Paper: arXiv 2609.23378

Authors: Zijiang Yang

Abstract

Recursive differenced forecasting, the standard remedy for non-stationarity, predicts one-step changes and integrates them by cumulative summation. We show that this reconstruction is a discrete integrator with a pole on the unit circle, so the biased increment errors of a learned nonlinear model are summed without bound and the rollout diverges: at 336 steps its normalised MAE reaches 1.6-3.8 for every neural architecture tested, against 0.80 for a stable linear recursion. We then introduce leaky-integrator reconstruction, a training-free fix that moves the pole inside the unit circle with H(z) = 1/(1 - gamma z^-1), gamma < 1, bounding the accumulation of the model’s own increment errors. Applied post hoc with a single fixed gamma=0.9 (no retraining, a two-line change to any deployed one-step or foundation-model forecaster), it beats the traditional recursive integrator at every horizon, with the mean gain over seven diverging architectures and twenty datasets growing from ~3% at H=24 to 23% at H=96, 37% at H=192 and 51% at H=336 (43-75% across those architectures; 78% with an oracle pole), bringing all of them to 0.87-0.97. Based on these extensive empirical experiments, adding a leaky integrator thus improves recursive differenced time-series forecasting.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper presents a novel and effective solution to a known problem in time-series forecasting, backed by extensive empirical validation across multiple architectures and datasets. The mathematical foundation is sound, and the practical implications are significant for deployed models.

Research Flowchart

  flowchart TD
    A["Research Goal: Tame Error Accumulation in Recursive Differenced TS Forecasting"] --> B{"Standard Approach: Recursive Differenced Forecasting"};
    B -- Problem Identified --> C["Problem: Traditional Integrator (pole on unit circle) causes unbounded error accumulation"];
    C --> D["Proposed Solution: Leaky-Integrator Reconstruction (pole inside unit circle)"];
    D -- Applied to --> E["Data/Models: 7 Neural Architectures, 20 Datasets, Post-hoc application with gamma=0.9"];
    E -- Evaluation --> F["Computational Process: Compare MAE of Traditional vs. Leaky Integrator across horizons"];
    F --> G["Key Finding: Leaky-Integrator consistently beats Traditional Integrator across horizons, architectures, and datasets, reducing MAE significantly"];

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a leaky‑integrator reconstruction, a training‑free fix that moves the pole inside the unit circle to bound error accumulation in recursive differenced forecasting.

Source sentence
We then introduce leaky-integrator reconstruction, a training-free fix that moves the pole inside the unit circle with H(z) = 1/(1 - gamma z^-1), gamma < 1, bounding the accumulation of the model's own increment errors.

Assumes 5

  • Recursive differenced forecasting reconstructs levels with a unit‑pole integrator (1/(1‑z⁻¹)). model
    Source sentence
    the reconstruction is a discrete integrator 1/(1 −z−1) with a pole at z = 1
  • The learned one‑step increment predictor is biased and its errors are correlated. statistical
    Source sentence
    a learned model’s biased, correlated increment errors are integrated without bound
  • The evaluation uses normalised MAE (absolute error divided by training standard deviation). statistical
    Source sentence
    We report normalised MAE: the absolute level error in units of the training standard deviation, |ˆyt+k − yt+k|/σtrain
  • Stable linear‑AR models are already stable and should remain unchanged (γ=1). model
    Source sentence
    Linear‑AR is stable and, correctly, left at γ=1
Show 1 more
  • The super‑√k diagnostic can identify when the fix should be applied. other
    Source sentence
    gated by the super‑√k diagnostic on a validation split so that stable or joint models are left at γ=1

Data

No dataset named in the text.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Recursive differenced forecasting diverges for all nonlinear architectures, reaching up to 3.8 normalised MAE at horizon 336. performance
    Source sentence
    on twenty datasets every recursively rolled‑out non‑linear architecture diverges (up to 3.8 normalised MAE at H=336)
  • Leaky‑integrator reconstruction with fixed γ=0.9 reduces 336‑step normalised MAE by 43–75% across the seven diverging architectures, bringing errors to 0.87–0.97. performance
    Source sentence
    a single fixed pole γ=0.9 pulls all of them back to ≈0.87–0.97 nMAE, a 43–75% error reduction
  • The gain from the leaky integrator grows with horizon, e.g., ~3% at H=24, 23% at H=96, 37% at H=192, and 51% at H=336. performance
    Source sentence
    the mean gain over seven diverging architectures and twenty datasets growing from ~3% at H=24 to 23% at H=96, 37% at H=192 and 51% at H=336
  • Linear‑AR models remain unchanged (γ=1) and retain their baseline normalised MAE of 0.80. performance
    Source sentence
    Linear‑AR is stable and, correctly, left at γ=1
  • The leaky integrator is inert for joint predictors that never integrate increments and can hurt stable recursions (linear‑AR +9.3%). negative
    Source sentence
    it is inert for joint predictors, which never integrate increments, and hurts stable recursions (linear‑AR +9.3%)

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 24 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (22)

  1. Time Series Analysis: Forecasting and Control (1970)
  2. Long Short-Term Memory (1997)
  3. Reversible Instance Normalization for Accurate Time-Series Forecasting Against Distribution Shift (2022)
  4. Are Transformers Effective for Time Series Forecasting? (2023)
  5. Informer : Beyond Efficient Transformer for Long Sequence Time-Series Forecasting (2021)
  6. A Time Series Is Worth 64 Words: Long-Term Forecasting with Transformers (2023)
  7. iTransformer : Inverted Transformers Are Effective for Time Series Forecasting (2024)
  8. Autoformer : Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (2021)
and 14 more
  1. N-BEATS : Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (2020)
  2. DeepAR : Probabilistic Forecasting with Autoregressive Recurrent Networks (2020)
  3. Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks (2015)
  4. Co-Integration and Error Correction: Representation, Estimation, and Testing (1987)
  5. Non-Stationary Transformers: Exploring the Stationarity in Time Series Forecasting (2022)
  6. A Comparison of Direct and Iterated Multistep AR Methods for Forecasting Macroeconomic Time Series (2006)
  7. A Bias and Variance Analysis for Multistep-Ahead Time Series Forecasting (2016)
  8. Exponential Smoothing: The State of the Art—Part II (2006)
  9. Forecasting Trends in Time Series (1985)
  10. Sequence to Sequence Learning with Neural Networks (2014)
  11. Professor Forcing: A New Algorithm for Training Recurrent Networks (2016)
  12. FEDformer : Frequency Enhanced Decomposed Transformer for Long-Term Series Forecasting (2022)
  13. N-HiTS : Neural Hierarchical Interpolation for Time Series Forecasting (2023)
  14. A Decoder-Only Foundation Model for Time-Series Forecasting (2024)

0 of 22 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.