20% → 25%
The difference is 5 percentage points; the relative change in the rate is 25%. These expressions answer different questions.
DATA / CALCULATIONS
Before writing “a 5% increase”, check what changed: a quantity, a rate or the composition of a population. An arithmetic result alone does not establish that two series are comparable.

AI-generated illustration
These four tools separate operations that are often confused. Enter non-negative numbers up to 1 000 000 000 000, with no more than eight decimal places. Either a comma or a dot can be the decimal separator; spaces can group thousands. Prefilled values are calculation examples, not study data. Counts for the missing-results calculation must be whole numbers.
Compare values in the same unit, with compatible coverage and periods. Absolute difference keeps the original unit; relative change uses the initial value as its base.
difference = final − initial; change (%) = 100 × (final − initial) / initial; index = 100 × final / initial
Enter rates on a 0–100 scale: use 20 for 20%. Subtracting gives percentage points. Relative change expresses a proportional change in the rate.
gap (points) = final rate − initial rate; relative change (%) = 100 × (final rate − initial rate) / initial rate
Use whole counts: observed cases and the corresponding population. Combine only disjoint groups with the same definition and compatible periods. A person present in both groups would be counted twice.
combined rate (%) = 100 × (cases A + cases B) / (population A + population B)
For a binary criterion fixed in advance, separate attempts, readable results and positive results among them. The calculation gives arithmetic bounds if every unreadable result could be negative or positive. It does not replace missing checks.
U = N − R ; min = C / N × 100 ; max = (C + U) / N × 100
Two fictitious batches of 100 records each are split into two disjoint profiles. The same binary criterion is assessed: in profile 1, A has 90% positive cases and B 80%; in profile 2, A has 40% and B 30%. These within-profile rates stay fixed in the experiment.
Fictitious teaching example, with no field data or client result. Counts are constructed to illustrate weighting.
Change the share of profile 1 in each batch; the remainder belongs to profile 2. Each batch keeps 100 records and neither profile is empty.
On a small screen, scroll the chart and table horizontally to see all values.
| Profile | Positive cases / count — A | A rate (%) | Positive cases / count — B | B rate (%) |
|---|---|---|---|---|
| Profile 1 | 9 / 10 | 90 % | 72 / 90 | 80 % |
| Profile 2 | 36 / 90 | 40 % | 3 / 10 | 30 % |
| Total | 45 / 100 | 45 % | 75 / 100 | 75 % |
A exceeds B within each profile, yet its overall rate is lower. The profile mix reverses the ranking here: an illustration of Simpson’s paradox.
Overall rate (%) = 100 × (profile 1 cases + profile 2 cases) / (profile 1 count + profile 2 count).
The common 50 / 50 setting gives 65% for A and 55% for B here. This teaching choice is not a universal reference weighting. In a real study, justify the groups, criterion and reference population; this calculation establishes neither causality nor statistical significance.
These fictional data illustrate baselines, signs and pooling limits. Open a case to see its values and calculation; the button loads them into the corresponding form.
Results are rounded to at most eight decimal places. This display does not measure data precision or uncertainty.
Going from 0 to 12 adds 12 units. Relative change and the base-100 index remain undefined: division by a zero baseline cannot yield an interpretable percentage.
Results are rounded to at most eight decimal places. This display does not measure data precision or uncertainty.
Going from 120 to 90 gives −30 units, or −25%. The final index is 75 with an initial base of 100; that is neither a 75% fall nor a loss of 25 units.
Results are rounded to at most eight decimal places. This display does not measure data precision or uncertainty.
From 80 to 100, the increase is 25%. Returning from 100 to 80 is a 20% fall because the denominator becomes 100. The difference remains 20 units in absolute magnitude.
Results are rounded to at most eight decimal places. This display does not measure data precision or uncertainty.
From 0.2% to 0.3%, the gap is 0.1 percentage point and the relative increase is 50%. This arithmetic proves neither causality nor significance; counts and the observation design are still needed.
19 / 1010 × 100 = 1.88118812 %
Results are rounded to at most eight decimal places. This display does not measure data precision or uncertainty.
Two disjoint, compatible groups have 9 cases out of 10 and 10 out of 1,000. Pooling gives 19 out of 1,010, approximately 1.88%. The simple average of 90% and 1%, which is 45.5%, does not describe this population.
Results are rounded to at most eight decimal places. This display does not measure data precision or uncertainty.
The same 100 people are observed on two dates, with 10 and then 20 cases. Separate rates can be calculated, but pooling them as 200 distinct people would count individuals twice. Compatibility is therefore not confirmed here.
12 / 30 × 100 ; (12 + 10) / 30 × 100
These bounds describe only possible assignments of missing results within this programme. They are neither a probabilistic estimate nor a confidence interval. They do not correct classification errors, duplicates, panel selection or dependence between trials.
Results are rounded to at most eight decimal places. This display does not measure data precision or uncertainty.
12 / 20 = 60% describes readable results only. Across 30 attempts, the positive share can range from 40% to about 73.33%, depending on the 10 unclassified results.
0 / 12 × 100 ; (0 + 12) / 12 × 100
These bounds describe only possible assignments of missing results within this programme. They are neither a probabilistic estimate nor a confidence interval. They do not correct classification errors, duplicates, panel selection or dependence between trials.
Results are rounded to at most eight decimal places. This display does not measure data precision or uncertainty.
All 12 results remain unknown: the possible range is 0–100%. The rate among readable results is undefined because its denominator is zero.
3 / 10 × 100 ; (3 + 0) / 10 × 100
These bounds describe only possible assignments of missing results within this programme. They are neither a probabilistic estimate nor a confidence interval. They do not correct classification errors, duplicates, panel selection or dependence between trials.
Results are rounded to at most eight decimal places. This display does not measure data precision or uncertainty.
Both bounds are 30%: no result is missing in this programme. Selection of the 10 attempts and variation across other programmes still need examination.
0 / 10 × 100 ; (0 + 2) / 10 × 100
These bounds describe only possible assignments of missing results within this programme. They are neither a probabilistic estimate nor a confidence interval. They do not correct classification errors, duplicates, panel selection or dependence between trials.
Results are rounded to at most eight decimal places. This display does not measure data precision or uncertainty.
The eight readable results are negative, but two remain unknown. The positive share across all 10 attempts can still range from 0% to 20%.
The difference is 5 percentage points; the relative change in the rate is 25%. These expressions answer different questions.
The increase is 20%, followed by a decrease of 16.66666667%. The second operation uses 120 as its base, so the percentages cannot simply be added.
The rates are 10% and 50%. If the groups can be combined, the result is 46/100 = 46%, not the simple average of 30%.
An observed zero can be a count; an unknown value must not be replaced with zero. Change from a zero base has no defined relative percentage.
Record indicator, unit, population, dates, series code and version. Check definitions before calculating.
Retain numerators, denominators, source file, extraction date and transformations.
Show the formula, result unit, rounding and what the figure does not establish.
A descriptive increase proves neither causation nor statistical significance.
Not to combine successive changes. Multiply the factors: +20% then +10% gives 1.2 × 1.1 = 1.32, or +32%.
Relative change with a negative base can be read in reverse. This calculator covers non-negative quantities; show a difference in units for balances that cross zero.
Only when appropriate weights allow it. For a population rate, use numerators and denominators of comparable, non-overlapping groups.
The calculator performs arithmetic. It does not analyse sampling design, margins of error or confidence intervals.
No. Published rounding can conceal a range of values. Do not infer more precision than the original data support.
DOCUMENTED REFERENCE
COMPARE-7_1_6.DAT contains the NIST outlier demonstration’s 90 values. Ordered data were checked 22 October 2026. This is a documented teaching dataset, not an income survey or site performance measurement.
| Reference | Value or form | What it identifies |
|---|---|---|
| n | 90 | Count all file observations, including repeats. Two values of 322 remain two observations; deduplicating numbers changes the distribution. |
| mean | 576.077778 | Arithmetic mean recalculated as sum divided by 90, displayed to six decimals. Not claimed as a new NIST-published measurement. |
| median | 559.5 | Usual even-sample median averages positions 45 and 46: 559 and 560. A central location, not a fictional person’s income. |
| min | 30 | Observed minimum. Being far below the mean does not automatically make a value erroneous. |
| max | 1441 | Observed maximum, flagged by NIST beyond an inner boxplot fence. Investigate rather than automatically delete. |
| Variable | Y · unit unspecified | File variable is Y with no subject unit specified. Do not relabel values euros, seconds or inhabitants. |
Official demonstration series; Y has no stated unit. Results below are recalculated by this site.
Source accessed .
All 90 values · Quantile algorithm 7 · p = 0.5 · Selected quantile : 559.5 · n = 90
| Retained count | 90 |
|---|---|
| Mean | 576.0778 |
| Usual median | 559.5 |
| Q1 | 436.25 |
| Q3 | 738.5 |
| Interquartile range | 302.25 |
| Selected quantile | 559.5 |
| Observations beyond fences | 1 |
Excluding 1441 is a sensitivity scenario, not source correction. There are no missing values. Usual even-sample median averages the two centre observations; type 1 Q(0.5) may differ. Display rounded to four decimals.
30, 171, 184, 201, 212, 250, 265, 270, 272, 289, 305, 306, 322, 322, 336, 346, 351, 370, 390, 404, 409, 411, 436, 437, 439, 441, 444, 448, 451, 453, 470, 480, 482, 487, 494, 495, 499, 503, 514, 521, 522, 527, 548, 550, 559, 560, 570, 572, 574, 578, 585, 592, 592, 607, 616, 618, 621, 629, 637, 638, 640, 656, 668, 707, 709, 719, 737, 739, 752, 758, 766, 792, 792, 794, 802, 818, 830, 832, 843, 858, 860, 869, 918, 925, 953, 991, 1000, 1005, 1068, 1441Source: NIST, COMPARE-7_1_6.DAT; calculations and chart by this site.
The comparator retains all values and offers exploratory exclusion of 1441. It neither corrects the source nor proves error. Means respond to extremes; medians still depend on data and the even-sample convention.
These are observations, not a dictionary of possible values. Removing repeats changes frequencies and the question; document transformations and counts.
Documentation consulted on .
DOCUMENTED REFERENCE
A total can change because its counting unit differs. PostgreSQL distinguishes input rows, non-null values and distinct values. The explorer shows all three without treating teaching IDs as statistical measurements.
| Reference | Value or form | What it identifies |
|---|---|---|
| COUNT(*) | All input rows | COUNT(*) includes every result row, including outer-join padding; not necessarily original objects. |
| COUNT(x) | Non-NULL x | COUNT(x) excludes NULL expressions, so a right-key count can be lower without losing left rows. |
| COUNT(DISTINCT x) | Distinct non-NULL x | COUNT(DISTINCT x) counts distinct non-null values, but cannot repair sums multiplied by many-to-many matches. |
| SUM(x) | Non-NULL x | SUM ignores NULL, not join repetitions. SUM(DISTINCT x) also removes legitimately equal amounts; it is not key-based deduplication. |
| AVG(x) | Non-NULL x denominator | AVG uses non-null inputs. Replacing missing values by zero changes terms and denominator; publish the observed count. |
| Empty input | COUNT = 0 · SUM / AVG = NULL | Empty input gives COUNT zero and SUM/AVG NULL. COALESCE-to-zero is a presentation decision, not evidence of observations. |
Define grain before and after joining. DISTINCT is not a universal repair: it acts on selected values. Resolve multiple relationships before per-object sums.
Several result rows can carry one ID. Repeated measures can still inflate sums or unintentionally weight averages.
Documentation consulted on .
DOCUMENTED REFERENCE
These unweighted binary formulas use a declared positive class. They can evaluate automatic presence annotations, without treating a brand mention as an accurate citation or useful outcome.
| Reference | Value or form | What it identifies |
|---|---|---|
| Precision | TP / (TP + FP) | Precision asks how many positive decisions have positive references. 1 − precision divides false positives by positive decisions. |
| Recall / TPR | TP / (TP + FN) | Recall asks how many positive references were found. High recall can coexist with many false alerts. |
| Specificity / TNR | TN / (TN + FP) | Specificity asks how many negatives were correctly rejected: recall of the negative class. |
| FPR | FP / (FP + TN) | FPR divides false positives by reference negatives. It is 1 − specificity, not 1 − precision. |
| Accuracy | (TP + TN) / (TP + TN + FP + FN) | Accuracy counts all correct decisions; class composition can hide minority-class errors. |
| F1 | 2TP / (2TP + FP + FN) | F1 is the harmonic mean of precision and recall, not their arithmetic average; it ignores true negatives. |
| Balanced accuracy | (TPR + TNR) / 2 | Balanced accuracy averages both class recalls equally, not by support, and is not a domain cost model. |
Zero denominators make some rates undefined. No predicted positives means undefined precision here, while F1 is zero if false negatives exist. Publish counts and zero-division policy.
No. FP/(FP+TN) and FP/(TP+FP) describe different populations.
Documentation consulted on .
DOCUMENTED REFERENCE
The official four-pair AP example yields ROC AUC 3/4 and AP 5/6. These are calculations on teaching data, not measured agent performance. They do not change when only the decision threshold changes.
| Reference | Value or form | What it identifies |
|---|---|---|
| Four-score example · ROC AUC | 3 / 4 = 0.75 | Two positives and two negatives give four pairs; positives outrank negatives in three, yielding ROC AUC 0.75 with distinct scores. |
| Four-score example · AP | 5 / 6 ≈ 0.833333 | Positives occur at descending ranks 1 and 3, so AP=(1+2/3)/2. Documentation rounds this to 0.83. |
| AP definition | Σ (R_n − R_(n−1)) × P_n | scikit-learn AP weights precision by recall increments; it does not linearly interpolate the PR curve. Trapezoidal PR area may differ. |
| ROC coordinates | x = FPR · y = TPR | ROC plots false-positive rate against recall across thresholds. The marker is one operating point, not a new area. |
| Ranking / threshold | score → decision | Thresholding changes labels without retraining or changing scores; decision metrics still change. |
| Scores / probabilities | ranking ≠ calibration | Strictly increasing score transformations retain order. Ranking does not prove that 0.8 means 80% observed success; calibration is separate. |
Name the metric and convention, preserve labels/scores and compare the same corpus. Four cases cannot estimate deployment performance or choose acceptable error costs.
No. Recall-increment sums and trapezoidal interpolation use different rules; specify implementation and definition.
Documentation consulted on .
DOCUMENTED REFERENCE
Mean document rates and a corpus rate answer different questions. This reader exposes denominators, including a reference verified to be empty.
| Reference | Value or form | What it identifies |
|---|---|---|
| corpus | 100 × ΣE / ΣN | Corpus rate is 100 × sum(E) / sum(N). Errors from a known empty reference stay in the numerator; an undefined local rate does not remove them. |
| document mean | mean(100 × Eᵢ/Nᵢ) | The simple mean weights documents equally. It is calculated here only when both N values are positive; no document is silently ignored. |
| example | 2/101 ≈ 1.980198% · mean = 50.5% | For E/N = 1/1 and 1/100, the corpus rate is 2/101, about 1.980198%. Mean document rate is (100% + 1%)/2 = 50.5%. Both calculations are correct but use different weights. |
| alignment boundary | document pairs preserved | Align within document pairs first. Concatenating pages before alignment can permit cross-boundary matches and change errors. Keep the profile, unit and pairing consistent. |
Original teaching example; no measurement of a real OCR engine.
Integers 0–1,000,000 per field. Zero N means a verified empty reference here, never missing data.
Tools calculate in your browser without transmitting entered text.
Comparison calculated · ΣE = 2 · ΣN = 101
Total N is zero: corpus rate undefined. Errors remain counted.
At least one zero N: mean rate undefined; no row removed.
| Document | E | N | Local rate (%) |
|---|---|---|---|
| 1 | 1 | 1 | 100 |
| 2 | 1 | 100 | 1 |
ΣN > 0 → corpus = 100 × (E₁+E₂)/(N₁+N₂) N₁ > 0 and N₂ > 0 → mean = (100E₁/N₁ + 100E₂/N₂)/2
| (E₁,N₁,E₂,N₂) | Total errors | Total reference units | Corpus rate (%) | Simple mean of both rates (%) |
|---|---|---|---|---|
| 1, 1, 1, 100 | 2 | 101 | 1.980198 | 50.5 |
| 1, 10, 1, 10 | 2 | 20 | 10 | 10 |
| 1, 1, 2, 0 | 3 | 1 | 300 | Undefined |
| 0, 0, 0, 0 | 0 | 0 | Undefined | Undefined |
| 2, 0, 3, 0 | 5 | 0 | Undefined | Undefined |
| 0, 100, 0, 100 | 0 | 200 | 0 | 0 |
| 2, 1, 0, 100 | 2 | 101 | 1.980198 | 100 |
Teaching counts for two documents, not real corpus observations. A missing reference is not a verified empty reference: do not default it to N = 0. A zero total N leaves the corpus rate undefined; errors remain visible.
When all N values are positive, weighting document rates by N equals sum(E)/sum(N). If a zero-length reference has insertions, its local rate is undefined: aggregating counts directly prevents losing those errors.
Documentation consulted on .
DOCUMENTED REFERENCE
These observations compare percentage-point and relative changes in a published series. The three countries are selected for reading, not as a digital-performance ranking.
The indicator measures the share of the population using the Internet in the previous three months. It measures neither connection time, access quality nor AI-assistant use.
IT.NET.USER.ZS · Extracted · Latest update reported by the API .
| Country | 2020 | 2021 | 2022 | 2023 | 2024 |
|---|---|---|---|---|---|
| France | 84.71 | 86.10 | 85.33 | 86.84 | 88.65 |
| Germany | 89.81 | 91.43 | 91.63 | 92.48 | 93.50 |
| Spain | 93.21 | 93.90 | 94.49 | 95.45 | 95.76 |
France: from 84.71% in 2020 to 88.65% in 2024; 3.95 percentage points, or 4.66% relative to the starting value.
Table values are rounded to two decimals; calculations use the received values before that rounding. Examine revisions and collection differences before making a causal explanation.
Source: ITU, distributed by the World Bank. Extract and chart by Alexis Roux ZG. CC BY 4.0
DOCUMENTED REFERENCE
These nine observations cover three selected dates in 2025 and three currencies. Extracted on 22 October 2026, this sample is fixed: it is neither a current quote, an annual average nor every day of 2025.
Rates are currency units per euro. Divide a foreign-currency amount by the rate to obtain euros; multiply in the other direction. This calculation does not adjust inflation or include exchange fees or margins.
Fixed extract, consulted on .
| Observation date | USD/EUR | GBP/EUR | JPY/EUR |
|---|---|---|---|
| 1.0321 | 0.83118 | 162.04 | |
| 1.1419 | 0.8434 | 162.98 | |
| 1.175 | 0.8726 | 184.09 |
Method: M in currency ÷ T in currency/EUR = M/T in EUR. Reverse: M in EUR × T = M×T in currency.
Use a dot or comma as the decimal separator and spaces for thousands. Negative amounts are accepted to examine a correction.
1 000.00 USD gives 968.90 EUR. Observation on 2025-01-02: 1 EUR = 1.0321 USD.
Arithmetic illustration: the entered amount is an exercise, not observed expenditure. Results use two decimal places; calculations use the stored rate. The ECB publishes reference rates for information and discourages their use for transactions.
Frequency D: daily. Three separate dates, no interpolation of missing days. Quote is currency per EUR, with unit multiplier 0 in the responses. All nine values retain source precision; derived calculations are not additional official rates.
Source: European Central Bank. Nine observations reproduced unchanged; conversions and presentation calculated by this site. The data are also freely available from the ECB.
DEEPER READING / PRACTICE
A percentage or rate must be calculated on the basis of public data. The calculation is interpretable only if the units, exclusions and the denominator are documented.
Updated on .
Check that the values are on the same scope and that the denominator corresponds to the question. Difference between zero result, absent value and calculation impossible.
An incorrect denominator can produce a plausible but misleading number. Rounding too early can also change the differences or the ranking observed.
The calculation sheet retains input, formula, unit, exclusions and round values. The commentary distinguishes the calculated result and scope of its interpretation.
PRACTICAL CASE / WORKED EXAMPLE
The pedagogical example concerns the preparation or interpretation of data. It shows why a technically possible calculation should remain linked to verified definitions, versions and transformation rules.
A rate is calculated on the listed lines, but the report presents it as a result for the entire population. The absent cases are not described.
Consider the observations available and missing. Check the denominator rule, keep the exclusions and test the calculation on a simple case. Examine the effect of coverage on the conclusion.
The result indicates its actual population and its limits. The absent values remain visible without being arbitrarily converted into zero or in negative cases.
They allow to check transformations and re-calculate a synthesis. Early rounding can change comparisons; the report must specify the rule applied to the final presentation.
DETAILED WORKSHOP / APPLICATION
Prepare values, units, formula and population. Identify the values absent, exclusions and conditions that make the denominator relevant to the question.
Make a simple calculation and then compare it to the expected result. Apply the rules to the body by controlling effective and round. Preserving important intermediate values.
If calculation is impossible for some cases, produce a separate status. A absence of result should not be converted to zero only to facilitate presentation.
The result is validated when formula, entrances and scope can be found. The interpretation is examined separately from the accuracy of the digital operation.
Transmit calculation sheet, versions and comments. Exclusions and rules changes remain visible to understand the differences between several results.
ANNOTATED WORKING DOCUMENT / REVIEW GRID
The file retains values, formula, units, version and round rule. It indicates the actual denominator and the cases excluded to make the result reproducible.
On a small screen, scroll the grid horizontally to read all three columns.
| Observation | Supported interpretation | Next step |
|---|---|---|
| The entries and the formula correspond to the question. | The calculation can be validated in the described scope. | Compare a simple case and keep the intermediate values. |
| The rate applies only to the listed lines. | The result does not automatically describe the entire initial population. | Presentation of available facts and absences with the conclusion. |
| Null and missing values are confused. | The treatment rule can significantly change the result. | Repeat the methods and then make the dependent calculations. |
The final verification must separate the accuracy of the operation and the relevance of its interpretation. A correct formula applied to an inappropriate denominator produces a result that can remain misleading.
REPORTING / EXAMPLE WORDING
A digital result must announce the observations actually used.
Separate validation of the operation and scope of its interpretation.
The calculation is based on the indicated values and exclusions. The effective population, the formula and the round rule are preserved with the result.