Methodology

The British Resilience Index is a non-partisan data project measuring stress across key systems of national life. It does not assign blame to political parties. It tracks whether core systems are improving, deteriorating or entering dangerous levels of strain.

The index combines 11 domains, each measuring a different national system, weighted by their estimated contribution to systemic resilience. Each domain score is the median of its component metric stress scores. The national score is a weighted sum of domain medians.

Current release: September 2026 · National score: 51/100 · Band: Fragile

The index does not:

• Attribute blame to any political party, government, or named individual.
• Predict policy outcomes or future conditions.
• Model cascading system failures or tipping points.
• Provide individual-level data on any person.
• Measure cultural, moral, or ideological conditions.
• Claim to know at what score a society becomes ungovernable.

It tracks whether measured systems are under more or less stress than in previous periods. The interpretation of why conditions have changed is left to the reader.

The index produces a single 0 to 100 reading. Higher scores indicate greater measured stress across national systems. The score is grouped into five bands:

• Stable (0–24): measured stress is low and within the normal historical range.
• Strained (25–49): under noticeable pressure, but not yet at dangerous levels.
• Fragile (50–74): significant strain; several systems are weak and vulnerable to shocks.
• High stress (75–89): severe strain by the standards of the measured record.
• Extreme stress (90–100): the most severe readings the scale describes.

The bands mark position on a relative stress scale. They are not calibrated against observed system failure: the index measures how stressed each indicator is against its own record and against published bounds, and it does not know the point at which a service or a system stops working. The top two bands were labelled Critical and Crisis until August 2026, which claimed a calibration the model does not have.

The score is not a prediction of societal collapse; it tracks whether measured systems are under more or less stress than in previous periods.

The index currently comprises 11 domains. Weights sum to 100. Each domain is scored as the median of its component metric stress scores. Weights represent the estimated share of systemic national resilience carried by each domain. These weights are currently set by editorial judgement and will be reviewed formally.

Each metric produces a stress score 0–100 from two sub-scores, over whichever of them could actually be measured:

Volatility left this blend in August 2026. It had carried a weight of 0.1, but of the 71 scored indicators it was pinned at its maximum for 14 and was a placeholder for 19, so on close to half the register a component that always pushed stress upward was either saturated or invented. It is still computed and published on each metric as a diagnostic, and it is the input to a separate volatility measure. The distinction is deliberate: this index describes the condition of a system, not how fast the system is currently moving.

A component that cannot be formed at all contributes nothing and the remaining weight is rescaled. Before August 2026 it contributed a neutral 50 instead, which meant an indicator with a short history could not score below 20 or above 80 while an indicator with a long one could span the full range. That was an artefact of history length rather than a statement about uncertainty, and uncertainty is carried separately by the confidence score.

• Level score (two thirds): where the current value sits relative to the metric’s own historical distribution. Metrics with ≥ 8 verified historical points use historical-percentile normalisation: a value at the 90th percentile of its own history scores ~90; the median scores ~50. Direction of badness is automatically accounted for. Shorter series fall back to provisional best/worst bounds, flagged in the metadata as the weaker method. Currently 46 of 73 metrics are scored on historical-percentile rank and 27 on provisional bounds.
• Trend score (one third): direction and velocity of recent change. trend = clamp(50 + 500 × worse_frac, 0, 100) where worse_frac is the signed fractional change (positive = worsening). A 10% deterioration or improvement spans the full 0–100 range.
• Volatility score (published, not scored): standard deviation of period-over-period fractional changes (if ≥ 4 history points; else 50). volatility = clamp(sd × 800, 0, 100). Shown on each metric for transparency; it does not enter the stress score.

trendDirection is derived from the trend score: >55 = deteriorating; <45 = improving; else stable.

The retrospective national trend line shown on the dashboard uses a level-only method (no trend or volatility component) applied consistently back to 2018. This ensures the historical trajectory uses a single method end-to-end. The current headline composite includes all three components and may differ from the final plotted point; this difference is disclosed beneath the chart.

Each metric carries a confidenceScore (0–1) reflecting source quality, update frequency, and known weaknesses. The index dataConfidence is the mean of all metric confidence scores.

Current index data confidence: 0.83. Every colour-coded confidence signal includes a text label; colour is never the sole indicator.

The national score is a weighted arithmetic mean of domain medians, normalised over the domains present in the view. Register weights sum to 100, so for the UK and the four nations (all eleven domains present) this is simply the weighted sum divided by 100; for views carrying a subset of domains (the nine English regions) the denominator is the present weight, so an absent domain is treated as no information rather than a score of zero.

Correction, July 2026 release: earlier releases divided by the full register weight in every view, which understated the nine English regional scores. Two further changes in this release: provisional seed placeholders are excluded from all scoring, and the headline movement is now the change since the previous published release.

Higher-weighted domains have a proportionally larger influence on the headline figure. The Stress Contribution chart on the dashboard visualises this: each bar equals score × weight.

These eleven domain weights are an editorial interim. A future version will derive them through a documented budget-allocation expert elicitation (Step 6 of the OECD/JRC Handbook on Constructing Composite Indicators), with the weights remaining the single editable source in the index data. Crucially, the headline does not hinge on the exact figures: the robustness analysis below re-scores the index under randomised weights (and other methodological choices) and the UK national score stays within 44–52 (median 48, headline 51), so it is robust to the weighting choice.

A single set of methodological choices, these domain weights, this normalisation method, the median domain rule, the weighted-arithmetic national aggregation, and the 0.6/0.3/0.1 blend, produces the headline figure. To test how much the headline depends on those choices rather than on the underlying data, the index is re-scored under 2,000 plausible methodological combinations (a Monte Carlo analysis, seeded for reproducibility).

The headline national score is 51. Across 2,000 plausible methodological combinations, varying the domain weights, normalisation method, aggregation rule, and the level and trend blend, it ranges 44–52 (median 48). The published headline can sit a point or two from the simulation median: the headline uses the production method exactly, while the median averages over many alternative methods, several of which score systematically lower. The score is therefore robust to methodological choice. The largest single driver is the domain aggregation rule (median vs mean vs geometric).

How to read that attribution. Each parameter is varied on its own with the others held at their production setting, and the share reported is that parameter’s share of the summed range. It is a range attribution, not a variance decomposition, so it measures no interaction between parameters. Two of the five families are sampled over a continuous range and three are enumerated exhaustively over a handful of discrete modes; a range widens with the number of draws, so the two groups are not strictly comparable and equal spreads are reported as ties rather than ranked.

This follows Step 8 (“Robustness and sensitivity analysis”) of the OECD/JRC Handbook on Constructing Composite Indicators, which treats a published range under alternative methods as a core requirement for a defensible composite rather than an optional extra. The headline figure remains the weighted mean of domain medians described above; the range quantifies the uncertainty around it.

Where several domains sit at or above 75 at the same time, the count is published as a signal. It does not change the national score. This release: 1 of 11 domains at or above 75.

Until August 2026 this was a scored penalty: +5 points on the national score when three or four domains crossed 75, +10 at five or more. It was retired for three reasons. The cliff was arbitrary, so three domains at 74 cost nothing while three at 75 cost five points, and two domains at 100 cost nothing at all. The magnitude had nothing to calibrate against, because this index carries no model of how a failure in one domain would propagate to another and says so plainly elsewhere on this page. And it was never applied to the retrospective trend line, so the headline and the published history sat on scales that could differ by as much as ten points.

The signal it was reaching for is real and is kept. What was removed is the claim to know what that signal costs. systemicPenaltyApplied remains in the published data, fixed at 0, so that archived releases continue to validate against the schema they were written under.

All metrics must be sourced. Sources are classified into five tiers reflecting data quality and methodological rigour:

Full source list: data sources page.

Each release is named for the month following its data cutoff. The cutoff is the 25th, and the release is published within ten days of it, so a release normally appears in the last week of the cutoff month or the first week of the month it is named for. Source publications up to the cutoff are incorporated, and anything published later appears in the next release. The index publishes every month regardless of how many indicators received new data, and each Monthly Review states how many did.

The cutoff, not the month in the name, is what tells you how current the data is. This release is named September 2026 and carries source data published up to 25 August 2026. It contains no data from September 2026 itself. The cutoff is recorded on every release, so the vintage is checkable rather than inferred from the name.

This wording was revised in August 2026 and its dates were corrected in September 2026. The policy previously promised publication in the first week of the month, which the record barely supports: June was published on 24 June, July on 7 July, August on 13 August and September on 26 August, each being the day the release went live. Only July fell inside a first week, and on its last day. A commitment the record barely honours is worse than a looser one every release meets, so the policy is now anchored to the data cutoff, which is the part that governs what is actually in a release and is published with it. June and July predate the cutoff convention; August 2026 was a transition release, named for its own cutoff month rather than the one after.

All releases are versioned and preserved. When source data is revised (as is common with ONS administrative series), the metric value is updated but the original release reading is retained in the metric history, labelled with its release period.

Methodology changes are disclosed in the Monthly Review section 10 of the relevant release. Significant methodology changes trigger a re-calibration notice and a comparison table showing how scores would have differed under the old method.

The following ten rules govern the construction of every metric, domain, and narrative in this index. They are reproduced verbatim from the project contracts.

1. The index does not assign blame to parties.
2. The index tracks conditions and trends.
3. Every metric must have a source.
4. Every metric must have a stated weakness.
5. Weightings are transparent.
6. Revisions are preserved.
7. Survey/perception data is separated from hard administrative data.
8. Culture-war claims are excluded from the core score.
9. Immigration is only measured through neutral capacity-pressure indicators, not moral framing.
10. The project distinguishes between deterioration, low baseline, and poor data quality.

Approved vocabulary: under pressure, elevated stress, fragile, critical, deteriorating, improving, material movement, data confidence, requires monitoring, stable, resilient.

Banned: “collapse”, “collapsing”, “broken”, any UK political party name, any named politician, “the government has failed”. “crisis” is allowed only inside a registered source’s own dataset name.

The index is called British, so it should say plainly how British each reading is. Several public service and housing series are published for England only, or for England and Wales, because the service itself is devolved and the devolved administrations publish their own measures on their own definitions. The United Kingdom view therefore leans on those series. The table states the lean, by the share of domain weight each geography carries, which is what actually moves the score.

Across 71 scored indicators in this view. A devolved measure is never blended into a United Kingdom figure to fill a gap: where a nation publishes its own equivalent, it is scored in that nation’s own view instead, which is why the four nation views carry different compositions and why they are compared at the stress score level rather than by raw value.

The following limitations are documented honestly. They do not invalidate the index but should be borne in mind when interpreting any specific reading.

• England-only bias: several public services and housing metrics use England-only administrative data, reflecting devolved structures. Scottish, Welsh and Northern Irish equivalents are registered for future addition.
• Survey lag: welfare and perception indicators (HBAI, CSEW, community surveys) carry a 12–18 month lag. Current readings may understate or overstate current conditions.
• Short-history metrics: some metrics use provisional best/worst bounds rather than historical-percentile normalisation; each metric carries a "normalisation" field for transparency, and bounds-based metrics migrate to percentile as their verified series lengthen (which can shift the headline slightly).
• Band boundaries are thresholds on a continuous score, not sharp categories: a national or domain reading within a point or two of a threshold should be read as sitting between bands rather than as a sharp category change.
• Percentile interpretation: a metric that has been chronically stressed will score around its own historical median (~50) even if the absolute level is high, because percentile is relative to its own track record. The trend component (a third of the blend) separately captures whether stress is worsening or improving.
• Domain weights: the eleven domain weights are an editorial interim, set by judgement rather than empirical optimisation. A future version will derive them via a documented budget-allocation expert elicitation (OECD/JRC Handbook Step 6). Crucially, the sensitivity analysis already demonstrates that the headline is robust to weight variation, re-scoring under randomised weights leaves the national score within the published robust range, so the interim weights are not load-bearing for the headline verdict.
• Volatility score: published on each metric as a diagnostic but excluded from the stress blend since August 2026. It saturates easily and defaults to 50 where a metric has fewer than four history points, which is why it no longer enters the composite.
• Metric independence: some metrics within a domain (e.g. housing affordability and social-housing waiting lists) are positively correlated; using the median partially mitigates double-counting but does not eliminate it.

The first of those, publication lag, is measured rather than asserted: how old the data is reports the age of every scored indicator and how far it varies by domain.

Each domain page includes an International context card showing how the UK compares to the OECD or G7 on a single headline indicator for that domain. These figures are sourced separately from the index metrics, primarily from the OECD, World Bank, and IMF, and are shown purely as context. They are un-scored: international benchmarks do not affect any domain score, the national score, or risk bands. They are not part of the British Resilience Index score. Each benchmark carries a confidence level (high / med) and a source link, and is labelled “context only, not part of the score.”

Where a domain has a reliable higher-frequency series, its page also shows a Faster signal · a monthly or quarterly indicator (for example CPIH inflation, job vacancies, public-sector borrowing, or mortgage approvals) showing which way the domain has moved since the last annual scored release. These are sourced from the ONS, Bank of England and others and, like the benchmarks, are un-scored context that never affects the index score.

The following improvements are planned. They are listed in rough priority order.

• Migrate the remaining bounds-based metrics to historical-percentile normalisation as verified historical series accumulate.
• Regional and sub-national breakdowns for all eleven domains.
• Live data ingestion pipeline with automated source refresh.
• Formal domain-weight elicitation exercise.
• Independent methodology review by a panel of statisticians.
• Confidence intervals on domain and national scores, not just point estimates.
• Machine-readable API for the full index and metric time series.

Further technical documentation is available in the repository’s methodology/ directory. Source ingestion specifications are in CONTRACTS.md.