ConformalNaive Forecasting Baseline

• ConformalNaive is a training-free, naive forecasting method that wraps last-value predictions with split conformal residual quantiles to define prediction intervals.
• It utilizes the simple random-walk forecast and empirical one-step absolute differences to produce symmetric, data-driven uncertainty bounds without model tuning.
• While it excels in one-step-ahead forecasting by offering competitive coverage and sharpness, its performance degrades in multi-step seasonal forecasts compared to horizon-adaptive alternatives.

ConformalNaive is the simplest possible conformal prediction interval for time series: a last-value, or random-walk, point forecast wrapped in a finite-sample split-conformal residual quantile. In the forecasting literature it is presented as a training-free, no-parameter baseline whose empirical strength depends sharply on horizon: it is a strong floor for one-step-ahead online forecasting, but at multi-step seasonal horizons it becomes the weakest method among those compared, which motivates the companion baselines ConformalSeasonalNaive and ConformalNaive+ (Manokhin, 8 Jun 2026).

1. Definition and terminological scope

ConformalNaive denotes a conformalized naive forecasting baseline rather than a theory of conformal geometry. Despite the lexical resemblance to Ulrych’s “conformal numbers,” which concerns conformal compactification in a hierarchy of hypercomplex projective spaces, bicomplex Vahlen matrices, and Clifford-algebraic representations of Minkowski and Anti-de Sitter space, the forecasting usage is distinct in object, construction, and application (Ulrych, 2017).

In probabilistic time-series forecasting, ConformalNaive is defined by two ingredients. The point forecast is the last observed value, so the predictor is exactly the standard random-walk or naive forecaster. The uncertainty set is then obtained by applying split conformal prediction to the method’s own absolute residuals. The resulting interval is symmetric around the last observed value and requires no model fitting, no hyperparameter tuning, and no learned base forecaster. The 2026 study frames this construction as a “floor”: a minimal benchmark that more complex learned or adaptive systems should clear before claims of improvement are considered persuasive (Manokhin, 8 Jun 2026).

The baseline is not presented as universally optimal. Its significance is comparative. The central empirical claim is that this trivial conformal floor is far stronger than the baselines often reported in learned probabilistic forecasting, particularly in one-step-ahead settings. A related implication is that omission of such a floor can materially distort the interpretation of gains attributed to sophisticated forecasting models.

2. Formal construction

Let the observed univariate series be $y_1,\ldots,y_T$, and suppose the target horizon is $H$. ConformalNaive uses the last observed value as its point forecast: $\mu_1 = y_T$ for one-step-ahead prediction, and in the multi-step batch setting

$\mu_h = y_T \quad \text{for } h=1,\ldots,H.$

Thus the point predictor is exactly

$\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$

Its nonconformity scores are one-step absolute differences,

$s_t = |y_t-y_{t-1}|,\quad t=2,\ldots,T.$

Given scores $s_1,\ldots,s_n$, the finite-sample conformal $(1-\alpha)$-quantile is defined as

$$Q_\alpha = s_{(k)}, \qquad k=\big\lceil (n+1)(1-\alpha)\big\rceil,$$

where $s_{(1)}\le \cdots \le s_{(n)}$, and $H$0 if $H$1. In the reported experiments $H$2, so $H$3 is the empirical 95% conformal quantile of the absolute residuals. The prediction interval is

$H$4

For one step this becomes

$H$5

The online protocol specializes split conformal prediction to sequential time series. The initial calibration set is the pool of past one-step absolute residuals. In the one-step experiments, this pool is initialized from a training window of 800 observations. At each forecast time $H$6, the method computes the current conformal quantile from all residuals accumulated so far, issues the interval centered at $H$7, observes the realized $H$8, and then appends the newly observed residual $H$9 to the pool. There is no model retraining. In the multi-step batch protocol the same residual pool and the same $\mu_1 = y_T$0 are reused for all horizons, so the interval width does not widen with $\mu_1 = y_T$1 (Manokhin, 8 Jun 2026).

For CRPS evaluation, the paper also defines a predictive-sampling procedure based on the same residual-quantile construction. For each level $\mu_1 = y_T$2, it forms the band $\mu_1 = y_T$3, samples uniformly over $\mu_1 = y_T$4, and thereby obtains a piecewise-linear CDF centered at $\mu_1 = y_T$5. Intervals for coverage and Winkler score, and samples for CRPS, are therefore derived from a single residual-quantile mechanism.

The description “training-free” is literal. No optimization is performed. The only operations are computing absolute differences, sorting residuals, and selecting an order statistic. Aside from the nominal miscoverage level $\mu_1 = y_T$6, there are no user-chosen parameters.

3. One-step online forecasting profile

The principal one-step evaluation uses 2,217 real univariate series from nine sources: the Monash forecast archive, the LOTSA collection, the LTSF traffic, electricity, and weather benchmark suites, METR-LA, BOOM, and nips/probts. Frequencies range from minutely to daily. Series shorter than 1,100 observations are excluded, leaving 2,217 of 2,373; the protocol uses an 800-point training window and a 300-step one-step-ahead online test window. Seasonal period $\mu_1 = y_T$7 is taken from dataset metadata rather than estimated. The main metric is the Winkler interval score, with lower values preferred (Manokhin, 8 Jun 2026).

The empirical picture is asymmetric. ConformalNaive dominates weak and moderate baselines, ties simpler learned conformal predictors, and is beaten only by adaptive-online or ensemble methods that explicitly track distribution shift.

Comparator class	Result against ConformalNaive	Regime
NaiveInterval / SeasonalNaiveInterval	ConformalNaive wins on 90% / 92% of series	One-step Winkler
NPTS / SeasonalNPTS	ConformalNaive wins on 73% / 64% of series	One-step Winkler
CSP	ConformalNaive wins on 71% of series; bootstrap 95% CI $\mu_1 = y_T$8	One-step Winkler
RCI / quantile regression	Statistical parity; median relative Winkler within 2%	One-step Winkler
SPCI / ACI / AgACI and related adaptive methods	These methods beat ConformalNaive by 9–33% relative Winkler	One-step Winkler

The CSP comparison is especially prominent: ConformalNaive wins on 71% of series, beats CSP on 7 of the 9 data sources, and the paired Wilcoxon test yields $\mu_1 = y_T$9. Against naive value-quantile baselines and the NPTS family, the Wilcoxon $\mu_h = y_T \quad \text{for } h=1,\ldots,H.$0-values are reported as smaller than $\mu_h = y_T \quad \text{for } h=1,\ldots,H.$1. Against RCI and plain quantile regression, win rates are around 47–48%, which the paper interprets as effective parity (Manokhin, 8 Jun 2026).

Coverage behavior is central to the method’s interpretation. At nominal 95% coverage, mean empirical coverage over the 2,217 series is reported as 0.91 for ConformalNaive. This is slightly below nominal, as expected under temporal dependence, but it is at or above every comparator in the one-step study, including methods that achieve better Winkler scores. Illustrative values reported in the same comparison are about 0.89 for SPCI, 0.90 for ACI, 0.86 for CQR, and 0.85 for CSP. The method therefore does not obtain competitive Winkler performance by being unusually narrow; rather, it combines comparatively strong coverage with sufficient sharpness to outperform many more elaborate baselines.

A common misconception in this setting is that a trivial random-walk conformal interval must be too weak to matter. The one-step results directly contradict that view. The evidence suggests that in many real series the last observed value is already a strong short-horizon predictor, and that the distribution of one-step absolute differences furnishes a stable estimate of local predictive uncertainty.

4. Multi-step seasonal horizons and horizon-adaptive variants

The multi-step evaluation uses the six GluonTS benchmark datasets on which DeepNPTS was introduced: electricity, exchange_rate, solar_energy, taxi, traffic, and wikipedia. Rolling-origin evaluation is performed with horizon $\mu_h = y_T \quad \text{for } h=1,\ldots,H.$2 for hourly datasets and $\mu_h = y_T \quad \text{for } h=1,\ldots,H.$3 for daily datasets, with 380 forecast windows per method, 100 predictive samples per method per window, and $\mu_h = y_T \quad \text{for } h=1,\ldots,H.$4. The metrics are CRPS and empirical 95% coverage (Manokhin, 8 Jun 2026).

In this regime the random-walk floor fails. ConformalNaive is last in CRPS among the compared methods. The paper attributes this inversion to two structural features of the method. First, the point forecast becomes stale as horizon grows because $\mu_h = y_T \quad \text{for } h=1,\ldots,H.$5 for all $\mu_h = y_T \quad \text{for } h=1,\ldots,H.$6; at seasonal offsets, the series may be in a markedly different phase. Second, the interval width does not widen with horizon because the same one-step residual quantile is reused for all $\mu_h = y_T \quad \text{for } h=1,\ldots,H.$7. This underestimates long-horizon uncertainty. In a horizon sweep on four hourly datasets with $\mu_h = y_T \quad \text{for } h=1,\ldots,H.$8, the method has very good CRPS and coverage near 0.99 at $\mu_h = y_T \quad \text{for } h=1,\ldots,H.$9, but by mid-horizon, around $\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$0, coverage falls to about 0.64.

The seasonal alternative is ConformalSeasonalNaive. It replaces the random-walk forecast and one-step residuals by a season-ago forecast and seasonal residuals: $\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$1 This baseline is weak at $\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$2, where season-ago values are often inferior to persistence, but it is substantially better aligned with seasonal multi-step structure. Coverage remains near nominal across horizons in the reported sweeps.

ConformalNaive+ is a horizon-adaptive selector between these two floors. It computes the seasonal-lag error

$\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$3

and, for each horizon $\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$4,

$\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$5

If $\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$6, it uses the ConformalNaive branch; otherwise it uses the ConformalSeasonalNaive branch. The rule is explicitly described as a one-line, training-free, horizon-adaptive selector. In the horizon sweep it tracks the minimum of the two floors: it follows ConformalNaive at $\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$7, switches quickly toward the seasonal branch as persistence degrades, and restores coverage into the approximate 0.85–0.99 range across horizons. The paper notes a remaining gap to a per-window oracle best-of-two envelope because the selector is global per horizon rather than per forecast window (Manokhin, 8 Jun 2026).

The aggregate multi-step scoreboard reflects this logic. CSP-Adaptive has the best CRPS rank overall, with CSP-Fixed also near the top. ConformalNaive+ and ConformalSeasonalNaive are mid-pack. NPTS and SeasonalNPTS rank lower. ConformalNaive is last. At nominal 95% coverage, ConformalNaive, ConformalSeasonalNaive, and ConformalNaive+ are all around 0.84–0.85; the NPTS family and CSP are around 0.89–0.95; DeepNPTS is worst calibrated at 0.66. Per-window Wilcoxon tests further show that ConformalNaive+ beats NPTS on about 64% of windows with $\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$8, slightly beats a protocol-matched DeepNPTS rerun on 51% of windows with $\hat y_{T+1}=y_T,\quad \hat y_{T+2}=y_T,\;\ldots$9, and loses to CSP.

5. Methodological interpretation

ConformalNaive is a direct instance of split conformal prediction with absolute error as the nonconformity score and the naive forecast as the deterministic center. Under exchangeability, split conformal prediction yields exact finite-sample marginal coverage $s_t = |y_t-y_{t-1}|,\quad t=2,\ldots,T.$0. The paper explicitly states, however, that it makes no finite-sample coverage guarantee for ConformalNaive or ConformalNaive+ in time series, because temporal dependence violates exchangeability. Coverage is therefore treated as an empirical metric rather than a theorem in this setting (Manokhin, 8 Jun 2026).

A related misconception is that the method inherits the classical split-conformal guarantee unchanged. It does not. Its favorable one-step performance is empirical. The paper links that performance to the regime in which calibration residuals and next-step residuals are sufficiently similar that the conformal quantile remains informative. This suggests that the success of the method depends on residual stability more than on formal exchangeability.

The study also places ConformalNaive within a calibration-then-sharpness framework. Calibration is assessed by empirical coverage relative to the nominal level. Sharpness is assessed by proper scores such as Winkler score and CRPS. On this view, one-step ConformalNaive is slightly conservative relative to many learned alternatives, with coverage around 0.91 at nominal 0.95, but sharp enough to remain competitive or superior on Winkler score. By contrast, DeepNPTS in the multi-step study achieves only 0.66 coverage at nominal 0.95, which the paper treats as severe overconfidence, while NPTS attains very high coverage, around 0.95, at the cost of poor sharpness and weak CRPS rank.

The method also differs from textbook naive intervals. The point forecast is the same naive or random-walk predictor, but classical statistical intervals often assume Gaussian residuals and widen with horizon, commonly as $s_t = |y_t-y_{t-1}|,\quad t=2,\ldots,T.$1. ConformalNaive replaces those assumptions with empirical split-conformal residual quantiles, but in the reported multi-step protocol it does not widen with horizon. This design choice is precisely what makes it strong as a one-step floor and weak as a long-horizon seasonal forecaster.

6. Role in probabilistic forecasting practice

The normative claim attached to ConformalNaive is that the matching conformal naive floor should be reported whenever a learned probabilistic forecaster claims gains. The argument is empirical rather than axiomatic. In one-step-ahead online forecasting, ConformalNaive beats naive interval baselines, NPTS, SeasonalNPTS, and CSP on a majority of series, and it matches simpler learned conformal predictors such as RCI and plain quantile regression. Only adaptive or ensemble methods that explicitly track distribution shift, including SPCI, ACI, and AgACI, consistently and materially outperform it. A plausible implication is that learned systems failing to exceed this floor at short horizons have not justified their additional complexity (Manokhin, 8 Jun 2026).

The appropriate baseline depends on regime. For short-horizon, especially one-step-ahead online forecasting, ConformalNaive is the relevant reference and can also serve as a deployable training-free method in its own right. For seasonal multi-step horizons, the random-walk floor is not an adequate benchmark by itself; the relevant reference set expands to ConformalSeasonalNaive, ConformalNaive+, and CSP. If horizon behavior is uncertain, ConformalNaive+ is presented as the sensible default among the floor methods because it routes low horizons to persistence and larger horizons to the seasonal branch.

The broader significance of ConformalNaive lies in baseline discipline. Its simplicity removes ambiguity about modeling capacity, optimization quality, or hyperparameter tuning. Because it is literally built from the last value and an empirical residual quantile, it establishes a lower bound on what can be achieved by a forecasting system with essentially no training machinery. The empirical record reported in the 2026 study indicates that this lower bound is nontrivial, especially in one-step settings, and that stronger methods should be evaluated against it rather than against weaker textbook or omitted baselines.