Variance Analysis and Driver Trees

The established framework for this is the flexible-budget decomposition set out in Horngren, Datar, and Rajan's Cost Accounting: A Managerial Emphasis, the text most analysts learned it from. The discipline is to avoid comparing actual results against a budget built for a different level of activity. You first flex the budget to the volume you actually sold, then read the variances against that flexed benchmark rather than against the original plan.

Under that framework, the gap between the static budget and actual revenue decomposes into a selling-price variance (the part explained by realizing a different price than planned, holding volume at the level actually achieved) and a sales-volume variance (the part explained by selling a different number of units than planned, holding price at budget). When a business sells more than one product, the sales-volume variance splits again into a sales-mix variance, which captures a shift in the proportions of what was sold, and a sales-quantity variance, which captures a change in total units at the budgeted mix. Each piece ties back to the total under one consistent favorable-or-unfavorable convention, which is what makes the decomposition checkable rather than a matter of opinion.

Decomposition is only half the craft. The other half is explaining the drivers in a way a decision-maker can act on, which is the substance of driver-based management reporting. The Institute of Management Accountants, in its Statements on Management Accounting, frames good management reporting as information that is actionable rather than merely accurate: alongside the figures, it should carry the business reasons behind the key variances. The CFA Institute curriculum makes the same move from the analyst's chair, in its treatment of the analysis of operating results, where decomposing a change into its drivers is how you judge whether a period's performance is high quality and likely to persist or a one-off that flatters the headline.

A decision-quality driver explanation tends to have four properties. It decomposes the movement, separating the price, volume, and mix pieces rather than reporting one lump sum. It sizes each piece, so the reader knows whether the story is mostly price or mostly volume. It grounds each named cause in something the data supports rather than a plausible guess. And it carries a decision-quality layer: at least one alternative explanation for the same numbers, so the analysis does not rest on a single confident story. A driver tree, the branching structure that hangs each named cause under the figure it explains, is the reporting form that makes the four visible at once.

Two of those four properties are arithmetic and two are judgment, and that seam is why this task suits an AI-assisted workflow. The decomposition and the sizing are deterministic: for a given set of units and prices under one convention, the price, volume, and mix pieces are fixed. The grounding and the alternatives are reasoning about cause under uncertainty. Part Two puts each kind of work where it belongs.

The most useful thing to hand a model for this task is an example of the destination. A price-volume bridge from a prior quarter, alongside the short driver memo that accompanied it, shows the house style: how each line is decomposed, how much detail a driver gets, and how the memo handles a movement whose cause is uncertain. With that example in the folder, the request shifts from "explain these variances" to "produce more of this known shape," which is a far easier task to get right.

The journey is clear at both ends. You start with a bridge of actual against budget by product line, built on the flexible-budget decomposition from Part One; you are going to a driver-based explanation a decision-maker can act on. Anchoring the far end with a real prior example keeps the model from guessing what a finished driver memo is supposed to contain.

Separating a revenue change into price and volume is deterministic work: for a given set of prior and current units and prices, one convention yields one price figure and one volume figure that sum to the total. Asking a language model to produce those figures in prose invites small, hard-to-catch slips. So the workflow puts the arithmetic in a bridge that computes each line's split once, and reserves the model for the non-deterministic work: reading the decomposition and reasoning about what caused it, which cause is likely to persist, and what a competing explanation would be.

A driver tree is the structure that makes this split visible. It breaks a top-line movement into branches: a price branch, a volume branch, and a mix effect branch that captures how a shift toward higher- or lower-priced lines moves the average even when each line's own price holds steady. Building the tree is arithmetic; naming and testing the driver at the end of each branch is judgment.

Two lines from Meridian's bridge make the split concrete, and they are a clean contrast because each moved for a single reason. Take the arithmetic one line at a time, using the convention that the price variance applies the price change to the actual units and the volume variance applies the unit change to the budget price.

Industrial Fittings, a pure price move. The line sold 40,000 units in both the budget and the actual, so volume did not move. The budgeted price was $205.00 and the realized price was $215.25, a $10.25 increase. The price variance holds units at the actual 40,000 and applies the price change: $10.25 times 40,000 is $410,000 favorable. The volume variance is (40,000 minus 40,000) times the $205.00 budget price, which is zero. Price plus volume is $410,000 plus $0, and that ties to the line's total variance of $410,000. The entire movement on this line is price.

Precision Components, a pure volume move. Here the price held at $200.00 in both the budget and the actual, while units rose from 32,000 to 36,000. The volume variance holds price at the $200.00 budget level and applies the unit change: (36,000 minus 32,000) times $200.00 is $800,000 favorable. The price variance is ($200.00 minus $200.00) times the 36,000 actual units, which is zero. Price plus volume is $0 plus $800,000, tying to the line's $800,000 total. The entire movement on this line is volume.

The two lines carry the same sign and nearly the same size, yet they are opposite stories. The $410,000 is a list-price increase that persists into next quarter until it is reversed; the $800,000 is 4,000 extra units whose recurrence depends entirely on whether that order repeats. A decision-maker who saw only a combined "+$1.21 million favorable" would miss that one half is durable and the other is a question mark. That contrast is the whole reason the split is worth doing, and why the arithmetic belongs in a bridge that runs it once and ties it out, freeing the analyst to reason about which story holds.

The lab folder is deliberately small: the price-volume bridge and the materials cost line, and little else. A full profit-and-loss statement would bury the two or three lines that actually moved and raise the chance the model latches onto something irrelevant. Minimum context is the fuel here: give the model the decomposed bridge so it reasons about cause rather than recomputing the split, and give it the cost line so a revenue beat does not hide a margin problem. The materials line earns its place precisely because Meridian's $500,000 unfavorable cost movement is large enough to reframe an $850,000 revenue beat; leave it out and the model reasons about half the picture and calls the quarter a clean win. The discipline is inclusion by relevance rather than by completeness: each file in the folder should be one the driver question actually turns on, and anything that does not change the answer is noise that can pull a confident model toward the wrong line.

Two pieces of scaffolding turn this from a one-off prompt into something you run each quarter. The first is a reusable skill, or a saved prompt, that encodes the house pattern: read the decomposed bridge, narrate each material line's driver in the finance team's voice, then, for each, propose at least one alternative hypothesis and name the evidence that would confirm or rule it out. Written once and reused, it saves you from re-specifying the decision-quality layer each time and keeps the output shape consistent across quarters.

The second is a short driver checklist the analyst runs before the explanation leaves the building:

• Do the price, volume, and mix pieces sum back to each line's total variance under one convention?
• Does each named driver match the sign and size of the piece it explains?
• Does each material line carry a genuine alternative hypothesis rather than a throwaway second option?
• Is the cost side read alongside the revenue beat, so a margin problem is not hidden behind a favorable top line?

The skill drafts; the checklist is where the analyst takes ownership. Neither one removes the judgment.

Download the folder and run the driver-tree pattern in whatever AI you use. The folder holds Meridian's price-volume bridge for the quarter, with actual against budget by product line and the price and volume effects already separated, plus the materials cost line. Let the bridge hold the decomposition, have the model explain each material line and then generate at least two alternative hypotheses with the evidence that would separate them, and come back for validation. The company and its numbers are fictional.

In a 2023 field experiment, researchers gave 758 consultants a set of tasks to complete with and without AI assistance. On tasks that sat inside the model's competence, AI help raised quality and speed noticeably. On a task deliberately designed to fall just beyond that edge, the pattern reversed: AI-assisted consultants were more likely to reach the wrong answer than those working without the tool, because the model produced fluent, plausible output that was subtly off, and people tended to accept it. The authors named the uneven boundary between what the model does well and what it does poorly the jagged frontier: capability is high in some places and low in others, and the border is hard to see from the inside.

Variance explanation lives near that border. A model can decompose and narrate a clean bridge well, and it can also generate a confident cause for a movement whose real driver is not present in the data. The guard is a decision-quality layer: alongside the primary explanation, require at least one alternative hypothesis that fits the same numbers, and make each hypothesis falsifiable by naming the evidence that would confirm or rule it out. A second hypothesis does not slow the work so much as it changes the question from "does this story sound right" to "what would tell these two stories apart," which is the question a decision-maker actually needs answered.

Concretely, if a line's volume jumped, the primary explanation might be a durable new customer, and the alternative might be a one-time stocking order; the evidence that separates them is the customer's contract structure and repeat-order history. Holding both until the evidence arrives is what keeps a confident story from hardening into a wrong decision. Work the checklist below against your draft before you would call it done.

The explanation below is what you would put in front of the CFO. It leads with the decomposition, sizes each line, separates price from volume, and refuses to let the favorable net stand in for the story.

Revenue. The quarter beat plan by $850,000, and that beat is three separate stories. Industrial Fittings rose $410,000 on price alone, from the list increase, with flat volume; this piece should persist. Precision Components rose $800,000 on volume alone, from one large new order, with flat price; this piece is real but its recurrence is open. Custom Assemblies fell $360,000 on lower volume, partly offsetting the gains. The headline is favorable, and its most durable component is the smallest of the three.

Cost and margin. Materials cost came in $500,000 unfavorable. Against an $850,000 revenue beat that is largely one order plus one price increase, a cost movement of that size is material to the margin read and should not sit outside the revenue story. Whether it reflects higher input prices or a shift in product mix changes what, if anything, to do about it.

The explanation above is the primary read. Decision quality comes from holding it alongside competing explanations for the same numbers, each with a test that would settle it. Two hypotheses matter most here.

Is the Precision Components volume gain a durable win or a one-time stocking order? The primary read treats the $800,000 gain as new business that recurs. The alternative is that the customer placed a single large order to build inventory, and volume reverts next quarter. The evidence that separates them: the order's contract structure (a multi-period agreement or blanket purchase order points to durable; a single one-off purchase order points to stocking), the customer's own sell-through and inventory position, and whether the backlog shows repeat orders behind the first. Until that evidence is in, the gain is reported as real but flagged as unconfirmed for run-rate.

Is the $500,000 materials cost rise driven by input prices or a mix shift? The primary suspicion might be that suppliers raised prices. The alternative is that the cost rose because the quarter produced more of a material-heavy line, the new Precision Components volume, rather than because unit input prices moved. The evidence that separates them: a purchase-price variance on unit input costs (are we paying more per unit of material?) set against the change in production mix (did material-heavy output rise?). If unit input prices held while material-heavy volume grew, the driver is mix rather than price, and the response differs in each case.

The lab is clean by design: each line's movement is entirely price or entirely volume, the bridge ties exactly, and the folder is minimal. Real bridges are messier, and three failure modes recur.

The joint variance gets buried. When price and volume both move on the same line, the change is not a clean sum of two parts; there is a joint term, the price change times the volume change, that some conventions fold into price, some into volume, and some report on their own. If the convention is left unstated, two analysts split the same $600,000 line differently and neither can reconcile to the other, so name the convention on the bridge before anyone reasons about cause.

The real driver is not in the bridge. The bridge shows that Precision Components gained 4,000 units, but not why; the reason, whether a distributor pre-buying ahead of a price increase, a competitor's stockout, or a one-time project, lives in the sales team's head rather than the CSV. Treat the decomposition as the question and not the answer, and confirm the driver with the business before it reaches the CFO.

The model narrates a cause the data does not support. Ask a model to explain a flat-price, higher-volume line and it will sometimes produce a confident "share gains in the mid-market segment" that no figure in the folder backs, which is the fluent-but-wrong output the jagged-frontier finding warns about. The tie-out catches broken arithmetic and the second hypothesis catches an unsupported story, so neither is an optional extra.

The workflow does not remove the analyst's judgment; it clears the mechanical work so the judgment has more room.

Rate your own explanation against the rubric below. An honest read of the rubric separates the parts of the decomposition you can run cold from the ones that still need practice. Your scores roll up to the workflow maturity dashboard on the course hub.