Reservoir-Weighted Conformal Prediction That Accounts for Its Own Memory

Conformal prediction wraps any forecaster with prediction intervals that cover the truth at a chosen rate, say 90%. Its guarantee assumes that past errors are exchangeable with the next one, roughly meaning that their order does not matter. Time series break this assumption. A recent method, ResCP, repairs it cheaply. It passes the series through an echo state network, which is a fixed, randomly wired recurrent network whose internal state acts as a fading summary of the input history. It then builds each new interval mainly from past errors made when that state looked similar to the current one. ResCP reports intervals up to 60% narrower than standard baselines and runs in 52 seconds where a Transformer-based competitor needs over 4,500. It offers no coverage guarantee.

This proposal argues that ResCP's similarity weighting has a specific, predictable failure mode, and that the reservoir itself supplies the means to fix it. A reservoir state changes slowly from one step to the next, so the most similar past states are usually the most recent ones. The errors at those steps are also the most strongly correlated with the current error. Similarity weighting therefore puts most of its weight on a few nearly duplicate data points. The interval behaves as if it were calibrated on a handful of samples, and coverage should break down exactly when the series has long memory or when the reservoir is tuned for long memory. The fix is to exclude a recent buffer of past steps whose length is read off the reservoir's known forgetting rate, and to report how many effectively independent residuals the weights actually use. This is the first step toward a finite-sample coverage bound for state-weighted conformal prediction on mixing processes.

Core Claim

A mixing process is one whose dependence between two time points fades as they move apart. A contractive reservoir driven by such a process produces states in which steps more than $b$ apart are approximately functions of non-overlapping input windows, up to an error of order $c^b$. Here $c<1$ is the reservoir's contraction factor, which is known from its weights by construction. The input's own mixing rate is a different matter: estimating it from one trajectory is ill-conditioned, and that is the field's central practical obstacle. The contraction rate, by contrast, is free. We claim that the coverage error of state-weighted conformal prediction separates into three parts. The first is a localization term that ResCP already exploits. The second is a dependence term controlled by the buffer length $b$ and by the effective sample size of the weights. The third is a tail term $c^b$ from the reservoir. Removing the buffer, as ResCP effectively does, makes the dependence term large precisely when the reservoir sits near the edge of stability, the regime the reservoir literature recommends for memory (Edge of Stability Echo State Network; Gallicchio et al.).

Concretely, with reservoir state $h_t$, weights on past residuals $s \le t-b$ are

$$w_s \propto \exp\big(\cos(h_t,h_s)/\gamma\big), \qquad \mathrm{ESS}_b = \Big(\textstyle\sum_s w_s\Big)^2 \Big/ \textstyle\sum_s w_s^2 ,$$

where $\gamma$ is a temperature and $\mathrm{ESS}_b$ is the Kish effective sample size of the buffered weights. The target theorem bounds the coverage gap by a localization term, plus a term of order $\sqrt{b/\mathrm{ESS}_b}$, plus $\mathrm{ESS}_b\,\beta(b)/b$, plus a term proportional to $c^b$. Here $\beta(b)$ is the input's β-mixing coefficient at lag $b$. The $\sqrt{b/\mathrm{ESS}_b}$ form comes from the blocking argument of Meir as used by Gonon et al.: weighted mass spread over blocks of length $b$ behaves like roughly $\mathrm{ESS}_b/b$ independent draws.

Research Program

The work has three parts.

The estimator. We build the buffered estimator, with $b$ chosen as the smallest lag at which $c^b$ falls below a tolerance tied to the miscoverage level. We add an effective-sample-size floor: if $\mathrm{ESS}_b$ drops below a threshold, the temperature $\gamma$ is raised until it does not. This keeps the cost of ResCP, a single matrix-vector update per step plus a weighted quantile.

The theory. We aim to prove the decomposition above for β-mixing inputs in the causal-Bernoulli-shift class treated by Gonon et al.. That class covers VARMA, GARCH and ARFIMA-type processes, and we reuse its blocking machinery and contraction assumptions. The proof would extend the finite-state Markov coverage-gap analysis and thinning argument of Zheng & Proutiere to continuous reservoir states with non-uniform weights. Their result gives a gap of order $\sqrt{t_{mix}\ln n/n}$ for unweighted split conformal on Markov chains, and it is the template we generalize. KOWCPI already proves validity under strong mixing for kernel-reweighted intervals, but it uses sliding residual windows and solves a convex program at each step. We differ in using a recursive state whose dependence horizon is known rather than estimated.

The tension with long memory. The reservoir theory of Eringis et al. and Gonon et al. requires strict contraction, while practice favors spectral radius near one. Our decomposition makes this trade-off quantitative for uncertainty quantification. Longer reservoir memory gives sharper localization but needs a longer buffer and so yields a smaller $\mathrm{ESS}_b$. We predict an interior optimum in spectral radius for interval width at valid coverage, which is a falsifiable statement.

Scope and an honest caveat. The bound still contains the input's $\beta(b)$, which is unknown. We do not claim to remove it. We claim to remove the reservoir's own contribution and to expose the remaining dependence through a quantity, $\mathrm{ESS}_b$, that can be monitored. Where $\beta$ decays polynomially, the analysis of Wang et al. for static random features suggests that rates degrade in ways that adding features cannot fix. We expect the analogous result here: no buffer length restores i.i.d.-like behavior, and the effective-sample-size diagnostic should flag this.

Baselines

The naive reference is split conformal prediction on a fixed calibration window. It ignores dependence and is expected to under-cover in high-correlation periods, as Zheng & Proutiere observe on French electricity prices.

The strongest existing method on width is KOWCPI. On the Atlanta solar benchmark it reports an average interval width of 48.8 at 90% target coverage, against 84.2 for SPCI and 143.9 for adaptive conformal inference. On ELEC2 it reports a width of 0.22 at 90% coverage. KOWCPI is the strongest competitor because it is the only method in the file that combines localized weighting with a mixing-based validity proof. For ResCP, the file reports lowest Winkler scores (a metric that adds width to a penalty for misses) and the 52-second runtime, but no numerical Winkler values; no reported result was found on Lacuna for ResCP's Winkler score on any benchmark. ResCP is the direct parent of our method and the comparison that tests the central claim.

We will also run K-split conformal prediction from Zheng & Proutiere with its adaptive choice of thinning, NexCP-style exponential time decay, and adaptive conformal inference.

Fair-comparison protocol. All methods share the same point forecaster, the same train, calibration and test splits, and the same hyperparameter search budget, measured in configurations tried on a validation segment. For reservoir methods, the reservoir is shared and fixed across ResCP and our variants. Wall-clock time is measured on identical hardware.

What counts as success. Beating the field means two things. First, worst-window rolling coverage must stay within two points of the 90% target where ResCP falls below it. Second, average width on Atlanta solar must come within 10% of KOWCPI's 48.8, or below it, at a fraction of KOWCPI's runtime.

Ablations. Four ablations isolate the key ingredient. ResCP with the buffer only, and ResCP with the effective-sample-size floor only, separate the two changes. A fixed buffer, tuned on validation, against the reservoir-derived buffer tests whether the contraction rate is the right signal. Replacing the reservoir state with static random features of a lagged window, following Wang et al., tests whether recurrence matters at all.

Why not just…?

Why not just use adaptive conformal inference, which guarantees long-run coverage with no assumptions? This is a real threat to the coverage half of the claim. Adaptive conformal inference adjusts its quantile online and achieves average coverage regardless of dependence. Its weakness is width and local behavior: on Atlanta solar its intervals average 143.9 against KOWCPI's 48.8, and long-run averages can hide sustained runs of under-coverage. Our experiments will report worst-window coverage alongside average width so the two can be compared. We will also test our weights with an adaptive conformal wrapper on top; if that combination dominates, it becomes the recommended method.

Why not just use ResCP as it is, since it already covers at the target empirically? ResCP's reported coverage comes from benchmarks where its reservoir settings were tuned and dependence was moderate. Our claim predicts failure specifically as spectral radius approaches one or as input autocorrelation grows. The synthetic sweep below tests this directly. If ResCP's coverage holds there and its effective sample size stays large, our motivation is wrong, and we will say so.

Why not just thin uniformly with K-split conformal? Thinning restores validity on Markov chains but discards $(K-1)/K$ of the calibration data and ignores where the system sits in its dynamics. Our buffer removes only the recent past while keeping all distant points, weighted by state similarity. The comparison on French electricity prices and on synthetic regime-switching data, where errors depend on the regime rather than elapsed time, will show whether state localization pays for itself beyond thinning.

Why not just use KOWCPI, which already has the theory? It may well match our width. The question is cost and memory length: its per-step optimization over sliding windows scales poorly on long streams, and its windows truncate history. If KOWCPI is both narrower and fast enough on our streams, the contribution shrinks to the theory and the edge-of-stability analysis. We will report that outcome plainly.

Proposed Experiments

Real data. We evaluate on the benchmarks used by ResCP, KOWCPI and Zheng & Proutiere: Solar (50 US locations, plus Atlanta for direct comparison with KOWCPI), Beijing PM10, the 8-currency Exchange data, ACEA Rome consumption, ELEC2, and French electricity prices. The metrics are marginal coverage, worst rolling-window coverage, Winkler score, mean width, $\mathrm{ESS}_b$ and wall-clock time.

Synthetic data with controlled dependence. We use AR(1) processes with autoregressive coefficient swept toward one, which gives geometric mixing that slows continuously. We use ARFIMA processes for polynomial mixing. We use ResCP's piecewise AR with change points to test recovery after a shift.

Decisive checks. The most decisive check is a joint sweep of reservoir spectral radius and input autocorrelation. Our account predicts four things:

  • ResCP's coverage falls as either parameter grows.
  • The decline tracks the drop in its unbuffered effective sample size.
  • The buffered method stays valid along the same sweep.
  • The buffered method's width shows an interior optimum in spectral radius.

The second check is the ARFIMA regime. There we expect every method, including ours, to show coverage that converges slowly with calibration size, with $\mathrm{ESS}_b$ correctly signalling the problem.

Each of these predictions can fail cleanly. A failure would tell us whether the reservoir's own memory is really the hidden source of miscoverage in state-weighted conformal prediction.