A prediction interval is a range around a forecast that is expected to contain a future or unobserved value at a specified confidence level. Higher confidence usually produces a wider interval. section 3.2.2
Conformal prediction constructs intervals nonparametrically from historical forecast errors. It uses a user-specified confidence level and does not assume a fixed statistical distribution; time-series use requires care because exchangeability may not hold. section 3.2.2
Prediction intervals support several forecasting and monitoring tasks:
- Quantify forecast risk so decisions can account for a range of plausible outcomes. section 3.2.2
- Represent probabilistic forecasts by selecting quantile bounds from sampled future values. section 4.1 section 5.5.2
- Flag anomalies when observed values fall outside a chosen interval. section 3.8 section 6.4
- Trade recall against precision in anomaly detection: narrower intervals flag more points, while wider intervals reduce false positives. section 3.8.2
Website visits
For a point forecast of 300 daily visits, an 80% prediction interval might span 200–400 visits, while a 90% interval might span 100–500 visits. section 3.2.2
Forecast-based anomaly detection
TimeGPT uses a 99% interval to label observed taxi-ride values outside its historical forecast bounds as anomalies. Chronos and Moirai apply the same basic logic with rolling forecasts and sampled intervals. section 3.8 section 5.8 section 6.4
