In short: Trade spend allocation is a budgeting problem with a clean statement: a fixed budget, a set of candidate events across accounts, brands and periods, and a measured return for each, allocated to maximise total incremental margin. It works only where uplift has already been measured net of cannibalisation and pull-forward, so that measurement has to exist before any allocation is attempted. Treating it as moving money from where it is towards where it should be keeps the size of each move visible and the plan executable. Check the funded calendar against what the network can serve before committing it, because that step is what separates a plan from a calendar.
This piece assumes you already have a measured history: for each event you ran, the net incremental volume after cannibalisation and pull-forward, and the margin it produced against what it cost. If you do not have that, the measurement comes first and nothing here will work without it.
Given that history, the question is where the money goes next year. It is a budgeting problem rather than an analytical one, and it has a structure worth using.
The allocation, stated properly
You have a fixed budget, a set of candidate events across accounts, brands and periods, and a measured return for each event type in each context. Allocate to maximise total incremental margin.
The naive version ranks events by return on investment and funds down the list. That is close to right and it misses three constraints that bind in practice.
Diminishing returns within an account. The fourth event of the year at an account returns less than the first, because the buyers who respond to a deal have been reached. Treating each event as independent overstates the value of concentration, and a ranked list will pile spend into whichever account had the best measured return.
John Little set out the general form of this in Management Science in 1970, arguing that marketing response should be modelled as a saturating curve fitted from whatever combination of data and management judgement is available, rather than as a constant rate of return. His point applies directly here: a return on investment figure is an average over the events you ran, and the decision in front of you is about the next event, which sits further along the curve than the average does. Ranking by average and funding down the list quietly assumes the curve is flat, and the whole reason to reallocate is that it is not.
Customer commitments. A large share of trade spend is contractually committed through annual terms, and only the discretionary remainder is genuinely allocable. Optimising the whole budget when two thirds of it is committed produces a plan that cannot be executed.
Supply feasibility. The events you fund generate volume the network has to serve. Concentrating spend into one quarter because that is where the returns were measured can produce a demand profile that exceeds capacity, and an event that stocks out returns nothing while costing everything.
That third constraint is the one most often discovered after the calendar is locked.
Reallocation rather than optimisation
A useful framing for a live business is that you are moving money from where it is to where it should be rather than building a calendar from scratch, and the size of the move matters as much as its direction.
Three practical reasons to reallocate incrementally rather than reoptimising wholesale.
Measured returns carry uncertainty, and a wholesale reallocation acts on that uncertainty at full confidence. A twenty percent shift acts on it proportionately.
Customer relationships absorb change at a limited rate. Cutting an account's support by half because last year's events underperformed will produce a reaction that is not in the model.
And the measurement improves as you go. Each cycle of reallocation generates new observations, particularly at accounts where the level of support changed, which is exactly the variation that identifies the response curve.
A reasonable operating rhythm moves ten to twenty percent of discretionary spend per year, directed by the measured record, with the changes concentrated where the evidence is strongest.
That sounds modest until you price it. Say two million of the budget is genuinely discretionary and you move ten percent of it, so two hundred thousand. The money comes out of fourth and fifth events at heavily promoted accounts, where the measured marginal return is around 0.9 times spend in incremental margin, and goes into first events at accounts running almost no activity, where it is around 2.1. The gain is two hundred thousand multiplied by the difference of 1.2, which is two hundred and forty thousand of incremental margin on a budget that did not change.
Then discount it honestly, which is the part that makes the case rather than weakening it. Those two return figures are estimates with error attached. If each carries something like plus or minus 0.4, the pessimistic version of the same move is 1.7 against 1.3, a difference of 0.4, and a gain of eighty thousand. The move is still positive across the whole plausible range, which is precisely the property you want from a decision made on uncertain measurement, and it is the property a fifty percent reallocation would not have. Moving a small share of the budget is what keeps the pessimistic case above zero.
Where the money usually should go
Patterns that recur once businesses start measuring properly.
Away from deep discounts on high-penetration products, which mostly buy volume from buyers who would have purchased anyway.
Toward events that generate trial on products with strong repeat rates, where the value accrues after the event and does not appear in a four-week read.
Away from the fourth and fifth events at a heavily promoted account, toward first events at accounts with little activity, where the response curve is steepest.
Toward mechanics with lower stockpiling potential when the product is storable, since forward buy converts spend into timing rather than volume.
None of these is universal, and the point of measuring is that your business will have its own version. The patterns are worth knowing as prior expectations to test rather than as conclusions to apply.
The first two have a long empirical grounding. Blattberg and Neslin, surveying the sales promotion literature in 1990, reported what remains the central finding of that literature, though van Heerde, Gupta and Wittink revised the decomposition materially in 2003: the great majority of the volume lift on a promoted item comes from brand switching and from stockpiling by existing buyers, with category expansion contributing a much smaller share. That is the reason deep discounts on high-penetration products read well on a four-week uplift and poorly on an annual margin line, and it is worth having the reference to hand when the four-week number is being waved around.
The supply handshake
The step that separates a plan from a calendar is checking the funded set against what the network can serve.
For each candidate calendar, project the demand it generates by period and location, then check it against capacity and inventory. Where it does not fit, the choice is to move the event, reduce it, or accept a service risk, and each of those has a cost that belongs in the allocation decision rather than being discovered by the supply team in February.
This is not a hard calculation and it is organisationally awkward, because it requires the commercial calendar to be visible to supply planning before it is committed. In many businesses the calendar is agreed with customers first and shared afterwards, which makes the check a report rather than an input.
Moving that one step earlier is usually worth more than any improvement to the allocation method.
The reason it is worth more is that a supply failure inside an event does not stay contained to that event. Work it through. An event is expected to lift an account from a baseline of 12,000 cases in the period to 40,000, and the plant can supply 30,000 across the window. The event runs at seventy-five percent availability, so the display empties in week two, the account applies whatever service penalty the terms allow, and the measured incremental volume comes in around 18,000 against a spend sized for 28,000. The measured return on that event is roughly two thirds of what the mechanic actually delivers.
Then that number goes into the response history and drives next year's allocation. The model reads a poor return for that mechanic at that account and moves money away from it, so a mechanic that works gets defunded because of a capacity constraint that had nothing to do with it. The symptom is a distinctive one and it is worth looking for in your own history: an event type whose measured return dropped sharply in a single year, at one or two accounts, with no change in price, depth or timing. That pattern is almost always a service problem rather than a response problem.
The fix is a join. Before an event read enters the response model, attach the service level achieved during the event window, and either exclude events that ran below a threshold or carry the shortfall as a correction. Without that join the measurement system slowly learns to avoid every event large enough to strain the network, which is a preference for small events dressed up as evidence.
What to measure afterwards
Two things, tracked over time rather than per event.
Realised return against modelled return, by event type. The gap between them tells you how much to trust the model's recommendations, and it should narrow. If it does not, the measurement is missing something structural rather than being noisy.
Spend concentration. What share of the budget sits with the top accounts and the top events, and whether it is moving in the direction the evidence supports. Concentration tends to drift upward through the year as in-year requests get funded from wherever there is slack, and tracking it catches that.
Make that one specific enough to argue about. Record the share of discretionary spend held by the top five accounts at the point the calendar is signed, and again at the year end after every in-year request has landed. A gap of more than a few points means the plan you approved and the plan you executed are different plans, and the difference was decided one urgent request at a time by whoever had budget left. That is a reallocation nobody chose, running in the opposite direction to the deliberate one described above, and it is frequently larger. Where the two are of comparable size, the honest conclusion is that your allocation method is being outvoted by your in-year approval process, and the approval process is the thing to fix first.
The limit
Historical returns describe how events performed in a market that has moved on. Competitive activity, channel mix and consumer price sensitivity all shift, and an allocation built entirely on last year's measured returns is fitting the previous environment.
The defence is to keep some deliberate variation in the calendar: a portion of spend allocated to test rather than to exploit, so the model keeps learning about parts of the response surface it would otherwise stop observing. An allocation that only funds what worked before stops generating information about anything else, and its estimates decay quietly.
There is also a category of value that no event measurement captures well. Trade spend buys relationship and shelf position as well as volume, and an account whose events return poorly may still be worth supporting for reasons that show up in distribution rather than in an event read. That argument is legitimate and it is also the argument that defends every underperforming account, so it should be made for specific accounts with a specific rationale and checked against distribution outcomes afterwards.
Start by identifying what share of your trade budget is genuinely discretionary this year. It is usually smaller than people assume, and it defines what any allocation exercise can actually change.