Health Scores

Overview Benefits Scoring Scores Across Periods Score Locations Score History & Frequency How to Use Health Scores Health Scores and Predictions Historic Performance

Overview

A Health Score is a numeric value attached to each period that reflects our confidence the data accurately represents real-world behaviour for that window. The higher the score, the higher our confidence. We surface this so you can quickly see when:

we’ve identified a known issue,

a manual intervention occurred, or

patterns suggest the data may be less reliable than usual

Benefits

Make confidence comparable. Past release notes flagged concerns but weren’t numeric or easily comparable across
periods or tickers. Health Scores give you a clear, standardised signal.

Make confidence quant-friendly. Quant teams get the same nuance our notes provide, but now in a form that can be
ingested, filtered, and modelled alongside the data.

Scoring

Each index is assigned a health score between 0–100% using an internal rubric, reflecting our overall confidence in the data for that period. Scores account for data completeness, consistency, and any manual interventions applied.

Score	How to Interpret
100%	100% Data aligns closely with real-world activity. No anomalies or interventions detected.
90-99%	A minor issue was identified and corrected. The result is expected to accurately reflect real-world behaviour.
75-89%	A fix was applied with moderate certainty, or small inconsistencies were observed. Totals should be directionally accurate but may not be exact.
60-74%	Structural or behavioural changes were detected. No manual fix applied. Treat totals as indicative rather than precise.
40-59%	Known shifts in data capture or coverage. Underlying structure may differ significantly from prior periods. Use caution when comparing value
<40%	Severe or uncorrected issues make the dataset unsuitable for quantitative analysis. Avoid drawing conclusions from this period.

Scores Across Periods

Each period inherits the lowest health score of any issue that occurred within that period.

Example

Consider a standard Q1:

January: minor issue (90%)

February: major issue (40%)

March: no issues (100%)

The overall Q1 health score would be 40%, reflecting the lowest confidence period within the
quarter.

Score Locations

Health scores appear anywhere you already consume our index data:

Systematic delivery: In TS_INDEX as the INDEXHEALTH column (per ticker, per period). Any score less than 100% includes a reference to a note giving context.

Excel delivery: Included in the Tabular Data tab, with the same note references for any score less than 100%.

Score History & Frequency

Historical coverage: Available from July 2023 onward, across all frequencies.

Update frequency: Every dataset we release includes a health score. Our analysts assess the score during data preparation, and it is delivered at the same time as the index so you can interpret quality and signal together.

How to Use Health Scores

Treat the score as an additional, quantitative confidence signal. In providing this extra layer of transparency, we believe that it gives you more opportunity to inform trading decisions:

Model weighting: Use the score as a dynamic weighting factor to introduce trust to modelling decisions.

Backtesting: Historical scores (from July 2023 onward) enable you to simulate re-sizing trades around data reliability.

Operational guardrails: Set thresholds to standardise how models respond to known issues.

Health Scores and Predictions

When index health is marked down, we see a higher share of our weaker errors. However, among our most accurate forecasts, this link is weak. We see no significant difference in the share of top predictions across health score groups.

Higher Health Scores Correspond to Lower Normalised Error*

Extremes are likelier when health is lower: Worst-error cases (NE ≥ 2) are more common when HS < 100%.

Top-tier accuracy occurs in both groups: NE < 0.1 is similar across groups.

Why the nuance? We sometimes withhold predictions when health is very poor or when there are major uncaptured business shifts. We also use custom models when relationships shift, and some KPIs are naturally linear, making them easier to predict even if health isn’t perfect. External business dynamics, such as price changes or AOV shifts, can still move KPIs independent of index health.

*What is Normalised Error (NE)? To compare errors across different tickers and KPIs, we scale each absolute error by that KPI’s typical error. This puts diverse KPIs on a comparable scale, so buckets (e.g., NE ≥ 2) indicate how large an error is relative to what’s typical for that KPI.

Historic Performance

Across 4,079 monthly index periods*, 122 were marked down (<100%), a historical mark-down rate of 2.99%. As quarters inherit the lowest score from their component months, this mark-down rate is higher than at a finer granularity.

Distribution by severity (monthly, Jul 2023–May 2025)

Score	Total	Percent of Total	Percent of Total where score <100%
100%	3,957	97.01%	-
90-99%	35	0.86%	28.69%
75-89%	13	0.32%	10.66%
60-74%	45	1.10%	36.89%
40-59%	13	0.32%	10.66%
<40%	16	0.39%	13.11%
Total	4,079	100%	100%

*We assessed all monthly indices from 1 July 2023 to 31 May 2025, with health scores applied in June 2025 (latest period end: 17 June 2025). This snapshot doesn’t include projects launched after that date or any monthly indices released from 1 July 2025 onward.