NRQL Predictions: Forecasting Future Trends in New Relic

2026-05-20observabilityNRQLMLforecasting

Overview

NRQL predictions in New Relic uses your time series' historical data patterns to predict future trends, providing insights into how metrics might behave in the future.

[!IMPORTANT] NRQL predictions are only compatible with time series queries using the TIMESERIES clause.

How It Works

The system fits a machine learning model to historical data and projects it forward. It supports both seasonal and non-seasonal time series.

Algorithm

NRQL predictions use the Holt-Winters (triple exponential smoothing) algorithm:

Seasonality	Description	Min Data Required
Hourly	Each minute behaves like the same minute in past hours	2 hours
Daily	Each hour mirrors the same hour from yesterday	2 days
Weekly	Each day repeats weekly patterns	2 weeks

Usage

Basic prediction query:

FROM Transaction SELECT count(*) WHERE error IS TRUE TIMESERIES PREDICT

With a custom prediction window:

FROM Transaction SELECT count(*) WHERE error IS TRUE TIMESERIES PREDICT BY 30 minutes

With specified seasonality:

FROM Transaction SELECT count(*) WHERE error IS TRUE TIMESERIES PREDICT holtwinters(seasonality: 1 hour)

When to Use Predictions

Disk space running out as log volume increases
Memory leaks slowly consuming container resources
Projecting future infrastructure costs based on growth trends

System Architecture

Hyperparameters

For advanced users, you can tune the model:

Parameter	Effect	Range
`alpha`	Level smoothing — higher = more weight on recent values	0 to 1
`beta`	Trend smoothing factor	0 to 1
`gamma`	Seasonal smoothing (not for non-seasonal)	0 to 1
`phi`	Trend damping — lower = flatter long-term forecast	0.98 to 1

Example with all parameters:

FROM Transaction SELECT count(*) WHERE error IS TRUE TIMESERIES
PREDICT holtwinters(alpha: 0.2, beta: 0.5, gamma: 0.5, phi: 0.99)
BY 1 hour USING 2 hours

[!TIP] The default PREDICT clause (no extra keywords) gives the best results for most use cases. Only customize if you need fine-grained control.