Reading Percentiles in Monthly Stats

Monthly Stats files store the full percentile distribution of each metric across all NDT tests in a given geography and month. Understanding what percentiles mean — and a critical quirk in how latency and loss are stored — is essential for correct analysis.

💡

Try it yourself: Open the country-level explorer on Binder

What a Percentile Means Here#

A percentile column metric_pN answers: “What was the metric value at or below which N% of tests fell?” For latency and loss, the labels are flipped (see below).

For example, download_p50 is the median download speed — half of all NDT tests in that geography/month were slower, half were faster.

Available percentiles in Monthly Stats: 1, 5, 10, 25, 50, 75, 90, 95, 99.

The distribution tells a richer story than a single number#

A wide gap between p25 and p75 signals high variability — some users have very fast connections while others are very slow. A narrow band means more uniform service.

The Polarity Quirk: Latency and Loss Are Inverted#

This is the most important thing to understand about Monthly Stats percentiles.

For download and upload speed, higher is better. download_p95 is the 95th percentile of speed — only 5% of tests were faster. It represents near-best performance.

For latency and loss, lower is better. Monthly Stats flips the labels for these metrics, so that a higher label always means better performance: latency_p95 is the 5th percentile of latency values, meaning only 5% of tests had lower latency. It therefore represents near-best latency, like download_p95 represents near-best speed.

In summary:

Why does this matter?#

If you read latency_p95 as a standard 95th percentile (near-worst latency), you will draw the wrong conclusions: it is near-best latency. Geographies with high latency_p95 have worse latency, not better.

The Internet Quality Barometer (IQB) handles this by applying the polarity correctly when computing scores: it uses the same label for every metric, by default download_p95, upload_p95, latency_p95, and loss_p95, all of which represent near-best performance.

Code example#

# Correct: for latency, LOWER values are better, so nsmallest finds best
best_latency = df.nsmallest(10, "latency_p50")[["country_code", "latency_p50"]]

# Correct: for download, HIGHER values are better
best_download = df.nlargest(10, "download_p50")[["country_code", "download_p50"]]

# Wrong: this finds countries with worst latency, not best
worst_approach = df.nlargest(10, "latency_p50")  # ← these are the SLOWEST countries

Which Percentile Should I Use?#

The right choice depends on your question:

Use p50 (median) when you want the typical user experience. The median is robust to outliers and represents a realistic “what does this network feel like for most people?” summary.

Use p95 for download/upload when you want near-peak performance — useful for understanding what fast connections look like in an area. The IQB framework uses p95 download/upload as an optimistic baseline for its quality thresholds.

Use p95 for latency/loss too when you want near-best values, because the labels are flipped. This is the IQB convention.

Use p25–p75 to understand variability and equity. A location where download_p75 / download_p25 is very large has high inequality between its fast and slow users.

Sample Count and Reliability#

The sample_count column records how many NDT tests contributed to each row. Rows with very few tests have noisy percentiles:

# Filter out low-confidence rows before analysis
df_reliable = df[df["sample_count"] >= 100]

As a rough guide:

• < 30 tests: very unreliable — avoid using percentile estimates
• 30–100 tests: usable with caution, especially for p50
• > 100 tests: generally reliable for p25–p75
• > 500 tests: reliable for extreme percentiles (p5, p95)

Finer-grained slices (city, city+ASN) will have far more low-count rows than country-level data, especially for smaller cities or minority ISPs.

Further Reading#

• M-Lab Monthly Stats Dataset — dataset overview, access, and structure