Conformal time series decomposition with component-wise exchangeability

Derck W. E. Prinzhorn, Thijmen Nijdam, Putri A. van der Linden, Alexander Timans

TL;DR

Conformal prediction provides rigorous, distribution-free uncertainty bounds, but its foundational assumption—data exchangeability—breaks down in sequential time series. This paper introduces a "divide-and-conquer" approach: decompose a time series into trend, seasonal, and remainder components, assign each component to a matching exchangeability regime, apply tailored conformal methods, and recompose the prediction intervals. The resulting framework substantially tightens prediction intervals on structured time series, though its real-world performance hinges heavily on decomposition quality.

What Problem Does This Paper Address?

Conformal prediction (CP) typically requires exchangeability (i.e., data distributions must be invariant to permutations, a generalization of i.i.d.). Time series violate this due to temporal autocorrelation, non-stationarity, and regime shifts. Existing remedies (like EnbPI or Adaptive Conformal Inference) treat the raw series as an indivisible non-exchangeable sequence, using rolling recalibrations or feedback loops to maintain nominal coverage.

However, time series are rarely uniform in their temporal dependency: a long-term trend drifts non-exchangeably, but seasonal patterns repeat cyclically, and noise remainders often resemble uncorrelated white noise. Forcing the entire time series into a single non-exchangeable framework discards valuable structural inductive biases.

Methods

The authors propose a three-stage pipeline combining classical Time Series Decomposition (TSD) with component-specific CP:

  1. Decomposition: The series is split additively using LOESS (STL) into: $$Y_t = T_t + S_t + R_t$$ representing trend, seasonality, and remainder.

  2. Component-wise CP tailored to Exchangeability Regimes:

  3. Remainder ($R_t$) $\rightarrow$ Globally Exchangeable: Treated as uncorrelated white noise. Calibrated using standard cross-conformal prediction ($CV^+$), which assumes global exchangeability.
  4. Trend ($T_t$) $\rightarrow$ Non-Exchangeable: Exhibits strong serial correlation. Calibrated using time series-specific CP methods with coverage guarantees under distribution shifts (EnbPI or ACI).
  5. Seasonal ($S_t$) $\rightarrow$ Locally Exchangeable: Samples across different periods at the same cyclic phase (e.g., peak vs. trough) are hypothesized to be roughly exchangeable. The authors design weighted CP schemes based on temporal period distance:
  6. BinaryPoint: Hard binary weights restricted exclusively to calibration points at the exact same periodic offset ($i \pmod \tau = \tau_{n+1}$).
  7. BinaryLocal: Extends BinaryPoint to include a small temporal neighborhood ($k$ steps) around the target phase.
  8. ExpLocal: Applies continuous exponential decay weights ($\lambda e^{-\lambda \Delta \tau}$) decaying with distance from the target phase within the cycle.
  9. Recomposition: Prediction interval (PI) boundaries are summed across components: $$\hat{L}(X_{n+1}) = \hat{L}_T + \hat{L}_S + \hat{L}_R, \quad \hat{U}(X_{n+1}) = \hat{U}_T + \hat{U}_S + \hat{U}_R$$ A union bound guarantees a loose nominal coverage lower bound of $1 - 3\alpha$.

Datasets

The authors evaluated the method across four setups:

  • Synthetic Data: Known additive process with linear trend ($0.1t$), a clean sine-wave season ($\tau = 30$), and Gaussian noise ($\mathcal{N}(0, 1)$).
  • San Diego Energy Consumption: 5 years of hourly electricity usage (Watt-hours) featuring a shifting trend and relatively consistent daily/weekly seasonality.
  • Rossmann Store Sales: Daily retail sales featuring irregular, layered seasonality and abrupt drops on holidays.
  • Beijing Air Quality: Hourly $\text{PM}_{2.5}$ pollutant measurements exhibiting high volatility, low signal-to-noise ratio, and no clean cyclic pattern.

Key Insights and Ideas

  • Exchangeability is not all-or-nothing: Treating time series as completely non-exchangeable is unnecessarily pessimistic. Isolating the remainder isolates the "white noise," justifying standard exchangeable CP ($CV^+$) without sacrificing empirical coverage.
  • Seasonality enables local exchangeability: In clean periodic data, time steps separated by exactly one period behave exchangeably. Exploiting this via BinaryPoint substantially narrows prediction intervals where standard monolithic CP is overly conservative.
  • The "garbage-in, garbage-out" caveat of TSD: When seasonal decomposition is noisy or misses multi-frequency cycles (as in retail sales), rigid local exchangeability assumptions backfire. In such cases, softer decay schemes (ExpLocal) become necessary to maintain stability.

Key Results

  • Synthetic Benchmark: For a target coverage of $90\%$ ($\alpha = 0.1$), the component-wise BinaryPoint method shrank the average prediction interval width (PIAW) to 29.76 (Linear Regressor) compared to 41.69 for raw EnbPI, and to 9.52 compared to 11.92 for Gradient Boosting, while easily exceeding nominal coverage ($94.6\%$).
  • Real-World Trade-offs:
  • On Energy Consumption, the decomposition approach produced tight intervals close to the raw baselines (PIAW 234.97 for BinaryPoint vs. 222.87 for raw EnbPI, with $97.1\%$ coverage).
  • On erratic datasets (Sales and Air Quality), ExpLocal outperformed hard BinaryPoint among decomposed variants (Sales PIAW: 5248.19 for ExpLocal vs. 6045.37 for BinaryPoint).
  • However, directly applying monolithic baselines (EnbPI/ACI) to the raw signal yielded narrower PIs on complex real-world data than the recomposed models. This discrepancy stemmed from decomposition errors compounding across components and additive interval recomposition producing conservative overcoverage ($>95\%$ to $100\%$).

Limitations and Future Work

  • Decomposition Dependency: Classical additive STL struggles with multiple overlapping seasonal frequencies, holiday anomalies, and multiplicative dynamics. Poor decomposition directly inflates interval widths.
  • Conservatism in Recomposition: Simple additive combination of component intervals is overly conservative (yielding empirical coverage far exceeding nominal $1 - \alpha$).
  • Future Directions: Incorporating advanced multi-seasonal decomposition models (e.g., MSTL, RobustSTL, OnlineSTL), accommodating multiplicative recomposition, exploring Mondrian-style adaptive weighting, and establishing tighter non-asymptotic coverage bounds for decomposed combinations.

Summary updated

Abstract

Conformal prediction offers a practical framework for distribution-free uncertainty quantification, providing finite-sample coverage guarantees under relatively mild assumptions on data exchangeability. However, these assumptions cease to hold for time series due to their temporally correlated nature. In this work, we present a novel use of conformal prediction for time series forecasting that incorporates time series decomposition. This approach allows us to model different temporal components individually. By applying specific conformal algorithms to each component and then merging the obtained prediction intervals, we customize our methods to account for the different exchangeability regimes underlying each component. Our decomposition-based approach is thoroughly discussed and empirically evaluated on synthetic and real-world data. We find that the method provides promising results on well-structured time series, but can be limited by factors such as the decomposition step for more complex data.
BibTeX
@inproceedings{prinzhorn2024conformal,
  author = {Derck W. E. Prinzhorn and Thijmen Nijdam and Putri A. van der Linden and Alexander Timans},
  title = {{Conformal time series decomposition with component-wise exchangeability}},
  booktitle = {COPA},
  year = {2024},
  url = {https://openreview.net/forum?id=yJfM2k7hQp},
}

Loading related content...

Figure