Noise or Signal? Deconstructing Contradictions and An Adaptive Remedy for Reversible Normalization in Time Series Forecasting

Source

• Raw Markdown: paper_noise-or-signal-2025.md
• Paper PDF: paper_noise-or-signal-2025.pdf
• Preprint: arXiv 2510.04667v1
• Official senior-author page: Yang Yang, Zhejiang University
• Related official baseline implementation: RevIN

No official code repository, project page, peer-reviewed venue, or author X announcement for this paper was verified during the 2026-08-05 ingest. The RevIN repository above is the original baseline implementation, not an implementation of R${}^2$-IN, R${}^2$-IN+, or A-IN.

Status And Credibility

Fanzhe Fu and Yang Yang submitted the paper to arXiv on 2025-10-06. The canonical record still listed only v1, with no journal reference or peer-reviewed venue, at ingest time. Both authors list Zhejiang University in the paper. Yang Yang’s official page identifies him as a professor of artificial intelligence working on large-scale graph and time-series data, and lists Fanzhe Fu as a PhD student.

This is therefore a recent Alex-curated preprint from a credible time-series research group, but its evidence maturity is limited. There is no released implementation, the main experiments use only DLinear, the seed and variance protocol is not reported, and the paper’s own result tables contain material inconsistencies. Its most durable value is the normalization failure analysis and the warning that a spike cannot be classified as noise or signal from a robust statistic alone.

Core Claim

The paper studies reversible instance normalization for long-horizon forecasting. Standard RevIN normalizes each input window with its own mean and standard deviation and reverses those statistics on the forecast:

$${\mu }^{(i)}=\frac{1}{L}\sum _{t=1}^L{x}_t^{(i)},{\sigma }^{(i)}=\sqrt{\frac{1}{L}\sum _{t=1}^L(x_t^{(i)}-{\mu }^{(i)}{)}^2+ϵ}.$$

The authors argue that normalization choices face four unresolved contradictions:

1. Noise versus signal: a spike can be corrupt noise or the first observation of a new regime.
2. Past versus future: lookback-window statistics need not match forecast-horizon statistics after a structural break.
3. Statistic versus distribution fit: mean and median encode different centers, especially under skew.
4. Robustness versus the Gaussian $k$-factor: scaling MAD by $1.4826$ assumes normality even though MAD is introduced because non-Gaussian contamination is expected.

The paper’s main empirical conclusion is cautionary rather than a successful adaptive-method result: naive median/MAD R${}^2$-IN obtains the best reported average rank, while the proposed dataset-level A-IN rule performs worst.

Methods

Method	Location statistic	Scale statistic	Selection rule
RevIN	Mean	Standard deviation	Fixed
R${}^2$-IN	Median	$1.4826\cdot \mathrm{MAD}$	Fixed
R${}^2$-IN+	Median	$k_{\mathrm{emp}}\cdot \mathrm{MAD}$	Fixed, with per-instance $k_{\mathrm{emp}}$
A-IN	Depends on selected method	Depends on selected method	Dataset-level switch: R${}^2$-IN+ for low CPR, RevIN for high CPR

The diagnostic portrait uses:

$$k_{\mathrm{emp}}=\frac{\mathrm{std}(\mathbf{x})}{\mathrm{MAD}(\mathbf{x})},$$

plus PELT-derived Change Point Risk (CPR), skewness, and kurtosis. The main protocol says that DLinear sees a lookback of 336 and predicts horizons 96, 192, 336, and 720 over eleven datasets, with MSE, MAE, and average rank.

Main Evidence

RevIN failure on Electricity

For Electricity at horizon 96, the Phase 2 table reports:

Configuration	MSE	MAE
DLinear without instance normalization	13.2793	0.3959
DLinear + RevIN	104.0162	0.7849
DLinear + R${}^2$-IN	12.9196	0.3914

The RevIN MSE increase relative to plain DLinear is $683.30\%$, consistent with the paper’s rounded $683\%$ claim. The large MSE/MAE gap is also a warning that a small number of extreme errors may dominate the squared-error aggregate; the paper does not provide window-level quantiles or affected-channel counts.

Average-rank result

The released figure reports average rank across the stated tasks:

Configuration	Average rank $\downarrow$
DLinear + R${}^2$-IN	2.08
DLinear without normalization	2.67
DLinear + R${}^2$-IN+	2.80
DLinear + RevIN	2.83
DLinear + A-IN	4.17

This supports a narrow conclusion: under this DLinear recipe and the paper’s aggregation, the simple robust baseline ranks first and the proposed adaptive heuristic ranks last. It does not establish R${}^2$-IN as a universal normalization method.

Independent Claim Audit

R${}^2$-IN+ algebraically restores standard-deviation scaling

The paper defines

$$k_{\mathrm{emp}}=\frac{\mathrm{std}(\mathbf{x})}{\mathrm{MAD}(\mathbf{x})}.$$

Therefore, whenever MAD is nonzero,

$$k_{\mathrm{emp}}\mathrm{MAD}(\mathbf{x})=\mathrm{std}(\mathbf{x}).$$

As written, R${}^2$-IN+ is not robust in its scale statistic: it combines median centering with ordinary standard-deviation scaling. This makes the phrase “corrected robust normalization” too broad and helps explain why the method need not inherit the outlier resistance of fixed-factor median/MAD normalization.

Large $k_{\mathrm{emp}}$ is not by itself proof of outliers

The paper interprets enormous $k_{\mathrm{emp}}$ values as evidence of extreme outliers. But the ratio can also explode when MAD is zero or near zero, which is common for sparse, quantized, or locally constant channels. The paper reports Electricity $k_{\mathrm{emp}}=5.22\times 10^8$ together with NaN skewness and kurtosis, but does not specify zero-MAD handling. A reproducible diagnostic must separate large numerator, small denominator, missing-statistic, and true contaminated-window cases.

A-IN is a dataset-level heuristic, not a learned adaptive normalizer

A-IN chooses one normalization strategy for an entire dataset before training. Its failure demonstrates that this CPR-to-method rule is poor; it does not show that every learnable, per-window, regime-aware, or context-conditioned normalizer will fail. The later reversed-rule ablation hardcodes the better method on one randomly sampled Electricity subset and therefore does not validate a general adaptive selection rule.

Reported tables drift across phases

Repeated configurations have different values without a stated seed or rerun protocol. Examples at Electricity horizon 96 include:

• R${}^2$-IN MSE: 12.9196 in Phase 2, 13.0741 in Phase 3, and 12.9196 in the final table;
• R${}^2$-IN+ MSE: 12.9334 in Phase 3 and 12.9662 in the final table;
• A-IN MSE: 127.6675 in Phase 3 and 194.9510 in the final table.

The ablation uses an unspecified random test subset and reports yet another scale of values. These are not small formatting-only differences, especially for A-IN, and prevent a clean reconstruction of the rank calculation.

Limitations And Gotchas

• The paper says it evaluates eleven datasets, but detailed prediction tables show only ETTh2, Exchange, and Electricity. The full per-dataset/per-horizon ledger behind the average-rank figure is absent.
• All main results use DLinear; there is no evidence that the ordering transfers to Transformers, state-space models, probabilistic forecasters, JEPA encoders, or action-conditioned world models.
• No random-seed count, uncertainty interval, optimizer recipe, split manifest, or official code release is reported.
• MSE and average rank can overemphasize different aspects of performance. The paper does not report per-window tail metrics, median performance, calibration, or statistical significance.
• CPR construction depends on PELT settings, but the operational penalty and sensitivity protocol are not sufficiently specified for independent reconstruction.
• The claimed novelty of being first to systematize the four contradictions is not established by a broad normalization-literature audit.
• R${}^2$-IN’s best average rank is a model-and-protocol-specific result, not evidence that extreme observations should generally be discarded.
• A-IN’s failure is a negative result about one counter-intuitive rule, not a negative result about adaptation as a class.

Relevance To Telemetry And FM-JEPA

The paper does not study telemetry world models, control, JEPA, flow matching, or probabilistic forecasting. Its transfer value is an upstream information-preservation warning.

Operational spikes can be:

• corrupt or missing-data artifacts;
• workload bursts;
• action transients after restart, rollout, scaling, or traffic shift;
• first evidence of queue saturation or cascading failure;
• rare state transitions that define a separate future mode.

A robust normalizer can protect a point forecaster from corruption while simultaneously erasing the branch cue that a latent-state or stochastic world model needs. Conversely, mean/std RevIN can let one corrupt window dominate the representation and inverse transform. The safe design is not one universal statistic; it is an explicit normalization contract:

raw telemetry + units + window statistics + missingness/outlier flags
  -> causal model input transform
  -> latent state / future distribution
  -> raw-unit decoder and operational constraints

For Probabilistic JEPA Predictor Using Flow Matching, normalization SHOULD be an ablation axis rather than an invisible preprocessing constant. A useful comparison includes no instance normalization, causal RevIN, median/MAD normalization, and a stats-as-context design. Rare-event and regime-identity probes must be evaluated before the flow predictor receives credit: a stochastic head cannot recover a mode that the encoder’s input transform has already deleted.

Foundation TSFM Relevance

Agenda slot	Verdict	Evidence	Missing pieces
Data diversity, curriculum, and long tail	warning	Distinguishes corrupt outliers from spikes that may announce a regime change and shows one RevIN MSE failure under an extreme Electricity window.	Needs labeled corruption-versus-event slices, high-dimensional channels, multiple architectures, and rare-state retention metrics.
Representation quality: semantic state versus numeric detail	warning	Shows that window statistics can materially alter the numeric signal before the model sees it.	Does not probe latent representations, state accessibility, or downstream decision relevance.
Benchmark hygiene	partially closes	Exposes normalization strategy as a first-order benchmark variable and reports both a strong simple baseline and a failed proposed heuristic.	Missing code, seeds, complete result tables, zero-MAD handling, and stable repeated values.
Control and counterfactuals	insufficient evidence	None; all experiments are passive point forecasting.	Needs actions, control inputs, intervention-conditioned futures, probabilities, and closed-loop utility.

Links Into The Wiki

• Foundation Time-Series Model Research Agenda
• Time-Series Benchmark Hygiene
• Observability Time Series
• High-Dimensional Time Series Forecasting
• Probabilistic JEPA Predictor Using Flow Matching
• MOMENT
• Toto
• Contradictions And Open Tensions

Open Questions

• How often do large $k_{\mathrm{emp}}$ values in forecasting benchmarks come from true outliers versus zero or near-zero MAD channels?
• Does median/MAD normalization improve robustness while degrading event, change-point, or rare-regime recall?
• Can a model carry normalization statistics, units, and outlier hypotheses as context instead of irreversibly removing them?
• Do the reported rankings survive multiple seeds, complete result release, and stronger backbones?
• For distributional forecasts, how should location and scale transforms preserve calibrated uncertainty and joint cross-channel structure after inversion?