RareCP: Regime-Aware Retrieval for Efficient Conformal Prediction
Manuel Heurich, Maximilian Granz, Tim Landgraf · 2026
Read RareCP: Regime-Aware Retrieval for Efficient Conformal Prediction by Manuel Heurich, Maximilian Granz, Tim Landgraf (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-08857.txt.
Introduction
Manuel Heurich, Maximilian Granz, and Tim Landgraf make time series bounds that are both true and tight. Conformal bounds swear a set rate of hits. Plain forms waste width when errors drift with ease, or jump twixt modes. RareCP splits the two roots. A set of pros each learns one error mode. A small net re-tunes the match key as drift sets in.
For each new cast, it pulls the closest past checks. It builds a bound from their errors, weighted, and not the same on both sides. On a top cast test set, bounds grow lean while the hit rate holds. Part by part tests credit each piece.
Catalog record
| CREATOR | Manuel Heurich, Maximilian Granz, Tim Landgraf |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 126 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.08857
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
RareCP: Regime-Aware Retrieval for Efficient Conformal Prediction Manuel Heurich, Maximilian Granz, Tim Landgraf Abstract Recent advances in uncertainty quantification for time series forecasting show that conformal prediction can provide reliable prediction intervals, yet standard conformal methods are often inefficient under temporal dependence, drift, and heterogeneous error behavior. Existing methods typically either update miscoverage rates over time or learn unconstrained calibration weights, without explicitly separating two central sources of nonstationarity: smoothly drifting error distributions and co-existing distinct error regimes. We introduce RareCP , a regime-aware retrieval method for adaptive conformal time series prediction. RareCP learns local calibration representations through a mixture of cosine-attention experts that each capture distinct error regimes, while a compact hypernetwork adapts the kernel parameters to track temporal drift. Given a new forecasting context, RareCP retrieves the top-k most relevant calibration examples, assigns similarity weights, and forms a weighted conformal quantile over their signed residuals, yielding asymmetric prediction intervals. The adaptive kernel is trained using a smooth interval score objective, with a parameter-space anchor to a lightweight teacher kernel to preserve stable local representations. On the GIFT-Eval benchmark, RareCP improves interval efficiency over recent conformal baselines and foundation model uncertainty estimates while maintaining empirical coverage. Ablations confirm that regime-specific experts, drift-adaptive kernels, sparse retrieval, and teacher anchoring each contribute to the final performance. 1 Introduction Time series data is ubiquitous across critical domains including finance, healthcare, and industrial operations. While deep neural networks have dramatically advanced forecasting capabilities, reliable uncertainty quantification (UQ) remains essential for risk-aware decision-making. Traditional probabilistic forecasting methods rely on asymptotic guarantees or impose restrictive assumptions on the data-generating process, such as distributional families, linearity, independence, or stationarity [ 47 ] . This gap between forecasting accuracy and reliable uncertainty estimation motivates distribution-free UQ methods that remain valid under realistic conditions. Conformal Prediction (CP) offers a powerful, model-agnostic UQ framework that provides finite-sample coverage guarantees without distributional requirements, under the exchangeability assumption that the joint distribution of samples is invariant under reordering [ 45 ] . These properties have led to widespread adoption across diverse domains, including image classification, graph-based modeling, and natural language processing. However, the sequential structure of time series data fundamentally violates exchangeability, introducing substantial challenges for CP in forecasting [ 54 , 50 ] . Existing adaptations generally fall into two categories: (1) adaptive miscoverage methods that dynamically adjust the target coverage rate based on observed under- or overcoverage, and (2) weighted CP approaches that assign calibration weights to leverage varying sample relevance. Adaptive conformal methods explicitly address temporal drift by updating the miscoverage level over time, and weighted conformal prediction tackles distribution shift by reweighting calibration samples. However, adaptive methods typically react through changes in the target miscoverage level, while weighted approaches often rely on fixed or non adaptive similarity weighting schemes. We instead argue that time series forecasting errors, i.e., the signed residuals between point forecasts and realized values, exhibit structural biases that can be built directly into the calibration design. These structural biases take two forms: (i) gradual drift, where the similarities of residuals evolves as the underlying process changes, and (ii) regime structure, where residual distributions switch between distinct contexts rather than varying smoothly along a single trajectory. Figure 1: RareCP adapts its 80%-interval prediction locally to the residual regime on m4_daily. Different residual regimes recruit different experts, so the interval tightens in calm stretches and widens across volatile ones rather than inflating uniformly. See Appendix Figure 4 for other examples. We address these limitations with RareCP , a regime-aware retrieval adapter for conformal time series prediction. RareCP calibrates a fixed forecasting backbone by retrieving past signed residuals that are most relevant to the current forecasting context under a learned linear cosine-attention similarity. The retrieved residuals define local lower and upper conformal quantiles, yielding asymmetric prediction intervals that adapt to temporal drift and recurring error regimes (see Figure 1 ). This gives conformal prediction a retrieval-based mechanism for improving interval efficiency while retaining a direct connection to coverage through the size and locality of the calibration set. Our key contributions are: • Conformal prediction is competitive with state-of-the-art quantile forecasters. In the single-step-horizon quantile prediction setting, we show that conformal prediction equipped with RareCP is competitive with the native quantile-based prediction intervals of state-of-the-art time series forecasting models, while remaining backbone-agnostic and applicable to forecasters that only provide point predictions. • Retrieval conformal prediction with an explicit efficiency-coverage trade-off. We introduce sparse residual retrieval for conformal calibration and analyze its effect through a local-CDF framework. The trade-off is controlled by K_{n} , the number of retrieved calibration residuals from a calibration set of size n : smaller K_{n} can improve efficiency by focusing on local residual behavior, whereas increasing K_{n} improves effective sample size and supports asymptotic local quantile and coverage consistency under stated assumptions. • An architecture matched to time series residual structure. We build a linear cosine-attention retrieval architecture designed for two structural biases of forecasting residuals: time-drifting similarities and clustered error regimes. This allows RareCP to adapt conformal intervals locally without relying on unconstrained calibration weights. 2 Related work 2.1 Conformal prediction Conformal prediction (CP) is a model-agnostic, distribution-free framework with finite-sample coverage [ 45 , 39 ] . Classical variants include SplitCP , which calibrates on a held-out set under exchangeability [ 27 ] ; cross-conformal prediction, which aggregates multiple calibration folds for tighter regions [ 46 ] ; and Jackknife+ , which builds intervals from leave-one-out predictions while retaining finite-sample guarantees [ 7 ] . A broad literature extends CP across online, quantile, and tree-based settings [ 31 , 13 , 53 , 37 , 9 , 36 , 23 , 22 ] . A separate line tightens intervals via differentiable conformal surrogates, either end-to-end [ 41 , 6 ] or as post-hoc boosting on a fixed model [ 49 ] . 2.2 Conformal prediction in sequential settings Sequential data violates exchangeability, motivating CP variants that adapt calibration to drift [ 8 ] . One line adjusts the miscoverage rate online: ACI updates \alpha at a fixed learning rate [ 15 ] , AgACI removes manual rate tuning by aggregating multiple learning-rate experts [ 54 ] , DtACI reweights those experts by recent performance [ 16 ] , and Conformal PID Control treats the threshold update as a PID loop [ 2 ] . A second line upweights calibration points by relevance to the test distribution: via covariate or label likelihood ratios [ 43 , 34 ] , recency decay [ 8 ] , or kernel-based weighted quantile regression on nonconformity scores with asymptotic guarantees under mixing [ 26 ] . Learned variants replace these hand-crafted weights with similarity in a representation space or attention over latent nonconformity scores [ 4 , 42 , 10 ] , and 25 structure that similarity weighting as a mixture-of-experts gate over calibration regimes. A separate residual-modeling line builds intervals from bootstrap or ensemble residuals ( EnbPI ) and from residual-distribution estimates ( SPCI ) [ 50 , 51 ] . Retrieval augmentation [ 28 ] has been adopted in time series forecasting via multi-resolution retrieval, relational retrieval with synthesis, diffusion guidance, and LLM-based conditioning [ 18 , 33 , 21 , 29 , 52 ]