Non-stationary Diffusion For Probabilistic Time Series Forecasting

NS-Diff addresses a significant limitation in existing diffusion models for time series: the assumption of constant variance. By utilizing the Location-Scale Noise Model (LSNM), it allows the diffusion process to adapt to heteroscedasticity (time-varying volatility). Quantitatively, the project is in its infancy with 0 stars and 3 forks, appearing to be a fresh release tied to an arXiv paper (2505.04278). From a competitive standpoint, it enters a crowded field of diffusion-based forecasters like TimeGrad and CSDI. While the mathematical approach is a clever refinement, it currently lacks a 'moat' beyond the specific algorithmic implementation. It is highly susceptible to displacement by time-series foundation models (TSFM) like Amazon's Chronos or Google's TimesFM, which are increasingly dominating the niche. These larger models could easily integrate non-stationary noise handling if it proves consistently superior. The defensibility is low (3) because it is a research-centric reference implementation rather than an infrastructure tool; its value lies in the methodology which can be replicated in more robust libraries like GluonTS or Darts. The displacement horizon is set to 1-2 years, as this is the timeframe in which TSFMs are expected to absorb these types of architectural improvements.