🤖 machine learning

The Spectrum Is Not Enough: When Context Helps Time-Series Forecasting

This paper argues that spectral predictability indices fail to determine the value of adding context (such as retrieval or foundation models) because they ignore phase-dependent, beyond-second-order structures, and instead proposes a "coverage deficit" diagnostic to accurately assess when such contextual enhancements will improve time-series forecasting.

Original authors: Mert Onur Cakiroglu, Mehmet Dalkilic, Hasan Kurban

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mert Onur Cakiroglu, Mehmet Dalkilic, Hasan Kurban

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather. You have a magic crystal ball that tells you exactly how "predictable" a storm is. If the crystal ball says "highly predictable," you might think, "Great! I should bring my biggest, most expensive super-computer to forecast the next week." If it says "low predictability," you might decide to just guess with a simple rule of thumb.

This paper, titled "The Spectrum Is Not Enough," comes along and says: "Stop! Your crystal ball is lying to you about whether you need the super-computer."

Here is the twist: The crystal ball only looks at the rhythm of the weather (how often it rains, how strong the wind blows on average). It ignores the timing of the rhythm. The authors prove that two weather patterns can have the exact same rhythm but completely different futures, and the crystal ball cannot tell them apart.

The Magic Trick: The "Phase-Randomized" Twin

To prove their point, the authors perform a magic trick. They take a real time-series (like electricity usage data) and create a "surrogate twin."

The Original: A real data stream with a specific rhythm and a specific order of events.
The Twin: They scramble the order of the events (the "phase") but keep the rhythm exactly the same.

It's like taking a song, keeping the exact same beat and volume for every note, but shuffling the notes so they play in a random order. To the "rhythm-crystal ball" (which the paper calls spectral indices), the original song and the shuffled song are identical twins. They get the exact same score.

But here is the kicker:

On the original song, adding "context" (looking at more history or using a fancy AI) helps a lot.
On the shuffled twin, adding that same context hurts or does nothing at all.

In one specific dataset called ECL (electricity usage), the authors found that using a "retrieval" tool (a plug-in that fetches past patterns) improved predictions by +33% on the real data. But on the shuffled twin? It made predictions 35% worse.

The crystal ball saw no difference between the two, yet the result was a massive swing from "great help" to "terrible harm." This proves that predictability scores based on rhythm alone cannot tell you if adding more context will help.

The "Impossible" Rule

The authors state a hard rule: No index built only from the power spectrum (the rhythm) can ever predict the value of "beyond-spectrum" context.

Why? Because the "value" of context often comes from non-linear patterns—complex, repeating shapes in the data that aren't just simple waves. When you shuffle the order (phase randomization), you destroy these complex shapes and turn the data into something that looks like random noise (Gaussian). The fancy AI models and retrieval tools rely on these complex shapes. Once the shapes are gone, the tools lose their power.

The paper shows that:

Spectral Indices (The Crystal Ball): Stay frozen. They give the same score for the original and the twin.
Context Value (The Real Benefit): Collapses. It goes from positive to negative when the complex shapes are scrambled.

The New Tool: The "Coverage Deficit"

Since the old crystal ball is useless for this specific job, the authors built a new diagnostic tool called the Coverage Deficit.

Think of it like checking if you have a magnifying glass and if you are looking at the right spot.

The Structure Term: Does the data have those complex, non-linear shapes that the fancy tools need? (The paper measures this by seeing if a "nearest-neighbor" guess is better than a simple straight-line guess).
The Coverage Term: Is your current "window" of time too short to see the whole shape? If you are only looking at half a cycle of a pattern, you need context to see the rest.

The authors tested this new tool on seven different benchmarks (including traffic, weather, and electricity data). They found that while the old crystal ball (spectral indices) failed to predict whether context would help, the Coverage Deficit successfully guessed the answer.

When the Coverage Deficit was high, adding context (like a retrieval plug-in or a foundation model) actually helped.
When it was low, adding context did nothing or made things worse.

What About the "Big AI" Models?

You might have heard that huge pre-trained AI models (Foundation Models) are amazing at forecasting. The paper doesn't say they are bad; it just explains why they work.

The authors found that these big models get most of their success from simple, second-order patterns (the rhythm) that the crystal ball can see. That part of their success survives the magic trick. However, the tiny bit of extra success they get from "beyond-linear" tricks (the complex shapes) disappears when the data is scrambled.

In short: The big models are mostly just very good at reading the rhythm. The part that makes them "smart" in a non-linear way is fragile and depends on the specific order of events, not just the rhythm.

The Bottom Line

The paper concludes with a clear message for anyone trying to build a forecasting system:

Don't just ask, "How predictable is this series?"
Ask, "Does my current window miss a complex shape that context could reveal?"

If you only look at the spectrum (the rhythm), you are blind to the very thing that makes adding context useful. The authors have provided a new, label-free way to check this before you even start training your model, ensuring you don't waste time on fancy tools that won't help your specific problem.

The results are not just a guess; they are backed by rigorous mathematical proofs and controlled experiments where the "twin" data was generated to be mathematically identical in rhythm but different in structure. The numbers are stark: on the ECL dataset, the difference between the original and the twin was a 69.8-point gap in performance, a massive swing that the old metrics completely missed.

Technical Summary: The Spectrum Is Not Enough

Problem Statement

The paper addresses a critical disconnect in time-series forecasting: the conflation of series-level predictability with the configuration-level value of added context. Practitioners increasingly rely on spectral predictability indices (e.g., $\Omega$ , spectral coherence) to decide whether to deploy heavier mechanisms—such as longer lookback windows, retrieval-augmented forecasting, or foundation models. The prevailing assumption is that a high spectral predictability score licenses the use of complex, context-heavy models.

The authors argue this assumption is unsound. While spectral indices measure how predictable a series is in principle (based on its power spectrum), they fail to predict whether additional context (beyond a standard window) will improve forecasts. The value of context depends on the operating point (window length, prediction horizon) and the presence of beyond-second-order structure (nonlinear, phase-dependent dynamics) that the power spectrum cannot capture.

Methodology

Theoretical Framework: Impossibility via Surrogates

The core theoretical contribution is an impossibility result: no index built solely from the power spectrum (or autocovariance) can predict the value of beyond-spectrum context.

Mechanism: The authors utilize phase-randomized surrogate data (and amplitude-adjusted surrogates, IAAFT). These surrogates preserve the power spectrum and marginal distribution of the original series but randomize Fourier phases.
Consequence: By construction, any spectral index yields identical scores for the original series $x$ and its surrogate $\tilde{x}$ . However, the surrogate is asymptotically Gaussian. Since Gaussian processes are fully characterized by their second-order statistics, any "beyond-second-order" value (exploited by retrieval or foundation models) collapses to zero on the surrogate.
Result: If a mechanism's value depends on phase-dependent structure, its performance will drop significantly on the surrogate, while the spectral index remains "frozen." This proves that spectral indices are blind to the specific structure that makes context valuable.

The Coverage-Deficit Diagnostic ( $\Gamma$ )

To address the deployment question, the authors introduce a label-free, configuration-level diagnostic called the Coverage Deficit ( $\Gamma$ ). It consists of two terms:

Structure Term ( $\Gamma_{cov}$ ): Measures the "beyond-spectrum" potential. It combines:
- $\Delta_{nl}$ : A conservative estimate of the gap between linear and analog (nearest-neighbor) prediction, representing the non-linear structure the spectrum misses.
- $u(S)$ : The fraction of a "state-identifying motif" that the current lookback window $S$ fails to observe.
- Formula: $\Gamma_{cov} = \Delta_{nl} \cdot u(S)$ .
Novelty Term ( $\Gamma_{oov}$ ): Measures the probability that a test window tuple is out-of-vocabulary relative to the training distribution (symbolic index). High novelty implies context is stale.

Decision Rule: Context is predicted to help if $\Gamma_{cov}$ is high (exploitable structure exists and is currently missed) and $\Gamma_{oov}$ is low (data is in-distribution).

Experimental Protocol

The authors evaluate their claims across seven standard benchmarks (ETTh1/2, ETTm1/2, Weather, ECL, Traffic) using three context-extending mechanisms:

Longer Linear Window: A purely second-order mechanism (increasing $S$ ).
Analog Retrieval: A window-keyed retrieval plug-in (purely beyond-second-order).
Foundation Model: Zero-shot Chronos-Bolt reading a long context (mixed second-order and beyond-linear).

Experiments include:

E1 (Spectrum-Controlled Gap): Comparing performance on original vs. IAAFT surrogates to verify the collapse of context value while spectral indices remain constant.
E2 (Sign Prediction): Leave-one-dataset-out (LODO) classification of whether context helps, comparing the Coverage Deficit against spectral indices ( $\Omega$ , SCP) and other baselines.
E3–E6: Correlation analysis, mechanism-specific breakdowns, out-of-distribution transfer, and ablation studies.

Key Results

1. The Spectrum-Controlled Gap

On benchmarks where retrieval provides value (ECL, Traffic), the value of context collapses under phase randomization:

ECL: Retrieval value drops from +33.0% on the original series to −35.0% on the surrogate.
Traffic: Value drops from +33.8% to +0.7%.
Spectral Indices: The spectral predictability index $\Omega$ remains unchanged (difference $|\Delta\Omega| \le 0.014$ ), confirming it cannot detect the loss of value.
Mechanism Split:
- Longer Windows: Value survives phase randomization (e.g., ECL: 56.7% $\to$ 55.1%), confirming it relies on second-order structure visible to the spectrum.
- Foundation Models: Their total value splits; the dominant second-order part survives, while the "beyond-linear margin" collapses (e.g., ECL margin: +9.0% $\to$ −8.8%).

2. Diagnostic Performance

The Coverage Deficit ( $\Gamma$ ) successfully predicts the sign of context value where spectral indices fail:

Beyond-Spectrum Mechanisms (Retrieval/Foundation Margin): The structure term ( $\Delta_{nl}$ ) achieves 78.1% balanced sign accuracy for retrieval and 73.2% for foundation model margins. In contrast, $\Omega$ performs at or below chance (39.5% and 52.4%, respectively).
Second-Order Mechanism (Longer Window): Spectral indices (specifically SCP) outperform the structure term (79.2% vs. 46.9%), confirming the diagnostic correctly identifies that longer windows rely on spectrum-visible structure.
Out-of-Distribution: On withheld datasets (Exchange, ILI, M5), the diagnostic correctly identifies that context offers little value (mostly negative or zero), whereas a naive prior based on in-distribution trends would incorrectly suggest context helps.

Significance and Claims

The paper claims three primary contributions:

Distinction of Concepts: It formally separates series-level predictability from configuration-level context value, proving that the former (via spectral indices) cannot determine the latter.
Controlled Comparison: By using surrogate pairs that fix the spectrum and marginal, the authors isolate the "blind spot" of spectral indices without relying on correlation, providing a rigorous test of what these indices can and cannot see.
Deployment Diagnostic: The Coverage Deficit offers a practical, label-free tool for practitioners to decide when to add context. It specifically targets the "beyond-spectrum" structure that retrieval and foundation models exploit, which is invisible to traditional spectral analysis.

The authors emphasize that they do not introduce a new forecaster. Instead, the contribution is a theoretical distinction, a controlled empirical comparison, and a diagnostic to guide deployment decisions. The work suggests that relying on spectral scores to justify complex models is theoretically unsound, and that the value of context is a property of the operating point (window size vs. motif length) and the specific nonlinear structure of the data, not just the series itself.