,\nu) = OT_\varepsilon(\mu,\nu) - \tfrac12 OT_\varepsilon(\mu,\mu) - \tfrac12 OT_\varepsilon(\nu,\nu)\), usable as a training criterion between empirical measures.

True intersection: Spelta et al. (JRSS-A 2025)

Spelta, Pagnottoni & Pecora — Short-term forecasting with optimal transport, JRSS-A 2025 (DOI 10.1093/jrsssa/qnaf036). They (i) propose a Sinkhorn-divergence loss treating two batches from the same dataset as sharing a distribution, for filling missing low-frequency entries, and (ii) add that Sinkhorn step as a complement to the EM algorithm of a Dynamic Factor Model, with Monte Carlo and US GDP nowcasting evidence. This paper was not in the arXiv priority brief (paywalled journal; no arXiv abs in cached hits) — it was found via Jev Ultrafast on the landing page.

3. Adjacent clusters (arXiv hits — what each is not)

Cluster	Why adjacent, not true ∩
N-BEATS + Sinkhorn alignment — Lee et al., 2305.15196	Stack-wise Sinkhorn on N-BEATS features for domain generalization — no DFM / ragged-edge nowcasting.
SWIFT Wasserstein factorization — Afshar et al., 2010.04081	Tensor CP under Wasserstein loss — static factorization, not dynamic factors for forecasting.
R-WAE — Han et al., 2101.07496	Recurrent Wasserstein AE for sequential disentanglement — generative seq model, not DFM nowcasting.
COT-GAN — Xu et al., 2006.08571	Causal OT + Sinkhorn for sequential GANs — adversarial generation, not factor state-space.
Spatial Sinkhorn forecast verification — Francis et al., 2412.16063	Sinkhorn as verification metric for precipitation fields — eval, not DFM training.
Nonlinear filter Sinkhorn stability — Mandal et al., 2208.10810	Sinkhorn/Wasserstein to probe filter stability (PF/EnKF) — diagnostic, not panel DFM loss.
q-fin martingale Sinkhorn	Sinkhorn for martingale OT / IV surfaces / SV calibration — finance math, not DFM nowcasting.
Chaos OT regularization — Melo et al., 2604.21097	Sinkhorn/WGAN OT to match attractor statistics — emulator fidelity, not macro panel factors.

4. Research recipe: DFM + Sinkhorn observation loss

Goal: Improve ragged-edge multi-series short-horizon forecasts by replacing or augmenting Gaussian observation likelihood with a Sinkhorn divergence between predicted and observed panel distributions, following the Spelta-style "distribution-preserving imputation" idea.

Assumptions

Approximate low-rank factor structure: \(y_t = \Lambda f_t + e_t\), \(f_t\) VAR / state-space.
Missingness is ragged-edge / publication lag, treated as missing-at-random conditional on the information set at forecast origin \(t_0\).
Sinkhorn acts on empirical measures over cross-section and/or residual batches.
Entropic regularization \(\varepsilon\) chosen for stability; debiasing (Sinkhorn divergence) preferred.

Protocol (leakage-safe)

Walk-forward: expanding or rolling origin; use only data available at each origin.
No same-period leakage: never train or evaluate with contemporaneous realizations that arrive after the decision time.
Imputation vs prediction: if Sinkhorn fills missing entries inside EM (Spelta-like), freeze filled values for out-of-sample scoring; score point (RMSE/MAE) and distributional (Sinkhorn / CRPS) metrics separately.

Candidate nulls / baselines

Classical Gaussian DFM + Kalman (MSE) — Banbura/Giannone/Reichlin-style.
Same DFM with Student-t / Huber observation loss (robustness without OT).
Sinkhorn-only nonparametric imputer without factors (Spelta stand-alone arm).
Factor model + MSE, then Sinkhorn only as evaluation metric (not train).
Feature-aligned N-BEATS + Sinkhorn (2305.15196) on the same panel.

Success criteria

Statistically better short-horizon scores vs baseline (1) on held-out origins, with Diebold–Mariano.
Gains concentrated where raggedness is severe (ablate missingness rate).
Ablation: remove Sinkhorn step → performance collapses toward MSE-DFM.

5. Top citations

Stock & Watson — dynamic factor / diffusion-index forecasting (classical DFM).
Banbura, Giannone & Reichlin — DFM nowcasting with ragged edge / Kalman.
Cuturi (2013) + Genevay/Feydy Sinkhorn divergence — OT loss technology.
Spelta, Pagnottoni & Pecora (JRSS-A 2025) — DFM-EM + Sinkhorn loss nowcasting. DOI 10.1093/jrsssa/qnaf036.
Lee et al. 2305.15196 — strongest adjacent forecasting+Sinkhorn deep baseline.

6. Open questions

Is Sinkhorn best as an EM augmentation (Spelta), a direct observation loss in gradient-based state-space learning, or only an evaluation metric?
How to define the ground metric on heterogeneous financial panels (scales, signs, mixed frequencies)?
Computational scaling: Sinkhorn per EM iteration vs large \(N\) panels — when does it dominate Kalman cost?
Does distributional matching help tail events / crises, or only average nowcast RMSE?
Identification: can Sinkhorn over-smooth factor dynamics and hurt multi-step forecasts?
Replication: open re-implementation of Spelta DFM+Sinkhorn on a generic public panel (e.g. FRED-MD style) under strict vintage rules.

7. Tooling notes

OpenAlex / feynman rank and raw arXiv API hit 429 earlier in the workflow; abs HTML cache + arxiv_hits.json + priority_brief.md were the reliable literature substrate.
Jev Ultrafast successfully opened the Spelta JRSS-A landing page — critical because the true ∩ paper was journal-first, not in the arXiv brief.
Feynman (GLM via NVIDIA NIM) was used for one-shot synthesis from the brief head.