Reservoir Conformal Prediction for Time Series

• Reservoir Conformal Prediction (ResCP) is a method that uses reservoir computing to dynamically weight calibration residuals, thereby adapting prediction intervals to local temporal dynamics.
• It employs fixed, random Echo State Networks to generate high-dimensional reservoir states which capture fading-memory summaries for effective uncertainty quantification.
• The approach achieves asymptotic conditional coverage with computational efficiency, making it ideal for online forecasting and resource-constrained environments.

Searching arXiv for the main paper and closely related conformal prediction baselines to ground the article in current literature. Reservoir Conformal Prediction (ResCP) is a training-free conformal prediction method for time series forecasting that uses reservoir computing to dynamically reweight conformity scores, with the goal of producing prediction intervals that adapt to local temporal dynamics while remaining computationally scalable (Neglia et al., 6 Oct 2025). It is introduced in "ResCP: Reservoir Conformal Prediction for Time Series Forecasting" (Neglia et al., 6 Oct 2025) as a response to a central limitation of classical split conformal prediction in sequential settings: calibration residuals and test points are not exchangeable in time series, and older residuals may become uninformative under temporal dependence, non-stationarity, and heteroskedasticity. ResCP addresses this by comparing the current reservoir state with historical calibration states, then using the resulting similarity structure to form a weighted empirical residual distribution. Under reasonable assumptions, the method is shown to achieve asymptotic conditional coverage, and the paper reports empirical effectiveness across diverse forecasting tasks (Neglia et al., 6 Oct 2025).

1. Sequential uncertainty quantification problem

Conformal prediction offers a framework for building distribution-free prediction intervals for exchangeable data, but time series violate exchangeability because observations are sequentially dependent and frequently subject to distribution shift (Neglia et al., 6 Oct 2025). In this setting, using all calibration residuals equally can produce conservative intervals that ignore local temporal dynamics. The motivating claim of ResCP is therefore not merely that temporal dependence exists, but that the calibration residual distribution should be localized to histories that resemble the present state of the system.

The forecasting setup is defined with a scalar target $y_t$ and optional exogenous covariates $u_t \in \mathbb{R}^{D_u}$. A base point forecaster produces $H$-step-ahead predictions

$$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$

for window length $W \ge 1$ and horizon $H \ge 1$ (Neglia et al., 6 Oct 2025). The base forecaster is arbitrary; the experiments use GRU-RNN, decoder-only Transformer, and ARIMA. ResCP then operates on residuals

$r_t = y_t - \hat{y}_t,$

using raw residuals rather than absolute values, which permits asymmetric intervals through separate lower and upper quantile estimation (Neglia et al., 6 Oct 2025).

A central design choice is that ResCP remains a split-conformal scheme with a disjoint calibration set. This allows it to retain the procedural simplicity of split conformal prediction while modifying only the way calibration residuals are aggregated. A plausible implication is that the method is intended as a plug-in uncertainty layer around an existing forecaster rather than as a replacement for the forecasting model itself.

2. Reservoir computing formulation

ResCP uses reservoir computing, specifically Echo State Networks (ESNs), to transform recent history into high-dimensional dynamical states. The paper emphasizes that the reservoir weights are fixed and random rather than trained, so the method avoids fitting a complex sequential model while still extracting temporal structure from the input stream (Neglia et al., 6 Oct 2025). In the base ResCP variant, the ESN input $x_t$ is chosen to be the residual $r_t$; variants can include exogenous covariates.

The ESN state update is

$$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$

where $u_t \in \mathbb{R}^{D_u}$0 is the reservoir state, $u_t \in \mathbb{R}^{D_u}$1 and $u_t \in \mathbb{R}^{D_u}$2 are fixed random matrices, $u_t \in \mathbb{R}^{D_u}$3 is a random bias, $u_t \in \mathbb{R}^{D_u}$4 is the leak rate, and $u_t \in \mathbb{R}^{D_u}$5 is typically $u_t \in \mathbb{R}^{D_u}$6 (Neglia et al., 6 Oct 2025). The initialization of $u_t \in \mathbb{R}^{D_u}$7 is chosen to ensure stable dynamics through the Echo State Property; the paper notes spectral radius $u_t \in \mathbb{R}^{D_u}$8 as a sufficient condition, while also observing that driven reservoirs can operate with $u_t \in \mathbb{R}^{D_u}$9 under certain settings.

The function of the reservoir is representational rather than predictive. ResCP does not require a learned readout from reservoir states, because the states are used only as a similarity space for weighting calibration residuals. This is a consequential distinction from many ESN applications, where a readout is trained for prediction. The paper also defines a related variant, ResCQR, in which a linear readout is trained on top of reservoir states to directly predict quantiles when informative covariates are available and sufficient calibration data exist (Neglia et al., 6 Oct 2025).

The key intuition is that the reservoir state provides a stable fading-memory summary of the recent past. If the current state resembles a past calibration state, then the residual observed after that past state is treated as more informative for the present uncertainty quantification problem.

3. Similarity weighting and interval construction

The weighting mechanism is the defining feature of ResCP. For a query at time $H$0, the method compares the current state $H$1 with stored calibration states $H$2. In the experiments, cosine similarity is used:

$H$3

although the method description notes dot-product as an option and states that cosine was preferred empirically (Neglia et al., 6 Oct 2025).

Given similarity scores $H$4, ResCP forms normalized softmax weights

$H$5

with temperature $H$6 (Neglia et al., 6 Oct 2025). Lower $H$7 concentrates mass on the most similar states; higher $H$8 approaches uniform weights and recovers vanilla split conformal prediction. To adapt under temporal drift, the method can optionally apply time-dependent discounting,

$H$9

where $$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$0 is time distance and the paper reports a simple linear decay $$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$1 working well in practice, combined with a fixed-size FIFO calibration window of size $$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$2 (Neglia et al., 6 Oct 2025).

The weighted residual distribution is represented through the weighted empirical CDF

$$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$3

and the corresponding weighted quantile

$$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$4

In practice, the paper approximates weighted quantiles via Monte Carlo sampling from the weight distribution and then computing the standard empirical quantile over sampled residuals, while also noting that direct weighted quantile computation without sampling is $$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$5 with cumulative weights (Neglia et al., 6 Oct 2025).

For target level $$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$6, ResCP can form a potentially asymmetric interval using lower and upper residual quantiles:

$$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$7

$$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$8

To tighten widths for skewed residuals, the paper also searches over $$\hat{y}_{t+H} = \hat{f}(y_{t-W:t},\, \mathbf{u}_{t-W:t}),$$9:

$W \ge 1$0

and then sets

$W \ge 1$1

This interval construction makes the method explicitly localized, weighted, and potentially asymmetric, while preserving the split-conformal regime (Neglia et al., 6 Oct 2025).

4. Theoretical guarantees

The theoretical contribution of ResCP is an asymptotic conditional coverage result obtained through consistency of the weighted empirical CDF under mixing and regularity assumptions (Neglia et al., 6 Oct 2025). The assumptions include a time-invariant $W \ge 1$2-mixing process for

$W \ge 1$3

with $W \ge 1$4 as $W \ge 1$5, together with a well-defined, causal ESN state map having fading memory.

The ESN stability assumption is expressed through a fading-memory metric $W \ge 1$6 and a constant $W \ge 1$7, such that

$W \ge 1$8

A further assumption requires the conditional CDF $W \ge 1$9 to be continuous in the fading-memory sense (Neglia et al., 6 Oct 2025). The weighting scheme is also constrained through an effective sample size

$H \ge 1$0

with $H \ge 1$1 decreasing slowly enough that $H \ge 1$2, and with weights concentrating within shrinking neighborhoods of $H \ge 1$3:

$H \ge 1$4

Under these conditions, the main theorem states

$H \ge 1$5

The proof idea decomposes the error into stochastic fluctuation and bias due to state mismatch. According to the paper, $H \ge 1$6-mixing controls covariances, boundedness yields variance bounds, the effective sample size condition makes variance vanish, and the concentration of softmax weights together with ESN fading memory and continuity of the conditional law controls the bias term (Neglia et al., 6 Oct 2025).

The corollary is asymptotic conditional coverage:

$H \ge 1$7

The paper interprets this as a conditional guarantee given the reservoir state, provided that $H \ge 1$8 sufficiently summarizes historical dynamics. In the degenerate case of non-informative states, ResCP reduces to marginal coverage (Neglia et al., 6 Oct 2025). This suggests that the quality of the guarantee in practice depends materially on whether the reservoir state is an informative summary of the relevant temporal regime.

5. Computational profile and operational workflow

ResCP is explicitly designed for computational efficiency in streaming settings. Per step, updating the ESN state costs

$H \ge 1$9

where $r_t = y_t - \hat{y}_t,$0 is reservoir sparsity, and the paper notes that in typical implementations this is near-linear in $r_t = y_t - \hat{y}_t,$1 (Neglia et al., 6 Oct 2025). Computing similarities against $r_t = y_t - \hat{y}_t,$2 stored states costs $r_t = y_t - \hat{y}_t,$3 with cosine or dot-product, softmax normalization costs $r_t = y_t - \hat{y}_t,$4, weighted quantile estimation by Monte Carlo with $r_t = y_t - \hat{y}_t,$5 samples adds $r_t = y_t - \hat{y}_t,$6, and direct weighted quantile computation without sampling is $r_t = y_t - \hat{y}_t,$7. Memory usage is $r_t = y_t - \hat{y}_t,$8 for stored reservoir states plus $r_t = y_t - \hat{y}_t,$9 for residuals.

The streaming workflow is straightforward. A base forecaster is trained on a training split, residuals are computed on the calibration split, and the ESN is driven on chosen inputs to obtain and store calibration states together with future residuals (Neglia et al., 6 Oct 2025). At prediction time, the current ESN state is updated, the base point forecast $x_t$0 is obtained, similarities and weights are computed over the active calibration window, and a prediction interval is formed from weighted residual quantiles. After observing $x_t$1 in an online setting, the realized residual is appended to the FIFO buffer and the oldest entry is dropped if necessary.

Because no model refitting is required, adaptation to shift occurs through time-dependent discounting and sliding-window updates rather than retraining. The paper contrasts this with learned time-series conformal methods such as SPCI and HopCPT and reports substantially smaller runtime for ResCP; on the Solar dataset, the runtime table gives approximately $x_t$2 for ResCP versus approximately $x_t$3 for HopCPT (Neglia et al., 6 Oct 2025). That comparison is specific to the paper’s experimental setup, but it underscores the intended operating regime of ResCP: online or resource-sensitive forecasting with evolving residual structure.

The paper also provides concrete hyperparameter guidance. It reports grid search over spectral radius $x_t$4, leak $x_t$5, input scaling $x_t$6, temperature $x_t$7 in $x_t$8, and window size $x_t$9, with model selection by minimizing the Winkler score on a validation slice comprising $r_t$0 of calibration while ensuring acceptable coverage (Neglia et al., 6 Oct 2025). The practitioner checklist gives representative settings such as $r_t$1, $r_t$2--$r_t$3, $r_t$4--$r_t$5, input scaling $r_t$6--$r_t$7, cosine similarity, and a starting temperature around $r_t$8.

6. Empirical behavior, comparative position, and limitations

The empirical study evaluates ResCP on Solar, Beijing Air Quality, Exchange rates, and ACEA, using SCP, NexCP, SPCI, HopCPT, and a GRU-based quantile regression baseline called CoRNN as baselines, together with GRU-RNN, decoder-only Transformer, and ARIMA base forecasters (Neglia et al., 6 Oct 2025). The reported evaluation metrics are Coverage gap $r_t$9, PI-Width, and Winkler score.

The main findings reported in the paper are that ResCP achieved approximately valid coverage and substantially narrower intervals than SCP and NexCP, with up to approximately $$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$0 width reduction, and that it often outperformed HopCPT while remaining training-free (Neglia et al., 6 Oct 2025). CoRNN performed strongly on large datasets with informative exogenous inputs, such as Solar, but degraded under smaller or shifting regimes such as ACEA and Exchange. ResCQR was reported as competitive when exogenous variables were beneficial. Sensitivity analyses further indicated that $$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$1 has an optimal range, too-small $$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$2 harms coverage, higher leak rates improve responsiveness, moderate input scaling avoids $$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$3 saturation, and adequate reservoir size is dataset dependent (Neglia et al., 6 Oct 2025).

Within the landscape of conformal methods for time series, the paper positions ResCP relative to several lines of work. NexCP uses data-independent exponential time-decay weights and is described as robust to shift but insensitive to local dynamics. SPCI sequentially re-estimates quantiles with learned regressors and is characterized as computationally heavier because it requires fitting at each step. HopCPT uses soft-attention weights via modern Hopfield networks and is described as powerful but expensive to train and sensitive to small calibration sizes. Kernel-based optimally weighted conformal prediction is presented as conceptually similar in its use of weighted residuals, but dependent on kernel bandwidth selection and potentially heavier computationally in high-dimensional covariate spaces. CQR, conformalized quantile regression, EnbPI, kNN-CP, and localized conformal prediction are all discussed as related approaches, with ResCP framed as a localized conformal method specialized to time series through ESN-based features (Neglia et al., 6 Oct 2025). A plausible synthesis is that ResCP occupies a middle position between purely time-decayed conformal baselines and more highly learned adaptive conformal methods.

The limitations identified in the paper are equally specific. Abrupt regime changes may not be captured by similarity in ESN state space; proposed mitigations include stronger time discounting, smaller windows, and reset strategies (Neglia et al., 6 Oct 2025). Poor reservoir design can yield uninformative states, motivating grid search and sensitivity diagnostics and suggesting architectures near the “edge of stability.” High-dimensional covariates can distort states if included unfiltered; the paper recommends ResCQR with a simple readout trained on calibration, or covariate preselection or transformation. Very small $$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$4 or very narrow windows reduce the effective sample size and can destabilize quantiles; the stated mitigation is to enforce a minimum $$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$5 by bounding $$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$6 from below and window size from above.

The extensions proposed in the paper include multivariate time series and joint uncertainty through ellipsoidal or copula-based sets, combining reservoir localization with heteroskedastic residual modeling, integrating ResCP weighting with ensemble or probabilistic base forecasts, online adaptation of $$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$7, window size $$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$8, and discount $$\mathbf{h}_t = (1 - l)\,\mathbf{h}_{t-1} + l\;\sigma\!\left(W_x \mathbf{x}_t + W_h \mathbf{h}_{t-1} + \mathbf{b}\right),$$9, and strengthening the theory toward finite-sample conditional guarantees through randomized localized conformal corrections (Neglia et al., 6 Oct 2025). These directions remain prospective, but they indicate that ResCP is intended as a general localization principle for sequential conformal inference rather than as a method restricted to a single experimental template.