Track Resources
- Teaching Dynamic factor nowcasting meets Sinkhorn loss — why, what was found, and research recipe
- Lit Memo Lit synthesis: DFM forecasting ∩ Sinkhorn-as-loss — cluster map and open questions
Overview
The research question
Two mature literatures sit next to each other:
- Dynamic factor models (DFMs) for large panels and ragged-edge nowcasting — observation equation \(y_t = \Lambda f_t + e_t\), Kalman filter/smoother, EM for parameters (Stock & Watson; Banbura, Giannone & Reichlin).
- Sinkhorn / entropic optimal transport as a differentiable loss between probability measures (Cuturi 2013; Genevay / Feydy-style Sinkhorn divergence).
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?
Key finding
Spelta, Pagnottoni & Pecora (JRSS-A 2025, DOI 10.1093/jrsssa/qnaf036) is the clearest intersection: they propose a Sinkhorn-divergence loss treating two batches from the same dataset as sharing a distribution, for filling missing low-frequency entries, and add that Sinkhorn step as a complement to the EM algorithm of a Dynamic Factor Model, with Monte Carlo and US GDP nowcasting evidence.
Adjacent work (not true ∩)
Most arXiv hits cluster around: deep time-series + Sinkhorn alignment (N-BEATS), Wasserstein tensor factorization, COT-GANs, martingale OT in derivatives pricing, Sinkhorn for precipitation verification — none combine DFM/Kalman nowcasting with Sinkhorn as the training/imputation loss.