- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper studies the validity and efficiency of split conformal prediction for time series, focusing on Conformalized Quantile Regression (CQR) and Conformalized Median Regression (CMR). Unlike the usual routes relying on conventional mixing conditions, the authors use functional dependence measures to establish non-asymptotic bounds on marginal coverage errors and interval-length deviations from their respective oracles. They further investigate Gaussian long-memory processes and derive sharper convergence rates for interval length, along with a matching lower bound under additional conditions. The theoretical results are illustrated through simulation studies and an application to electricity load forecasting.
Critical strengths
-
The use of functional dependence measures extends existing theoretical guarantees to a broader class of time series, including non-mixing and long-memory processes that are not covered by conventional mixing-based analyses.
-
Theorems 1–3 provide non-asymptotic guarantees for both marginal coverage and interval-length accuracy under temporal dependence, addressing an important gap in the existing literature.
-
The results in Section 4 are particularly interesting. Theorem 4 establishes sharper convergence rates for interval length under Gaussian long memory, while Proposition 5 provides a matching lower bound under additional conditions. The simulation studies also provide supporting evidence for these theoretical findings.
Critical weaknesses
-
Theorem 4 is limited to Gaussian linear processes with constant covariates, making it unclear how broadly the sharper rates apply to general time-series regression. Appendix D.4 provides some encouraging results for symmetric non-Gaussian innovations, but the theoretical extension remains open. The matching lower bound also relies on additional restrictions.
-
Assumption 1 requires uniform consistency of the fitted regression functions, which can be difficult to establish for flexible models such as neural networks. Although Appendices B.1–B.2 discuss consistency and misspecified learners, it is unclear what finite-sample rates can be expected for more general learning methods.
-
The connection between the theory and the NYISO application could be clearer. The theoretical analysis assumes stationarity and a fixed data split, while the experiments use rolling refitting. The annual CMR coverage also varies considerably (0.833–0.934). Given that the evidence for long memory is mainly based on estimated Hurst exponents, further discussion of the underlying assumptions would be helpful.
-
The efficiency results are relative to specific oracle intervals, which are not necessarily minimum-length intervals. The experiments also focus largely on verifying the theoretical rates, with few comparisons against existing time-series conformal prediction methods.
Questions for authors
-
Section 6 briefly discusses extending Theorem 4 beyond Gaussian processes. What are the main technical obstacles to obtaining similar rates for symmetric non-Gaussian innovations or models with nonconstant covariates?
-
For a fixed sample size , can the theoretical bounds offer any guidance on choosing the training-to-calibration ratio, especially under long memory?
-
Could the authors clarify how the stationarity assumption relates to the NYISO application, and what the theoretical results imply for coverage under repeated rolling refitting?
3 comments
Response to Reviewer zmGk (2/2)
W3 and Q3
Stationarity. We agree that real series are stationary only approximately. The raw data are the hourly loads, which have strong daily and annual cycles. The operator's day-ahead forecast carries these cycles. We use what remains after the forecast is subtracted, the percentage forecast error, with a separate daily series for each hour of the day. Such a series is much closer to a stationary one. We also checked the homogeneity of the errors over time, through the yearly median and the yearly interquartile range of the daily mean error. The reference level is 2.60% of the forecast, taken from 2005–2013, the years used to fix the design. In the twelve complete evaluation years, 2014–2025, the yearly median lies between 1.87 and 3.04. It is always within one percentage point of the reference, a band fixed before the evaluation. The interquartile range lies between 1.37 and 1.91, against 1.35 to 1.74 in 2005–2013. Only 2026, for which eight months are available, has a median outside the band.
Rolling refitting. The revision closes this gap. The end of Appendix A.4 now shows that, on a stationary series, the guarantees of Theorems 1–3 hold at every origin of a rolling scheme. It also shows that the fraction of covered targets converges almost surely to the common marginal coverage. This concerns the long-run fraction, not the coverage of a single year. The variation of the yearly coverage between .833 and .934 is therefore outside the guarantee. Under long memory, the coverage of a single year is expected to vary more than with independent errors.
Evidence for long memory. The Hurst exponent is a common diagnostic of long memory in applied work, and a series with a Hurst exponent above .5 is regarded as long-range dependent (Leland et al., 1994; Zhai et al., 2026). Long memory is not an assumption of Theorems 1–3, which cover short and long memory alike.
References. Leland et al. (1994) and Zhai et al. (2026) are already cited in the paper.
W4
Oracle intervals. Each procedure has its own target. CQR targets the interval between two conditional quantiles, and CMR targets the conditional median plus or minus a fixed radius. The oracle interval of a procedure is this target. Our theorems show that the calibrated interval attains its length at an explicit rate. In Section 4 and in Studies 2 and 3, the law of the response is symmetric and unimodal. The target is then also the shortest set with coverage among those that do not use the past. On real data no oracle is known, and the hindsight width serves as a proxy for the oracle width.
Comparison. The revision adds Table 3 in Appendix E. It compares CMR and CQR with NexCP (Barber et al., 2023), ACI (Gibbs & Candès, 2021) and a gapped split on the same targets, with the same forecast and the same information. The overall coverage of all rules lies between 89.5% and 90.2%. CMR and CQR have narrower intervals on average than NexCP and ACI, which have steadier yearly coverage. A gap between the blocks changes almost nothing.
Purpose of the experiments. This paper does not intend to propose a new interval. It provides coverage and length guarantees for the standard split CQR and CMR intervals, and the experiments are designed around these guarantees. They are not meant as a benchmark of methods. The simulations test the guarantees where the oracle interval is known. The real data then illustrate them on more than twelve years of targets that were not used to design the procedures. Both intervals cover close to the nominal level, and the mean width of CMR is within 2% of the hindsight width, our proxy for the oracle width.
References. Barber et al. (2023) and Gibbs & Candès (2021) are already cited in the paper.
Q2
Our bounds give orders, so the guidance concerns the orders of and , not their exact ratio. In the model of Section 4, the length error is of exact order for all sufficiently large and , for the centers of Proposition 5 (Theorem 4 and the revised Proposition 5). For , the two terms have the same exponent. The best order, , is then attained when and are both of order , at any fixed ratio. For , the calibration term dominates. The best order, , then needs of order , whereas may be of smaller order, down to . In both cases the ratio changes the constant only, and the theory does not determine the constant.
Thank you for the detailed response and additional experiments, which have addressed most of my concerns. I appreciate the authors' efforts in strengthening the paper and maintain my positive assessment.
Response to Reviewer zmGk (1/2)
We thank the reviewer for the careful reading and the constructive comments. All pointers refer to the revised version, in which the changes are marked in blue.
W1 and Q1
A more general theory of Section 4. The reviewer is right that Theorem 4 was stated for a Gaussian location model and that its extension was open. The revision adds the extension as Theorem 6, with a complete proof in Appendix C.8. Take iid symmetric innovations with a finite fourth moment and a Lipschitz density that is nonincreasing on [0,∞). The bound of Theorem 4 then holds in the form (17) for every 1/2<β<1. Theorem 6 also covers nonconstant covariates. The constant location is replaced by a regression function μ(x) fitted by any learner, with covariates independent of the innovations. The fitted function enters the length through its mean squared error at the calibration covariates, that is, at second order.
Technical obstacles (Q1). There were two. The proof of Theorem 4 used tools that are specific to Gaussian laws. With a covariate, the error of the fitted function varies from one calibration point to the next. The revision replaces the Gaussian tools by a bound on the functional dependence measure that needs only the symmetry of the innovations (Lemma 22 and Step 3 of the proof). It handles the fitted function by conditioning on the training block and on the covariates.
Why a linear process in Section 4. Theorems 1–3 already apply to long-memory processes without a model (Appendix B.5). Section 4 asks for more, a sharper rate with a matching lower bound. Such a rate depends on the form of the dependence, so it is stated for a specified class of processes. This is common practice in the long-memory literature (Dehling & Taqqu, 1989; Wu, 2003). The linear process is the standard class, and its form is representative. By Wold's decomposition, every centered, purely nondeterministic stationary process has a linear representation with uncorrelated innovations (Brockwell & Davis, 1991, Section 5.7).
Lower bound. The revision removes the condition that related m to n. The lower bound (16) now holds for all n≥n0 and m≥m0′. Its proof also gives a lower bound of order rm(β) for the calibration term, the known center included (Step 3 in Appendix C.6). These lower bounds are for the Gaussian location model.
References. Dehling & Taqqu (1989), Wu (2003) and Brockwell & Davis (1991) are already cited in the paper.
W2
Assumption 1 was stronger than necessary, and we have weakened it in the revision (proofs in Appendix A.4). For the coverage bounds (4) and (8) and for the CMR length bound (10), a small fitting error at the calibration and test covariates suffices. Uniform consistency is not needed. When these covariates, given the training block, have a bounded density with respect to the stationary covariate law, this error is controlled by the L1 or L2 error of the learner. An L1 or L2 rate of the learner is then sufficient (Appendix B.1).