How to Do Budget Variance Analysis Without Chasing Noise

Budget variance analysis compares actual results with the budget, line by line, to decide what to change. Most monthly variances are noise, and many of the signals that remain measure the plan’s error before the team’s. The order that works is to test each variance for a signal, correct the forecast, and only then judge performance.

The practice is a century old: in 1922 James O. McKinsey wrote that “After estimates are made, it is necessary that means be provided for checking their accuracy” (McKinsey, 1922). Yet Benchmarkit’s 2026 survey found that fewer than 20% of B2B SaaS companies formally track actual against planned growth with root-cause analysis (Benchmarkit, 2026). A report of actual minus budget, with anything over 10% flagged, cannot tell a random month from a real change, or a bad plan from a bad quarter. Walter Shewhart addressed the first problem at Bell Labs in a one-page memo dated 16 May 1924 (Control chart, Wikipedia); management accountants later named the second.

The parts of a budget variance, and who answers for each

Measure	Formula	What it tells you	Who answers for it
Variance	Actual − budget	How far the month landed from the plan	Nobody, until it passes the signal test
Percentage variance	(Actual − budget) ÷ budget	Size relative to the line	Nobody yet
Favourable or unfavourable	Revenue above budget, or cost below it, is favourable	Direction, not merit	Nobody yet
Price (rate) variance	(Actual price − budgeted price) × actual volume	Paid more or less per unit	Whoever buys
Volume variance	(Actual volume − budgeted volume) × budgeted price	Did more or less of the activity	Whoever drives the activity
Planning variance	Revised budget − original budget	How wrong the plan was, given what is now known	Whoever set the plan
Operational variance	Actual − revised budget	Performance against a corrected plan	The line owner
Signal test	A month outside the line’s own limits, or a run on one side of budget	Whether any of this deserves an explanation	Finance, before the meeting

Source: AFP glossary (price and volume formulas, sign convention); ACCA, Cordwell (2009) (planning and operational variances); Shewhart (1931) and the Western Electric rules (signal tests).

Budget variance analysis starts by separating signal from noise

Frederick Taylor wrote the rule for management reports in Shop Management, presented to American engineers in 1903 (Management by exception, Wikipedia). The manager should see every exception “to the past averages or to the standards pointed out, both the especially good and especially bad exceptions” (Taylor, 1911 edition). He left open how large an exception must be to mean anything.

Shewhart’s memo answered that, with a chart for telling two kinds of variation apart. Common-cause variation is the noise of a system working as usual. The other kind comes from something that has changed, and Shewhart’s word for it was assignable: a cause that can be named, found and removed (Common cause and special cause, Wikipedia).

For budget variance analysis, the tool is the individuals chart, applied one line at a time. Take the line’s monthly variances for the previous year and average the month-to-month changes (the moving range). The limits sit 2.66 times that average either side of the centre (Shewhart individuals control chart, Wikipedia). Engineers centre them on the process average; centring them on the budget makes the chart a test of the plan as well (our adaptation). A month outside the limits is a signal, and so is a run of months on one side. Everything else is ordinary variation, and explaining it month by month produces stories.

A flat threshold does neither job. Martus Solutions, a vendor of budgeting software for nonprofits, recommends thresholds “such as 10%” (Martus). SEC staff, writing about filed financial statements in Staff Accounting Bulletin 99, found that “exclusive reliance on this or any percentage or numerical threshold has no basis in the accounting literature or the law” (SEC, 1999). For management reports the case rests on arithmetic. A line whose monthly variance has a standard deviation of 12% of budget will breach a 10% threshold in about 40% of months by chance alone (our arithmetic), while a stable line can drift for most of a year without reaching it.

A flat 10% rule flags noise on a volatile line and misses drift on a stable one

Three lines from the monthly report of an invented Series A software company show the difference.

Line	Monthly budget	Months flagged by a flat ±10% rule	Control limits around budget	Chart signals this year	Reading
Paid acquisition	$60k	6 of 12	±31.0%	None	Noise: the year ended 2.0% over budget
Payroll	$400k	0 of 12	±2.8%	8 months in a row over budget (months 5 to 12)	Drift: $56k over in 8 months
Software subscriptions	$25k	1 of 12	±8.9%	Month 8, 58% over	Timing: an annual renewal billed in one month

Source: invented monthly series for a hypothetical company, computed by us. Limits: ±2.66 × the average moving range of the prior year’s monthly variances (Shewhart individuals chart), centred on the budget. Run rule: 8 consecutive points on one side (Western Electric, 1956).

On this report the flat rule asks for six explanations of advertising noise and none for payroll, where 8 straight months over budget have added $56k. The chart reverses the workload: one question about the renewal, one about payroll and none about ads. With normal, independent data, all four Western Electric rules together fire by chance about once in 91.75 points (Western Electric rules, Wikipedia). That is 2 or 3 false alarms a year on a 20-line report, and more on skewed or seasonal lines.

Acting on noise makes next month’s variance wider

W. Edwards Deming’s funnel experiment shows what responding to noise costs. A marble is dropped through a funnel held over a target. Under Rule 1 the funnel stays put. Under Rule 2 it moves after every drop by the opposite of the last miss, as a finance team does when it offsets one month’s overspend with the next month’s cut. In Bill McNeese’s account for BPI Consulting, “Using Rule 2 actually increases the variance of the process by a factor of 2 over Rule 1”, and the scatter is about 41% wider (SPC for Excel, 2006).

The mathematics is short. If each month’s result is the new random error minus the last one, its variance doubles and its standard deviation grows 1.41 times; our simulation of 200,000 drops gives a ratio of 2.004. Deming set out the rules in Out of the Crisis in 1986 (Advantive), and his list of Rule 2 cases, as McNeese summarises it, includes “Using variances to set budgets”. The doubling holds for a budget line on two conditions: its monthly errors are independent, and each offsetting cut lands in full the next month. Where both hold, a $60k line with a monthly standard deviation of 12% sees its typical swing grow from $7.2k to about $10.2k, and a cash buffer sized for a bad month must grow by the same 41%.

FloQast, a vendor of close software, teaches the reflex: in its example a company overspends January’s $10,000 marketing budget by $2,000 and trims February’s ads “to compensate for the overspending” (FloQast). That overspend had a cause; in a month with none, the same cut is Rule 2.

A run on one side of the budget indicts the plan

Leaving single months alone does not mean ignoring the run. The Western Electric rules, codified in the company’s 1956 quality-control handbook, treat it as a signal when “Eight consecutive points fall on the same side of the centerline”. Lloyd Nelson’s 1984 version uses nine, a pattern the standard summary of his rules calls “prolonged bias” (Nelson rules, Wikipedia). On a chart centred on the budget, prolonged bias has an obvious owner: the budget.

The arithmetic explains why the rules wait for eight. If the plan is unbiased, each month lands over or under it with even odds, so three months in a row on the same side happen one time in four, and eight in a row about once in 128 (our arithmetic). Treating three months running as a trend would flag a quarter of all budget lines at any moment, by chance.

Software growth plans show such a run. In Benchmarkit’s self-reported survey, B2B SaaS companies planned growth above what they had just achieved for four years in a row: a 26% median plan for 2026 against 20% achieved in 2025 (Benchmarkit, 2026). Four years alone would occur by chance one time in eight (our arithmetic), so the size and the sign of the gap carry as much weight as its length. Benchmarkit calls the pattern “a measurement and accountability gap”. When revenue runs under budget month after month, the first question is about the forecast.

Planning variances belong to whoever set the plan

Management accountants make that question explicit. Geoff Cordwell, ACCA’s examiner for performance management, sets out the split: “Planning variances are measured by comparing original budget figures with revised figures”, and “Operational variances compare the differences between actual and revised budget figures” (Cordwell, ACCA, 2009). The revised budget is the one the company would write with what it knows now, and performance is judged on operational variances alone. In Theresa Libby and Murray Lindsay’s survey, 25.7% of Canadian and 32.2% of US units with a strong budget emphasis already adjust year-end targets for uncontrollable variances (self-reported, with response rates of 13.6% and 1.5%) (Libby & Lindsay, 2010).

One quarter’s variances, split between the plan and the team ($ thousands)

The same hypothetical company planned to start 4 account executives on 1 January, each booking $75k of new ARR a month at a loaded cost of $20k a month. Two of them started on 1 March. Its cloud provider also raised list prices by 10%.

Line	Budget	Revised budget	Actual	Total variance	Planning variance	Operational variance	Who explains it
New ARR booked	900	600	680	−220 (U)	−300 (U)	+80 (F)	The plan owner, for the start dates; sales beat the corrected plan
Account-executive payroll	240	160	165	−75 (F)	−80 (F)	+5 (U)	The same hiring delay, seen as a cost
Cloud hosting	90	99	112	+22 (U)	+9 (U)	+13 (U)	Engineering, for usage growth

Source: our calculation for the hypothetical company. Planning variance = revised budget − original budget; operational variance = actual − revised budget (Cordwell, 2009). F = favourable, U = unfavourable.

Read raw, the quarter says sales missed by $220k, payroll saved $75k and hosting overspent by $22k. Split, it says the plan overstated bookings by $300k because it counted two salespeople for two months before they arrived. The sales team beat the corrected plan by 13.3%, and $13k of the hosting overrun belongs to engineering. The raw reading blames the team that outperformed.

That moves the argument to who writes the revised budget. Cordwell names the risk: “it is a common manipulation to leave the revised price at a level whereby favourable operational variances can still be achieved”. Finance should write it from evidence dated before the results; the person whose operational variance it defines should not.

Grading people on raw variances teaches them to manage the variance

Raw variances do a second kind of damage once they grade people, because the measure starts to move. Charles Goodhart put the idea more exactly in 1975 than the popular version does: “Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes”. The familiar line, “When a measure becomes a target, it ceases to be a good measure”, is the anthropologist Marilyn Strathern’s 1997 phrasing, after Keith Hoskin (Goodhart’s law, Wikipedia).

In budgets, two mechanisms do the collapsing. The first is the ratchet. Martin Weitzman called the use of current performance as a partial basis for future targets “an almost universal feature of economic planning” (Weitzman, 1980). Jan Bouwens and Peter Kroos found the response in a retail chain. Store managers with favourable sales in the first three quarters cut their sales activity in the fourth, and were then more likely to beat the next year’s target, set from current sales (Bouwens & Kroos, 2011).

The second is the expiring budget. Jeffrey Liebman and Neale Mahoney studied 14.6 million US federal purchases worth $2.6 trillion: an even pace would put 1.9% of spending in each week; 8.7% fell in the last week of the fiscal year. IT projects begun that week had “2.2 to 5.6 times higher odds of having a lower quality score” (Liebman & Mahoney, 2017). At the Justice Department, which could roll unused IT funds forward, 3.4% of IT spending fell in the last week, against 9.3% of its other spending.

Neither study looked at a startup. The mechanisms transfer because they need only two conditions, both of which a startup’s finance team can create: next period’s budget set from this period’s result, and unused money that disappears at a fixed date. A favourable variance in March that lowers a team’s April budget is the first; a November underspend clawed back in January is the second. The report then rewards landing on the number, and a spending line that hits its budget every month deserves the same question as one that misses it.

Should a startup wait eight months to confirm a run?

The case against this approach to budget variance analysis is speed. Fincome, a SaaS analytics vendor, argues that “Modern SaaS companies can’t afford to rely on quarterly or even monthly reviews”. It also warns that “Minor discrepancies can snowball into major issues if ignored” (Fincome). A seed-stage company with 14 months of cash cannot spend 8 of them confirming a run.

Fincome is right that late data costs money. McKinsey made the point in 1922: January’s variances are useful on 1 February and “of little value in controlling the budget program for February if not available until February 25”. A startup also changes its process with every hire, price change and funding round, so limits need re-basing after each structural change. And some variances matter whatever a chart says. SAB 99 lists qualitative factors, such as whether a small number “masks a change in earnings or other trends” or affects loan covenants (SEC, 1999). For a startup, the equivalent test is whether a variance moves the date the cash runs out.

Faster reaction to each month, though, is Rule 2 at a higher frequency. A large variance crosses the limits in the month it happens; only a small drift needs a run to show, and that drift is the snowball Fincome fears. The payroll line in the hypothetical report above is one, invisible to a 10% threshold and plain on a chart. Speed belongs in the monthly close and in the materiality test, and patience in the response to noise.

Changes to the monthly report and the board pack

Change	What it prevents
Chart each line against its own history, with limits 2.66 average moving ranges either side of budget, re-based after any structural change	Six explanations of advertising noise a year, and none for a payroll drift of $56k
Explain only signals: a month outside the limits, 8 months in a row on one side, or any variance that moves the cash-out date or a covenant	Stories written for ordinary variation
Offset last month’s overspend in next month’s budget only once a cause has been found	Deming’s Rule 2, which doubles the variance of the line
Split material variances into planning and operational parts, with finance writing the revised budget before the results are known	Charging the plan’s errors to the team that carried it out
Let teams carry underspends forward for a short period	A year-end surge, like the 8.7% of federal spending that falls in the last week
Keep the original budget, the revised budget and the actuals side by side, for instance in the annual budget template and variance tracker below	Losing the record of how wrong the plan was
In the board pack, list the lines with signals and the open runs	A board agenda set by whichever lines crossed 10%

A monthly variance report mixes noise, the plan’s error and the team’s performance. Done as one number, budget variance analysis chases the noise and charges the plan’s mistakes to the people who carried them out. Done in order, signal first, then the plan, then the team, it asks fewer questions and turns each report into a correction to the forecast. The order needs history, though. The limits come from the previous 12 months and must be re-based after each structural change, so a company in its first year of monthly reporting has to lean on the cash-out test until its charts have data.

Sources

• Advantive. Statistical Process Control 101: The Problem with Tampering.
• Association for Financial Professionals. Variance Analysis (glossary).
• Benchmarkit (2026). 2026 B2B SaaS and AI-Native Performance Benchmarks.
• Bouwens, J. and Kroos, P. (2011). Target ratcheting and effort reduction. Journal of Accounting and Economics 51(1–2), 171–185.
• Cordwell, G. (2009). Measuring planning variances. Student Accountant, ACCA, April.
• Fincome. Mastering budget variance analysis: a complete guide.
• FloQast. Budget Variance Analysis: Basics to Expert Insights.
• Libby, T. and Lindsay, R. M. (2010). Beyond budgeting or budgeting reconsidered? A survey of North-American budgeting practice. Management Accounting Research 21(1), 56–75.
• Liebman, J. B. and Mahoney, N. (2017). Do expiring budgets lead to wasteful year-end spending? Evidence from federal procurement. American Economic Review 107(11), 3510–3549; NBER Working Paper 19481.
• Martus Solutions. How to Perform Budget Variance Analysis (Step-by-Step Guide).
• McKinsey, J. O. (1922). Budgetary Control. New York: The Ronald Press Company.
• McNeese, W. (2006). Over-controlling a process: the funnel experiment. SPC for Excel, BPI Consulting.
• Securities and Exchange Commission (1999). Staff Accounting Bulletin No. 99: Materiality.
• Shewhart, W. A. (1931). Economic Control of Quality of Manufactured Product. New York: D. Van Nostrand.
• Taylor, F. W. (1911). Shop Management. First presented to the American Society of Mechanical Engineers in 1903.
• Weitzman, M. L. (1980). The ratchet principle and performance incentives. Bell Journal of Economics 11(1), 302–308.
• Wikipedia. Control chart; Common cause and special cause (statistics); Goodhart’s law; Management by exception; Nelson rules; Shewhart individuals control chart; Western Electric rules.

The figures come from the papers, regulators, surveys and practitioners listed above; the Benchmarkit and Libby and Lindsay numbers are self-reported survey data, and the vendor guides are cited only for what they recommend. Unattributed figures, the funnel simulation among them, are computed from the cited data. The company in the two worked tables is hypothetical, with monthly series invented to show the method, and SAB 99, which governs financial statements filed with the SEC, is used here only for its reasoning about thresholds.