📊 statistics

Rolling-Origin Conformal Prediction under Local Stationarity and Weak Dependence

This paper proposes and theoretically validates a rolling-origin conformal prediction method for time-series forecasting that adapts to local non-stationarity and weak dependence by calibrating on recent errors, achieving minimax-optimal coverage rates and demonstrating superior empirical performance over full-history calibration.

Original authors: Stanisław M. S. Halkiewicz

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Stanisław M. S. Halkiewicz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a weather forecaster. Your job isn't just to predict if it will rain tomorrow; it's to draw a circle on a map and say, "I am 90% sure the rain will happen somewhere inside this circle."

In the world of statistics, this circle is called a prediction interval. For decades, statisticians had a perfect tool to draw these circles, but it had one major flaw: it only worked if the weather was completely random and independent every single day (like flipping a coin). But real life isn't like that. Weather patterns depend on yesterday's weather, stock markets depend on last week's trends, and economies drift over time. When you use the old "coin-flip" tool on real-world data, your circles often end up too small (missing the rain) or too big (wasting resources).

This paper introduces a smarter, more adaptable way to draw those circles for time-based data. Here is the breakdown in simple terms:

1. The Problem: The "Stale Menu"

The old method (called Conformal Prediction) works like a restaurant that makes a "special" based on the average of every dish they've ever served in the last 10 years.

The Issue: If the chef suddenly started using spicy peppers last month, the 10-year average is useless. The "special" will be too mild for today's taste.
The Real World: In time series (like stock prices or temperature), the rules change. Volatility clusters (calm days followed by stormy days), and trends drift. Using old data to predict the future creates "stale" intervals that don't match reality.

2. The Solution: The "Rolling Origin"

The author proposes a new method called Rolling-Origin Conformal Prediction.

The Analogy: Instead of looking at the last 10 years of data, imagine the chef only looks at the last 30 days of orders to decide today's special.
How it works: The method constantly updates its "calibration" (the rule for how wide the circle should be) using only the most recent forecast errors. If the weather gets stormy, the method sees the recent big errors and immediately widens the circle. If the weather is calm, it shrinks the circle.

3. The Big Question: How Big Should the Window Be?

If you look at too little data (say, just the last 3 days), your circle might be wobbly and unreliable because you don't have enough information. If you look at too much data (the last 10 years), your circle is too slow to react to changes.

The paper solves the mystery of how many days you should look back.

The Formula: The author proves mathematically that the "sweet spot" isn't a fixed number. It depends on how fast the world is changing.
The Rule: If you have $T$ $T$ total days of data, the perfect window size is roughly $T$ $T$ raised to the power of $2/3$ $2/3$ (about $T^{0.67}$ $T^{0.67}$ ).
- Example: If you have 1,000 days of data, you shouldn't use all 1,000, nor just the last 10. You should use the last ~100 days.
Why it matters: This specific window size is proven to be the best possible choice. No other method can do better at balancing "reacting fast" vs. "being accurate."

4. The "Four-Part" Explanation

The author breaks down exactly why the prediction might be slightly off into four distinct parts, like a mechanic explaining a car's fuel efficiency:

Noise: Even with the right amount of data, there's always some randomness (like wind affecting a car).
Approximation: The math used to simplify the problem isn't perfect, but the error is tiny.
Drift: This is the "stale menu" problem. If the window is too wide, the data inside it is from a different "era" than today.
Model Error: The underlying forecast (the weather prediction itself) might be slightly wrong.

The paper shows that by choosing the perfect window size, you can minimize the total error.

5. Real-World Testing

The author didn't just do math on paper; they tested this on real data:

Six Real Series: Including US inflation, unemployment, stock market returns (S&P 500), and electricity usage.
93 Competition Series: A massive set of data from the famous "M4" forecasting competition.

The Results:

Better Performance: In 86% of the cases, this new "rolling window" method created better, tighter prediction intervals than the old "use-all-data" method.
Accuracy: The method kept the actual outcomes inside the predicted circle about 90% of the time (the target), even when the data was messy or changing.
Speed: It adapted quickly to sudden changes, like the volatility spikes seen in financial markets.

Summary

Think of this paper as a guide for building a self-adjusting safety net.

Old Way: Use a net made of the same size rope for every situation, regardless of how fast you are falling.
New Way: Use a net that automatically changes its size based on how fast you fell in the last few seconds.
The Breakthrough: The author figured out the exact mathematical rule for how much "recent history" to use to make that net perfect. It turns out, looking at roughly the last two-thirds of your data's "power" (specifically $T^{2/3}$ ) is the golden rule for staying safe in a changing world.

Technical Summary: Rolling-Origin Conformal Prediction under Local Stationarity and Weak Dependence

Problem Statement
Conformal prediction offers distribution-free prediction intervals with exact finite-sample coverage guarantees under the assumption of exchangeability. However, this assumption is fundamentally violated in time-series forecasting due to serial dependence, volatility clustering, and gradual distributional drift. Consequently, classical split conformal prediction often yields coverage levels significantly deviating from the nominal $1-\alpha$ . While practitioners frequently employ "rolling-origin calibration"—using only the $m$ most recent pseudo-out-of-sample errors to calibrate the conformal quantile—this approach lacks theoretical justification regarding the optimal window length $m$ and the precise magnitude of coverage deviation as a function of $m$ , sample size $T$ , and the degree of non-stationarity.

Methodology
The paper proposes and analyzes Rolling-Origin Conformal Prediction (ROCP). The method constructs a prediction interval for $Y_{T+h}$ by treating the $m$ most recent absolute forecast errors (scores) as a local calibration sample. The interval half-width is the empirical $(1-\alpha)$ -quantile of these $m$ scores.

The theoretical framework relies on:

Local Stationarity: The score distributions $F_{t,h}$ evolve according to a Hölder- $\beta$ smoothness condition (following Dahlhaus, 1997), where the drift rate is controlled by a constant $L_{F,h}$ .
Weak Dependence: The oracle score process is assumed to be uniformly $\alpha$ -mixing (or satisfy Wu's physical dependence framework), controlling the effective information content of the calibration window.
Quantile Regularity: The density of the score distribution is bounded away from zero and infinity near the target quantile.
Forecast Estimation: The gap between the oracle predictor and the estimated predictor is bounded by a term decaying with $T$ .

The core technical innovation is a four-term decomposition of the coverage error $\Delta_{T,m,h} = |P(Y_{T+h} \in \hat{C}_{T+h|T}) - (1-\alpha)|$ . This decomposition separates the error into:

(I) Quantile Noise: The stochastic variability of the empirical quantile from $m$ dependent observations.
(II) Bahadur Remainder: The error from linearizing the empirical quantile around a deterministic target.
(III) Drift Bias: The systematic bias arising from using scores whose distributions have drifted away from the current target distribution.
(IV) Estimation Error: The cost of using an estimated forecasting model rather than an oracle.

Key Contributions

Optimal Window Rule: The paper derives the minimax-optimal calibration window length $m^\star \asymp T^{2\beta/(2\beta+1)}$ . Under Lipschitz drift ( $\beta=1$ ), this simplifies to $m^\star \asymp T^{2/3}$ . This rule balances the trade-off between reducing quantile noise (favored by large $m$ ) and minimizing drift bias (favored by small $m$ ).
Minimax Optimality: Using a Le Cam two-point construction with Hölder- $\beta$ bump functions, the authors prove that the coverage error rate $O(T^{-\beta/(2\beta+1)})$ achieved by ROCP is minimax-optimal. No conformal procedure can achieve a uniformly better rate over the class of locally stationary processes.
Bahadur Representations: The paper establishes Bahadur representations for the empirical quantile under both $\alpha$ -mixing and the physical dependence framework, providing the necessary linearization to bound the coverage error.
Oracle Inequality for Adaptive Selection: The authors formalize Winkler cross-validation as an adaptive window selector. They prove an oracle inequality showing that the data-driven selector $\hat{m}$ achieves near-minimal validation loss, provided the window grid is sufficiently fine and the uniform concentration condition holds.
Volatility Scaling: A variant (VS-ROCP) is proposed where scores are normalized by estimated conditional volatility. Theoretically, this reduces the drift constant $L_{F,h}$ , potentially allowing for larger optimal windows and narrower intervals when non-stationarity is driven primarily by time-varying scale.

Empirical Results
The theory is validated on six real-world series (macroeconomic, financial, and energy) and a stratified sample of 93 series from the M4 competition:

Performance: Rolling-origin calibration outperforms full-history calibration in 86% of comparisons, with a median Winkler score improvement of 12.3%.
Coverage: The method maintains coverage within $\pm 2\%$ of the 90% target at short ( $h=1$ ) and medium ( $h=5$ ) horizons. At long horizons ( $h=22$ ), coverage degrades slightly (e.g., to 0.871), attributed to model misspecification rather than the calibration window choice.
Scaling Law: A log-log regression of the empirically optimal window $m^\star$ against sample size $T$ across the M4 dataset yields a slope of 0.614 (95% CI $[0.424, 0.805]$ ) after controlling for frequency fixed effects. This is consistent with the theoretical prediction of $2/3$ (for $\beta=1$ ).
Dependence Regimes: Financial series with strong GARCH clustering exhibit optimal windows much smaller than $T^{2/3}$ (ratio $\approx 0.05$ ), while macro series with slow drift exhibit larger windows, confirming the role of the series-specific constant $C_h$ .

Significance and Claims
The paper claims to provide the first rigorous theoretical framework for rolling-origin conformal prediction in time series. It moves beyond the "black box" of heuristic window selection by:

Replacing the impossible goal of exact finite-sample coverage (which requires exchangeability) with a precise, computable approximation error bound.
Establishing that the rolling-origin scheme is not just practically convenient but statistically efficient in the minimax sense.
Providing a principled, data-driven method (Winkler cross-validation) for selecting the calibration window that preserves the optimal convergence rate.

The authors note that the method is model-agnostic and adds no computational overhead to pipelines already performing rolling-origin evaluation. They acknowledge that extending these results to multivariate HPD regions remains an open challenge due to the complexities of high-dimensional density bounds and Bahadur representations.