Forecasting goes continuous: ODE attention and phase-aware residuals for long horizons

Attention inside an ODE solver

In a preprint, the CTIA-ODE authors embed inter-variable attention inside an ODE solver for long-term multivariate forecasting. The design enables dynamic querying of static exogenous context along the ODE integration path without explicit temporal alignment — useful for heterogeneous, asynchronous data streams that do not arrive on a tidy grid. A Linear Residual Augmentation module provides a macroscopic trend anchor to stabilize the continuous latent dynamics, and an Outlier-Robust Hybrid Loss mitigates the influence of heavy-tailed anomalies during optimization. m:…6434

The authors report state-of-the-art forecasting performance across eight real-world datasets. For developers working with irregular, multi-source streams, the selling point is the missing alignment step: no rigid temporal grid required before the model can reason across variables. The design's two stabilizers — the macroscopic trend anchor and the outlier-robust loss — both target the failure mode that usually kills continuous-time models in production: dynamics that drift and data with heavy tails. Neither is glamorous; both are the kind of engineering that decides whether an ODE-based forecaster survives contact with real data. m:…6434

Residuals are signal, not noise

A second preprint, Paloma, reconsiders the part of forecasting most pipelines throw away: the residuals. The authors show that residuals exhibit structured, phase-dependent deviations rather than unstructured noise, and build a technique that modulates them with phase rotation and amplitude scaling. The result is a general-purpose plugin for existing forecasting architectures — not a new model but an upgrade path — and the authors report improved forecasting performance across twelve datasets. The conceptual shift is the story: error is treated as structured signal with phase, and correcting it becomes a first-class modeling step rather than a post-hoc adjustment. m:…8531

Diffusion meets the order book

The third preprint takes the physics toolkit to market microstructure. LOBGrad applies a conditional diffusion model to probabilistic limit-order-book forecasting and reports improvements over encoder-free diffusion ablations across all reported queue metrics. It also reproduces key stylized facts of LOB microstructure — the empirical regularities any credible market model must match — and provides calibrated predictive intervals aimed at downstream execution tasks. For developers building trading or risk tooling, the calibrated intervals may matter more than the point forecasts: execution decisions price in uncertainty, and the paper targets that directly. The authors also report a scaling result on the theory side: under contractive dynamics, the dependence on the conditioning-window length can be reduced to a dependence on the latent-state dimension. In practical terms, that is a claim that longer history windows stop costing you linearly once the dynamics are well-behaved — relevant to anyone sizing the compute budget for a production LOB model. m:…0689

A shared direction

Three papers, one direction: forecasting is moving from discrete sequence models toward continuous, probabilistic structure — ODEs for time, phase for residuals, diffusion for distributions. The shared bet is that long horizons are a structure problem, not a scale problem. All three are preprints, so independent replication is still pending, but for applied developers the plugin angle of Paloma and the calibrated-interval angle of LOBGrad are the immediately testable ones. The ODE-attention machinery is the longer bet — and the one most likely to leak into adjacent domains next. m:…6434 m:…8531 m:…0689