PerformanceTimingConfidence

Instance properties

PerformanceTimingConfidence.randomizedTriggerRate Read only

A number indicating how often noise is applied when exposing the value.

PerformanceTimingConfidence.value Read only

An enumerated value indicating a broad confidence measure of whether a performance record reflects typical application performance, or is likely affected by external factors.

Instance methods

PerformanceTimingConfidence.toJSON()

Returns a JSON representation of the PerformanceTimingConfidence object.

Description

If a website has loaded after a browser "cold start" or session restore, its pages may load more slowly as a result. This can cause a significant difference between real-world dashboard metrics and performance observations in page profiling tools, making it hard for a developer to understand whether a performance issue is a legitimate concern or an outlier caused by external factors.

The PerformanceTimingConfidence interface allows developers to compensate for this problem by returning a browser estimate (in the value property) of the likelihood that a returned performance record represents typical application performance. This is a value of either "low" or "high", indicating the browser's confidence in the measurement.

To reduce the possibility of using the value for fingerprinting, noise is added to the estimate, meaning the value will deliberately be wrong for some proportion of results. The trigger rate for the noise is given in the randomizedTriggerRate property.

Since this can vary across records, per-record weighting is needed to recover unbiased aggregates, to improve data consistency, reduce the number of compound errors, and generally to produce a baseline against which the measured results can be evaluated.

Using the data

You should use the data as follows to extract meaninful information from the randomized values:

1. When collecting PerformanceNavigationTiming records, collect randomizedTriggerRate and value for each record.
2. When computing statistics such as 75th-percentile Largest contentful paint (LCP) or mean page load time, apply the weighting formulas explained below instead of a plain average — this gives you separate, corrected metrics for "typical" loads vs. "degraded" loads.
3. Use the "high" confidence mean/percentile as your "real" performance baseline, and use the "low" one to understand what typical data looks like in cold-start scenarios.

The procedures below illustrate how weighting based on value can be applied before computing summary statistics based on the confidence data.

Computing debiased means

To compute debiased means for both high and low values:

1. For each record:
   • Let p be the record's randomizedTriggerRate.
   • Let c be the record's value.
   • Let R be 1 when c is high, otherwise 0.
2. Compute per-record weight w based on c:
   • For estimating the high mean: w = (R - (p / 2)) / (1 - p).
   • For estimating the low mean: w = ((1 - R) - (p / 2)) / (1 - p).

     Note: w may be negative for some records; you should keep every record.

   • Let weighted_duration = duration * w (see duration).
3. Let total_weighted_duration be the sum of the weighted_duration values across all records.
4. Let sum_weights be the sum of the w values across all records.
5. Let debiased_mean = total_weighted_duration / sum_weights, provided sum_weights is not near zero.

Computing debiased percentiles

To compute debiased percentiles for both high and low:

1. Follow the computing debiased means steps to compute a per-record weight w.
2. Let sum_weights be the sum of the w values across all records.
3. Let sorted_records be all records sorted by duration in ascending order.
4. For a desired percentile (0-100), compute q = percentile / 100.0.
5. Walk sorted_records and for each record:
   • Compute cumulative weight cw per-record: cw = sum_{i: duration_i <= duration_j} w_i.
   • Compute debiased cumulative distribution function per-record: cdf = cw / sum_weights.
6. Find the first index idx where cdf >= q.
   • If idx is 0, return duration for sorted_records[0].
   • If no such idx exists, return duration for sorted_records[n].
7. Compute interpolation fraction:
   • Let lower_cdf be cdf for sorted_records[idx-1].
   • Let upper_cdf be cdf for sorted_records[idx].
   • if lower_cdf = upper_cdf, return duration for sorted_records[idx].
   • Otherwise:
     • Let ifrac = (q - lower_cdf) / (upper_cdf - lower_cdf).
     • Let lower_duration be duration for sorted_records[idx-1].
     • Let upper_duration be duration for sorted_records[idx].
     • Return lower_duration + (upper_duration - lower_duration) * ifrac.

Examples

Basic usage

This example uses a PerformanceObserver to retrieve confidence data from observed PerformanceNavigationTiming entries.

const observer = new PerformanceObserver((list) => {
  list.getEntries().forEach((entry) => {
    console.log(
      `${entry.name} confidence: ${entry.confidence.value}`,
      `Trigger rate: ${entry.confidence.randomizedTriggerRate}`,
    );
  });
});

observer.observe({ type: "navigation", buffered: true });

Specifications

Browser compatibility

See also

• PerformanceNavigationTiming