Conformal Prediction for Time-Series Forecasting
Also known as: conformal prediction, distribution-free prediction intervals, EnbPI, Konformal Tahmin (Conformal Prediction — Zaman Serisi)
Conformal prediction is a distribution-free wrapper that turns any point forecaster — ARIMA, a neural network, or a machine-learning model — into valid prediction intervals using only its residuals. The time-series form was popularised by Xu & Xie (2021) and the modern tutorial treatment by Angelopoulos & Bates (2023).
Key highlights
- Distribution-free: produces valid intervals without assuming a parametric error distribution.
- Model-agnostic wrapper — works on top of ARIMA, neural networks, gradient boosting, or any point forecaster.
- Block-bootstrap (EnbPI) variant respects temporal dependence so coverage holds for dynamic series.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use conformal prediction when you have a continuous time series, already have a trustworthy point forecaster, and need calibrated prediction intervals without committing to a Gaussian or other parametric error model. It expects roughly exchangeable residuals (approximate exchangeability) and a held-out calibration set, with at least about 50 observations. It is the natural choice when the base model is a black box whose native intervals are unreliable, and less suitable when exchangeability is grossly violated, such as under abrupt regime shifts, unless a time-series-aware scheme like block bootstrap is used.
Strengths & limitations
- Distribution-free: produces valid intervals without assuming a parametric error distribution.
- Model-agnostic wrapper — works on top of ARIMA, neural networks, gradient boosting, or any point forecaster.
- Block-bootstrap (EnbPI) variant respects temporal dependence so coverage holds for dynamic series.
- Validity rests on approximate exchangeability of residuals, which abrupt regime shifts can break.
- Requires a held-out calibration set, so it spends data that the base model could otherwise use.
- Coverage is marginal: intervals are calibrated on average rather than conditionally for every input.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What does it mean that conformal prediction is distribution-free?
The interval is built from the empirical quantile of the model's own residuals, so it needs no assumption that the errors are Gaussian or follow any particular distribution. Validity rests instead on the much milder condition of approximate exchangeability of the residuals.
Can I use it with any forecasting model?
Yes. Conformal prediction is a wrapper: you supply any point forecaster — ARIMA, a neural network, gradient boosting — and it calibrates intervals from that model's residuals. It does not change or re-fit the base model's mechanics.
Why a special version for time series?
Standard conformal prediction assumes exchangeable data, which is violated by autocorrelation. The EnbPI scheme of Xu & Xie collects residuals with a block bootstrap so that temporal dependence is respected and coverage holds for dynamic series.
Does it guarantee the interval covers every forecast?
No. The guarantee is marginal — coverage holds on average at the chosen level (1−α) — not conditional on each specific input. A band can still miss for an individual hard-to-predict point.
Sources
- 1.Angelopoulos, A. N. & Bates, S. (2023). Conformal Prediction: A Gentle Introduction. Foundations and Trends in Machine Learning, 16(4), 494-591.
- 2.Xu, C. & Xie, Y. (2021). Conformal Prediction Interval for Dynamic Time-Series. International Conference on Machine Learning (ICML).
You have read it. What now?
Cite this page
ScholarGate. (2026, June 1). Conformal Prediction (Time Series). ScholarGate. https://scholargate.app/econometrics/conformal-prediction-ts