Reader question: Thousands of overnight money-market transactions can occur at different rates and in different sizes. How does a benchmark administrator compress that distribution into one published interest rate without allowing a handful of unusual trades to dominate the answer?
The short answer is that benchmark construction is a problem in robust statistics plus data governance. The administrator first defines the market it intends to measure, filters transactions to an eligible set, validates the data, applies a specified aggregation rule, checks data sufficiency, and publishes the result under documented correction and contingency policies.
Three important overnight benchmarks make the mathematics unusually visible. The New York Fed’s SOFR uses a volume-weighted median of eligible Treasury-repo transactions. The Bank of England’s SONIA uses the volume-weighted mean of the central 50% of eligible transaction volume. The ECB’s €STR also uses a volume-weighted trimmed mean, removing the lowest and highest 25% of volume before averaging the central 50%.
What this page owns
This page owns the mathematical and algorithmic question: how transaction-level overnight funding data become a benchmark statistic. It is upstream of the floating-rate loan calculation, where a published reference rate is compounded or otherwise applied to a contract. It also connects to yield-curve construction, swap valuation, and cross-currency funding mathematics.
This is educational material, not a forecast of future interest rates and not advice about borrowing, lending, investing or hedging.
Step 1: define the underlying interest
A benchmark cannot be evaluated only by asking whether its arithmetic is correct. First ask: what economic quantity is it trying to measure?
SOFR measures the cost of borrowing cash overnight when the borrowing is collateralised by U.S. Treasury securities. SONIA measures overnight sterling wholesale funding in circumstances where credit, liquidity and other risks are minimal, using eligible unsecured transactions. €STR reflects wholesale euro unsecured overnight borrowing costs of euro-area banks.
These are related but not identical markets. A numerical comparison between the three therefore mixes currency, collateralisation, institutional composition, monetary-policy conditions and market structure. The algorithm begins with a definition before it begins with a formula.
Step 2: construct the eligible transaction set
Write the day’s raw observations as transactions:
(r1, v1), (r2, v2), …, (rn, vn)
where r is the transaction rate and v its volume.
The raw set is not automatically the benchmark set. Eligibility rules decide which instruments, counterparties, maturities, settlement conditions and minimum sizes belong in the population. Data-quality controls then look for malformed, duplicated, implausible or otherwise unusable records.
This is a crucial modelling choice. A flawless median of the wrong population is still the wrong benchmark.
SOFR: find the rate of the middle dollar
The New York Fed states that SOFR is calculated as a volume-weighted median. Conceptually:
- sort eligible transactions by rate;
- accumulate their transaction volumes;
- find the rate at which cumulative volume reaches the 50th percentile of total volume.
If total eligible volume is V, the algorithm looks for the rate associated with the “middle dollar” around 0.5V.
This differs from an ordinary median of transaction records. Ten tiny trades should not receive the same total influence as one transaction one hundred times their combined size if the benchmark is explicitly volume-weighted.
A small volume-weighted-median example
Suppose eligible overnight repo transactions are:
| Rate | Volume |
|---|---|
| 4.90% | 10 |
| 4.91% | 20 |
| 4.92% | 50 |
| 4.96% | 20 |
Total volume is 100. The cumulative volume reaches 50 within the 4.92% block, so the volume-weighted median is 4.92%.
Now move the last 20 units from 4.96% to 6.00%. The median is still 4.92%. The extreme tail changed sharply, but the middle of the volume distribution did not. That is the robustness property the median is exploiting.
SONIA and €STR: average only the central volume
SONIA and €STR use a different robust-statistics design. Transactions are ordered by rate, but instead of selecting only the 50th-percentile rate, the algorithm trims the tails in volume terms and calculates a volume-weighted mean of the central part.
For SONIA, the Bank of England describes the result as a trimmed mean based on the central 50% of the volume-weighted distribution. Eligible SONIA transactions are unsecured, one-business-day sterling deposits reported through the Bank’s money-market data collection, executed and settled under specified timing conditions, and at least £25 million in size.
For €STR, the ECB uses eligible unsecured fixed-rate overnight deposit transactions above €1 million. It removes the lowest 25% and highest 25% of transaction volume, then averages the remaining central 50%, applying pro-rata treatment where a transaction-rate bucket crosses a trimming boundary.
The trimmed-mean formula
After trimming, let wi be the retained volume of transaction or rate bucket i. Then the benchmark mean has the familiar form:
r̄ = Σ wiri / Σ wi.
The subtlety is in determining wi. At the 25% and 75% cumulative-volume boundaries, only part of a transaction bucket may belong to the central interval. The algorithm must allocate that boundary volume consistently rather than trimming whole records arbitrarily.
Median versus trimmed mean: neither is universally “better”
A volume-weighted median is highly resistant to extreme rates if those extremes do not move the middle of the volume distribution. It is also easy to explain: locate the rate of the middle unit of volume.
A trimmed mean deliberately uses more information from the centre of the distribution. It can respond to shifts across the central range while suppressing extreme tails. But the output depends on the chosen trimming fraction, and the administrator must periodically test whether that parameter remains appropriate for the underlying market.
So the correct question is not “Which statistic is best?” It is: Which statistic is robust, representative and governable for the market this benchmark claims to measure?
Inputs and outputs
Inputs include transaction rates, volumes, instrument classifications, counterparties, timestamps, settlement dates, maturity, collateral status where relevant, data-source identifiers, eligibility rules, exclusion rules, benchmark parameters and contingency thresholds.
Outputs should include more than the headline rate. Good benchmark transparency can include aggregate volume, percentile information, number or concentration of contributors, publication type, whether standard or contingency methodology was used, and later correction information where applicable.
The ECB, for example, publishes the €STR together with volume, number of active banks, transaction count, selected percentiles and the share of volume accounted for by the five largest active banks. Those accompanying statistics help readers judge the day’s data base rather than treating the single rate as context-free.
Data sufficiency: the calculation can be mathematically defined but statistically weak
Imagine a trimmed-mean algorithm that technically has transactions to process, but nearly all volume comes from one institution. The formula will still return a number. Yet that number may no longer represent a broad market.
This is why benchmark administrators define data-sufficiency and contingency policies. The system needs a rule for what to do when data are missing, contributor concentration becomes excessive, or the ordinary calculation cannot be trusted to represent the intended market.
For €STR, public methodology reviews explicitly test both the 25% trimming parameter and contingency thresholds, including the number of reporting banks and concentration of turnover among the largest contributors.
Data quality comes before robust statistics
A median is resistant to some outliers, but “resistant” does not mean “ignore data quality.” A mistaken transaction that falls near the centre of the volume distribution can matter more than an absurd but tiny tail observation.
Therefore the pipeline requires plausibility checks, format checks, reconciliation of data feeds and controlled expert judgment or exclusion policies where the administrator’s framework permits them.
The weak link can be upstream. If the input transaction population is incomplete or misclassified, no robust statistic can reconstruct observations that never entered the dataset.
Evidence polarity: what supports the benchmark?
Evidence for representativeness includes broad eligible transaction volume, diversified contributor participation, stable data-quality controls, a central distribution consistent with nearby money-market evidence, infrequent reliance on contingency procedures, transparent governance and periodic methodology reviews.
Evidence against confidence includes collapsing transaction volume, extreme contributor concentration, frequent late revisions, repeated contingency activation, structural migration of trading into instruments outside the benchmark scope, unexplained divergence from the market the benchmark claims to measure, or data errors concentrated near the statistic’s influential region.
A counterexample: the mean can move while the median does not
Take the earlier SOFR-style example. If a small tail trade jumps from 4.96% to 20%, an ordinary volume-weighted mean rises. The median can remain unchanged if the cumulative 50% point remains in the same central block.
This does not prove the median is always superior. If genuine market stress begins in the tails and later spreads inward, a median may move later than a full mean. Robustness always trades some sensitivity to extremes for protection against extremes.
Another counterexample: trimming can hide a genuinely bifurcated market
Suppose half the market funds near one rate and half near a very different rate because participants have split into two economically distinct groups. A central trimmed mean can compress that bimodal structure into a number that few institutions actually pay.
The benchmark may still be calculated correctly, but distributional diagnostics become essential. Percentiles, contributor breakdowns and market-structure analysis can reveal when “one number” is hiding important heterogeneity.
Diagnostics for benchmark algorithms
- Volume diagnostic: is eligible volume large enough relative to normal market conditions?
- Contributor diagnostic: how concentrated is the day’s volume?
- Distribution diagnostic: inspect percentiles, tails, gaps and multimodality rather than only the centre.
- Method comparison: compare median, trimmed mean and untrimmed mean as diagnostics even when only one is official.
- Leave-one-contributor-out test: measure how much the result changes if a dominant contributor is excluded.
- Boundary test: verify partial-volume handling at trimming thresholds.
- Feed reconciliation: confirm transaction counts and aggregate volumes across upstream data stages.
- Revision diagnostic: analyse the frequency and magnitude of corrected publications.
- Representativeness check: test whether the eligible market still matches the benchmark’s stated underlying interest.
What would falsify confidence in an implementation?
Confidence should be withdrawn if sorting and cumulative-volume logic reproduce the wrong percentile; if boundary trimming changes when identical transactions are split into smaller records; if independent implementations disagree on the same eligible set; if excluded trades re-enter after aggregation; if a benchmark labelled “standard calculation” actually used contingency data; or if published accompanying statistics cannot reconcile to the transaction set.
SOFR, SONIA and €STR are not literally interchangeable “risk-free rates”
The phrase risk-free rate is convenient market language, not a claim that every underlying transaction is economically free of every risk. SOFR is secured by Treasury collateral; SONIA and €STR are unsecured wholesale funding benchmarks in different currencies and market structures.
Contract conventions, compounding, observation shifts, lookbacks, fallback rules, spread adjustments and curve construction are separate layers. A benchmark algorithm produces the daily reference observation; it does not by itself determine the interest amount on a particular loan or derivative.
How this connects to the surrounding knowledge estate
Once the daily benchmark exists, floating-rate loan algorithms decide which observation dates and compounding rules apply. Yield-curve algorithms transform market instruments into discount and forward structures. Swap valuation then applies those curves to future fixed and floating cash flows. The chain is therefore: transactions → benchmark → conventions → curves → contract valuation.
Verification and update triggers
Preserve the benchmark-methodology version, eligible-market definition, aggregation parameters, exclusion policy, contingency rules, publication calendar and data-source lineage. Reassess the algorithm after an administrator consultation, methodology change, material shift in market structure, persistent volume decline, increased contributor concentration, recurring data-quality incidents, changes in reporting regulation, or repeated use of contingency calculation.
Time-sensitive example: the New York Fed continued to publish SOFR under its documented reference-rate framework in 2026 and issued an IOSCO-compliance statement on 17 July 2026. The Bank of England’s SONIA page was updated on 27 May 2026. The ECB’s public €STR page on 27 August 2026 continued to report the standard calculation with accompanying volume and contributor statistics. These details should be rechecked rather than hard-coded into long-lived software.
Primary and high-quality references
- Federal Reserve Bank of New York, Secured Overnight Financing Rate data and methodology overview.
- Federal Reserve Bank of New York, Additional Information about Reference Rates, including the volume-weighted-median methodology and contingency information.
- Bank of England, SONIA interest rate benchmark and SONIA Key Features and Policies.
- European Central Bank, €STR overview and current publication and the ECB’s published €STR methodology and policies.
- International Organization of Securities Commissions, Principles for Financial Benchmarks, for the governance, methodology, data quality and accountability framework used by major benchmark administrators.
Educational boundary: This article explains benchmark statistics and computational logic. It does not predict interest rates, recommend a financial product, calculate a reader’s borrowing cost or provide personalized financial advice.