Budget vs Actual: Does the Variance Measure Performance or the Budget?
The season has closed. The full-year budget-versus-actual report is on your desk, and one line stands out: food and beverage cost came in well above budget. The first instinct is to question the operation — who overspent, which department slipped, who writes the explanation.
That instinct hides an assumption about measurement: that the variance measures performance. A variance is a difference score. It is manufactured by subtracting one number from another, and of those two numbers one is the actual the operation produced and the other is the budget you wrote twelve months ago. Charging the whole difference to one side of that subtraction is a move with a name in the measurement literature, and the name is not a flattering one.
Management accounting calls the exercise variance analysis; on property it is usually just the budget vs actual variance report. This piece treats it as a measurement problem rather than a decision problem. The order of the questions matters: before "what caused this variance?" comes "what is this number measuring?" What follows covers the measurement properties of a difference score, whether budget accuracy can stand as an indicator in its own right, and the — genuinely unresolved — literature on what budget-based evaluation does to behaviour.
A variance is the difference between two things — and one of them is the budget
The arithmetic is not in dispute: variance = actual − budget. What is in dispute is which quantity that difference is taken to indicate.
Difference scores have a long track record in psychometrics, and it is not a good one. Cronbach and Furby (1970) showed that raw gain scores — formed by subtracting a pretest from a posttest — lead to fallacious conclusions, principally because such scores are systematically related to any random error of measurement. Having set out better ways to estimate change, the paper concludes that gain scores are rarely useful however they are adjusted or refined, and the authors state explicitly that the argument applies not only to change over time but to other differences between two variables.
Mark the limit of the transfer before leaning on it. Cronbach and Furby are writing about test scores, not ledger lines, and a budgeted figure and an actual figure do not have the same error structure — the actual is auditable, the budget is a bundle of assumptions. What carries across is not a number but a logic: a difference inherits the uncertainty of both terms that produced it. A small variance can be two large errors cancelling; a large one can be a single wrong assumption in a single term.
Its sibling piece takes the diagnostic route: the distinction between bad performance and a badly set target walks through which cause is live and what to do about it. The question here sits one step earlier — what quantity is the figure in front of you an indicator of?
What the variance measures: performance, or accuracy
The same variance figure has two candidate objects of measurement.
The first is operational performance: treat the budget as correct and read the difference as how close the operation came to it. The second is budget accuracy: treat the actual as correct and read the same difference as how well the budget was built. Both readings come out of one number, they run in opposite directions, and they land on different desks.
An indicator cannot be interpreted until the distance between its label and the concept it claims to represent has been written down — that is what construct validity asks of a scorecard line, and a variance column is no different. In most properties the column heading is simply "Variance." Because the heading declines to say which concept it stands for, the interpretation is settled by whoever speaks loudest in the review meeting.
The practical test is ownership: is the person who wrote the budget the person who produced the actual? In hospitality they frequently are — the department head builds next year's expense budget and then spends it. Both terms of the difference then come from the same source. In measurement language, you are subtracting two non-independent observations and reporting the result as an independent measure of performance.
Can budget accuracy be treated as an indicator in its own right
It can, on three conditions, and all three are routinely broken.
First, do not net the signs. The net sum of line-level variances is not an accuracy measure. A year that lands five percent light on revenue and five percent heavy on cost can look immaculate in the net. Accuracy is the sum of absolute deviations; the net is the profit effect. Two quantities, and they cannot share a column.
Second, definitions have to hold still across periods. If the content of a line changed, two years of variance are not comparable — and this is not a hypothetical risk. The lodging industry's reporting standard arrived in a 12th Revised Edition published by HFTP with an adoption date of 1 January 2026, in which the Utilities Schedule was replaced by a new Energy, Water and Waste (EWW) Schedule, with new schedules introduced for Payroll Full-Time Equivalents (FTE) and for Annual Mandatory Brand and Operator Costs. When the label stays recognisable while the scope moves underneath it, the difference between this year's line and last year's measures the definition change, not the operation.
Third, do not wire accuracy to a reward. The moment budget accuracy is scored, the cheapest route to a better score is an easier budget. That mechanism gets its own section below.
Separating systematic bias from random variation
A single period's variance cannot make this separation. It is not a matter of effort — you have one observation.
The separation needs a time series of the signed variance. If a line's variance oscillates around zero across years, what you are holding is noise, and explaining one year of it is assigning meaning to noise. If it falls the same way year after year, what you are holding is bias, and the explanation belongs in the assumption that generated the budget rather than in the operation that met it. Distinguishing a small effect with few observations is the subject of the statistical power literature (Cohen, 1992); a three-year series is not a strong test, but a one-year series is not a test at all.
Then there is structural asymmetry that looks like bias. Anderson, Banker and Janakiraman (2003) found, across 7,629 firms over 20 years, that selling, general and administrative costs rise on average 0.55% per 1% increase in sales but fall only 0.35% per 1% decrease, and they attribute this stickiness to managers deliberately adjusting the resources committed to activities. That evidence is cross-industry SG&A, not hospitality — it is not a hotel benchmark. The logic still applies: if your budget assumes cost moves symmetrically with revenue, then in a soft year part of the unfavourable cost variance is produced by the model, not by the team.
Directional bias: the budget that misses the same way every year
When a line misses in the same direction three years running, that is forecast bias, and the informative content of the line has changed sides. It no longer describes the operation; it describes whoever wrote the budget.
A hypothetical illustration (the figures are invented and are not a benchmark): at a seasonal resort, food and beverage cost closes 6%, then 7%, then 5% above budget across three years — three different executive chefs, three seasons, three supplier sets. That hypothetical pattern holds almost no information about performance. It says the per-guest consumption assumption inside the budget is set too low.
On the revenue side the most common source of the same bias is segment mix. Kimes (1989), setting out the logic of perishable inventory and the allocation of fixed capacity across fare and segment classes, names lodging among the industries that had adopted the approach. If the budget assumed one mix and the operation sold another, the rooms revenue variance measures that assumption rather than commercial performance. Where the segment assumptions behind a revenue budget are written out line by line, the distinction can be made; against a single total revenue line, it cannot.
The behavioural consequences of budget-based evaluation
The measurement properties of a variance are not independent of how the variance is used.
Hopwood (1972), in his empirical study of the role of accounting data in performance evaluation, found that a rigid, budget-constrained evaluative style went together with higher job-related tension and poorer social relations. The three-style taxonomy itself — budget-constrained, profit-conscious, non-accounting — is attributed to Hopwood's 1973 book; the 1972 paper is the empirical work.
The point that matters for measurement: evaluative style changes the process that generates the variance. A manager evaluated in a budget-constrained style manages both the operation and the record in order to shrink the number. The same figure, under two evaluative styles, is the output of two different data-generating processes — which is what the behavioural effect of setting a target looks like once it reaches the variance report.
The finding that would not replicate: context conditionality
Now the honest part of this literature.
Otley (1978) repeated Hopwood's study in a different organisation, one with a well-designed budgetary control system, and did not support Hopwood's findings. More than that: contrary to Hopwood's hypothesis, Otley found a budget-constrained style associated with higher managerial performance than the other evaluative styles. He attributed the divergence to organisational context and concluded that Hopwood's findings are conditional on it.
The contradiction is reported in detail in Noeverman's doctoral thesis (2008, Erasmus University Rotterdam), which also notes that a measurement artefact is an unlikely explanation: Otley's instrument was a carefully revised adaptation of Hopwood's, so both studies were most likely capturing the same construct.
So "tight budget-based evaluation hurts performance" is not a settled finding. What the variance report does to behaviour in your property depends on your context — team structure, season length, data quality — and only your own data will show it. That unresolved contradiction inside a single literature also sets a ceiling on how much interpretation one variance figure can be asked to carry.
The easy ways to manufacture a variance: leaving slack and shifting timing
If an indicator can move while the concept it measures stands still, that indicator has a surface exposed to manipulation. A budget variance has two.
Jensen (2003) analyses the counterproductive effects of using budgets and targets inside performance measurement and compensation systems: paying people according to how their performance relates to a budget makes them game the system, with both superiors and subordinates lying in the formulation of budgets and then gaming the realisation of them. His remedy is not to dismantle budgeting but to change how people are paid — taking the kinks, discontinuities and non-linearities out of the pay-for-performance profile. The early work setting out hypotheses about the propensity to create budgetary slack is Merchant (1985).
Read both mechanisms as validity threats rather than as moral failures. They reach the report in more than one shape:
| The variance you report | What the same number may also contain |
|---|---|
| A cost line closing under budget | Slack left in during formulation |
| A variance that improves sharply at year end | Maintenance or purchasing pushed past the cut-off |
| A variance that stays in a narrow band for three years | A target kept inside the band, rather than an operation kept on plan |
| A net variance very close to zero | Two large line variances cancelling each other |
| Revenue favourable, mix unfavourable | A segment mix other than the one budgeted |
The right column is what an explanation written without testing the left column's validity is likely to miss. It is a measurement check, not a diagnosis.
Single, multiple, composite — the limits of all three
There are three ways to build a variance report, and each has a documented weakness.
Ridgway (1956), in his short paper on the dysfunctional consequences of performance measurement, criticises all three: single measures drive wasteful behaviour, multiple measures create contradictory trade-offs, and composite measures produce tension along with role and value conflict. Reading him as an author who recommends indicator sets means reading only the first third of the paper.
The variance-report equivalents are clear enough. A net total variance is a single measure and conceals everything inside it. A line-by-line list is a multiple measure and leaves the contradictory trade-offs to the reader. A weighted "budget compliance score" is a composite measure and, when it falls, does not say which component fell.
Kaplan and Norton's (1992) scorecard proposal treats measures as an expression of strategy across four perspectives rather than as a list. That does not dissolve Ridgway's objection; adding measures is another trade-off, not an escape from one. The rule that survives is the same whichever form you choose: the components must stay openable.
The materiality test: which variance is worth investigating
Not every variance earns an investigation, and the requirement to write an explanation against every one of them is the most common practice that degrades explanation quality.
Financial reporting has already built a language for this question. IFRS Practice Statement 2, issued in September 2017 and available for application from 14 September 2017, sets out the general characteristics of materiality and offers a four-step process companies may follow in making materiality judgements. It is non-mandatory — it adds no requirement to IFRS Standards, and applying it is not a condition of stating compliance — and its scope is general purpose financial statements, not internal budget reporting. What transfers is not the procedure but the test: information is material if omitting or misstating it could change a user's decision.
In internal variance triage that reduces to three criteria: the size of the variance, the existence of a decision attached to it, and the persistence of the pattern. The third is the one usually skipped. A small variance falling the same way for the third year carries more information than a large one-off.
On thresholds: there is no verified source for "investigate variances above X percent" that could honestly be presented as an industry standard. The threshold is your policy decision. It needs to be written down; it does not need to claim to be universal.
The evidence standard for a variance explanation
A variance explanation is a claim, and claims answer to an evidence standard. Three lines are enough.
Mechanism. Which concrete event produced this difference? "Seasonality" and "market conditions" are not mechanisms; a statement that can explain any outcome explains none of them.
Observable trace. If that mechanism really was live, where else would it show? Higher per-guest consumption should leave a trace in purchase volumes; a shift in occupancy should leave one in the segment report. No trace means the explanation rests on a single source.
Falsifier. What would we see if this explanation were wrong? An explanation with no answer to that is a sentence typed into a report, not a finding.
One check comes before all three: was the content of the line the same in both periods? The chart-of-accounts revision above shows how readily one label can carry a different scope. If the definition moved, the variance is not a performance finding but a break in comparability, and it should be reported as one.
Revising a budget is not the same as lowering a target
The question of when to revise a budget cannot be answered without separating these two operations, and confusing them is expensive.
Revision updates the measuring instrument when an input assumption has been shown to be false: the energy unit price moved, the season lost two weeks, a property closed. Lowering the target holds the assumptions and reduces the difficulty.
The behavioural cost of the difference has been measured. Locke and Latham (2002) report a positive, linear function between goal difficulty and performance, with performance levelling off or decreasing only when the limits of ability were reached or when commitment to a highly difficult goal lapsed. The meta-analytic effect sizes carried in that review — d = .52 to .82 for goal difficulty, and d = .42 to .80 for specific difficult goals over "do your best" — are attributed there to Locke and Latham (1990); the 2002 paper is a review, not the meta-analysis itself.
The working rule comes to three lines: write down which assumption changed, hold the difficulty level, mark the break in the series. Unmarked, a revision and a lowered target look identical in the file — and next year nobody knows which definition produced the series.
Carrying what the variance taught into the next budget
The output of budget vs actual variance analysis is not a paragraph of explanation. It is three records that feed next year's budget.
Signed variance history. By line, at least three periods, with the sign kept. Only that series exposes bias.
Assumption log. Under every line, the assumption it was budgeted on: per-guest consumption, unit price, season length, segment mix. If a variance cannot be traced to an assumption, there is nothing to learn from it.
Definition log. When the scope of a line changes, the date and the reason. It is the only thing protecting comparability.
All three depend on the same precondition — comparing budget and actual in the same structure. When the budget sits in one file, the actual in another system and the definitions in a third place, the variance gets rebuilt by hand every year, and its interpretation stays tied to whoever rebuilt it.
Conclusion
A variance figure is the difference between two terms, and deciding which term to hold responsible is a measurement decision, not an arithmetic one. Explanations written before that decision look as though they describe the operation. Often they describe the budget.
The literature contributes two sentences. A difference score carries the uncertainty of both terms that produced it, so a single period's variance does not separate bias from noise. And the behavioural consequences of budget-based evaluation — as the failed replication between Hopwood and Otley demonstrates — are conditional on context, which means an effect measured in one organisation cannot be carried into your report as a fixed law.
What has to change in practice is small. Write down which concept the variance column represents. Keep the signed variance for three periods. Put the assumption under every line and the date beside every definition change. None of that is a research project; it is record discipline, and it can be in place before the next budget season starts.
The most useful exercise for this year: take your three largest variances and answer one question for each — is this number telling you about the operation, or about the budget? Getting "the budget" three times out of three is common, and it is not bad news. The side you wrote is the side you can fix.
If an outside reading of your variance report, your line definitions and your assumption records would be useful, the HRMIM team works regularly with the measurement and reporting structures of hotel groups. To talk through a single line in your own report, get in touch.


