SHADE: Shrinkage via Halted Alpha-stable Denoising for Heavy-Tailed Probabilistic Time Series Forecasting

What it claims

SHADE, from Jiahao Weng and Jiyu Liu (Renmin University of China, Kuaishou Technology; arXiv v1, 19 October 2026), argues that fully reversing a diffusion forecaster to zero noise produces predictive distributions that are too narrow in the tails when observations are heavy-tailed and low-SNR. The authors attribute this to a high-frequency score that is hard to estimate from finite heavy-tailed data. Over-reversal amplifies that error and gives rare observations vanishing density, hence extreme NLL. SHADE halts at a nonzero depth τ, and the residual stable kernel acts as distribution-level shrinkage. The gains are stated for α ≤ 1.3. On near-light-tailed or high-SNR data the model reverts to Gaussian diffusion or full denoising. Overall forecast error stays "comparable to strong baselines."

How it works

The denoiser follows CSDI's residual backbone, alternating attention over time and variables. The forward process uses LIM's alpha-stable Lévy diffusion (Yoon et al. 2023) and perturbs only the future window, with history held fixed. Other details from the page:

• The tail index α enters every residual block through FiLM (channel-wise scale and shift), not input concatenation.
• α is restricted to (1, 2]. α = 2 is Gaussian.
• The cosine schedule uses K = 50 steps and s = 0.008.
• Training uses an L1 noise-prediction loss, because only the first absolute moment is guaranteed for α < 2.
• Each window is normalized by the median and MAD of its history, not mean and variance.
• On real data, α is estimated on the training segment only.
• Data is split chronologically into training, halting calibration and test.

The theory uses CRPS, whose excess risk equals the squared L2 distance between predictive and true CDFs. This risk splits into a smoothing bias B(λ), which grows with depth, and an estimation error V(λ; s), which shrinks fastest at high frequency. A local power-law approximation gives the halting law τ* ≈ c(α)·SNR^b(α) with b < 0, so lower SNR means an earlier halt. When the local slopes are unstable, the authors fall back to numerical calibration or the context route.

flowchart LR
  H[History window] --> M1[M1 robust scale: MAD of level, diffs, seasonal diffs]
  M1 --> M2[M2 scale-calibrated halt]
  H --> M3[M3 context-amortized halt, no noise labels]
  M2 --> T{Halt depth τ}
  M3 --> T
  R[Stable reverse trajectory, states saved per depth] --> M4[M4 readout at τ]
  T --> M4

The two halting routes are alternatives. M1 with M2 is the auditable route, mapping tail index and SNR scale to depth. M3 learns a distribution over depths from pooled history and ψ(α). Inference runs one reverse trajectory and saves a state at every candidate depth, so halting does not require extra chains.

Evidence

The fetched text stops in the method section. The only figures available are those in the abstract: a 21.7% median CRPS improvement across synthetic noise settings with α ≤ 1.3, and a "significant" NLL drop. The abstract does not name the baselines for these numbers. The paper describes a full tail-index by SNR synthetic grid, five real data domains (only cryptocurrency and FX are named), and ablations on noise-family misspecification and equal-width counterfactuals. No tables or per-dataset numbers appear in the text.

Limits

• Benefits are claimed only for heavy tails (α ≤ 1.3). Elsewhere the method falls back to standard diffusion.
• α ≤ 1 is excluded by construction.
• The halting law is a local approximation that the authors say must not be extrapolated to λ = 0.
• The bias-error split assumes zero-mean endpoint error. The biased case is in Appendix B, which is not in the fetched text.
• No code, weights or licence are mentioned.
• Real-data evidence is described only as "further evidence of improved tail calibration."