Stationarity Transformations for Robust Time Series Forecasting

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 296 papers112 conceptsfrozen 2026-10-04

Most statistical and machine learning models share a fundamental assumption: the "rules" of the system being studied do not change over time. In the context of time series, this property is known as stationarity. A stationary series is one whose statistical properties—such as mean, variance, and autocorrelation—are constant regardless of when you observe them.

However, real-world data is rarely so cooperative. Economic indicators trend upward, electricity prices fluctuate with the seasons, and server loads spike during business hours. To model these "moving targets," researchers across diverse fields rely on a suite of transformations designed to strip away non-stationary noise and reveal the underlying signal.

The Intuition of Differencing

The most common tool for achieving stationarity is differencing. If a time series has a trend—meaning its mean changes over time—we can often stabilize it by looking at the change between consecutive observations rather than the absolute values.

Mathematically, first-order differencing is defined as $y'_t = y_t - y_{t-1}$. As Zheng Yin and Lei Dai demonstrate in their work on coal price forecasting, this simple subtraction eliminates underlying trends that would otherwise bias a model. By focusing on the rate of change, the model is forced to learn the cyclical patterns and fluctuations that actually drive the system.

This approach is ubiquitous. In social sciences, Linghe Kong Yiwen Fang and Yan Zhu use differencing to handle the upward trajectory of educational resources. In infrastructure, Chaojun Fu and colleagues apply it to virtual machine load prediction to prevent unnecessary migrations caused by temporary spikes.

Stabilizing Variance with Power Transformations

While differencing stabilizes the mean, it does not always address heteroscedasticity—a phenomenon where the magnitude of fluctuations changes over time (e.g., a stock that becomes more volatile as its price increases).

To handle this, researchers use power transformations like the Box-Cox method. André Bauer and the team behind Telescope use Box-Cox to ensure that the variance remains consistent across the series. Similarly, Patrick Emami applies these transformations in BuildingsBench to improve model convergence when dealing with highly volatile energy consumption patterns.

For data exhibiting exponential growth, such as China’s service industry, Jing Fan and Rui Shan show that logarithmic transformations can linearize the growth, making it manageable for standard multivariate models like ARIMAX.

Beyond Classical Statistics

Stationarity is not just a requirement for classical models like ARIMA; it is increasingly recognized as a critical preprocessing step for deep learning.

In the Memory in Memory (MiM) architecture, Yunbo Wang and Mingsheng Long introduce a differencing operation directly into the recurrent hidden states. By subtracting the previous state from the current one, the network stationarizes the internal signals, allowing it to capture "the change in the change"—essential for forecasting complex spatiotemporal dynamics like weather patterns.

Similarly, Changhua Pei and Quan Zhou utilize first-order differencing in KAN-AD, a framework using Kolmogorov–Arnold Networks. They find that removing the constant mean and drifting trends allows the model to focus exclusively on frequency-based, cyclical patterns, leading to a 15% boost in anomaly detection accuracy.

Automating the Pipeline

Because identifying the correct transformation requires rigorous testing (such as unit root tests mentioned by David J. Albers), recent work has focused on automation. The Merlion library, developed by Aadyot Bhatnagar and colleagues, includes automated routines to detect seasonality and select optimal differencing orders.

Perhaps the most advanced application of this principle is found in ANT (Adaptive Noise Schedule for Time Series). Kibok Lee and Seunghan Lee propose that the optimal way to train a diffusion model is to reduce the data's non-stationarity at a constant, linear rate. This ensures that every step of the model's "un-noising" process contributes equally to reconstructing the structured patterns from stationary white noise.

Papers in this version

Showing 60 of 96 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 107 papers
  2. Version 1 · 2026-10-04 · 96 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_unit_root_testing_ab3826fc,
  title = {unit root testing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unit-root-testing/rdr_ab3826fcc44c4ee2a557967686ad4837},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_stationarity_transformations_for_robust__ab3826fc_v1,
  title = {Stationarity Transformations for Robust Time Series Forecasting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stationarity-transformations-for-robust-time-series-forecasting/rdr_ab3826fcc44c4ee2a557967686ad4837/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}