Predictive Analytics for the Enterprise

Enterprises buy predictive analytics as though every forecast were a weather forecast: an estimate of an indifferent world that improves as the instrument improves. Some are. But the forecasts carrying the most money are not observations of anything. They are estimates of decisions your own executives have not taken yet — how fast a role gets filled, how much discount survives the quarter, how much inventory somebody commits to in March.

You can raise the statistical quality of that estimate indefinitely and change nothing, because precision was never the constraint. The constraint is whether the person who sets the hiring pace agreed the assumption, and whether anything happens if it proves wrong.

What actually happened

Polymath’s founder ran global human capital planning inside a private consumer packaged goods business of roughly $2–3 billion in revenue that later went public on the TSX, before moving into corporate strategy and analytics. Payroll ran across approximately 21 countries; the workforce P&L was above $200 million. Three forecasts sat in his remit, and the tooling made them look like one problem.

Consumer demand was the honest one: what would sell through, in which category, at which retailer, over which weeks. The team automated demand planning and built the reporting behind it, because on a series produced by millions of people who have never heard of the company’s planning cycle, the instrument really is the constraint.

The second was acquired workforce cost. Doing the human capital work on three bolt-on acquisitions, he built the workforce models that held mean absolute percentage error to about 3%. That figure flatters the modelling. An acquired population is unusually forecastable because its decisions have already been taken and dated: pay contractually fixed, headcount scheduled in an agreement, effective dates in a signed document. Little was left for a colleague to change.

The third was the company’s own workforce cost, held at about 98% forecast accuracy against that $200 million-plus P&L, and none of it was fixed in advance. Adaptive Planning was the instrument; it settled nothing about ownership. The forecast asserted that specific positions would be opened, filled at a specific pace, at specific structures, in specific countries — every clause a decision belonging to somebody with a name: the leader opening the requisition, the hiring manager judging a candidate good enough, the payroll owner working to a statutory calendar.

So the assumptions were set where the behaviour sat, in the units those owners used rather than in dollars, and the variance conversation was held with them. When a month came in wrong, the question was which decision had changed, not which coefficient. Nobody had been forecast at without their knowledge.

Inventory after the pandemic held both halves in one number. Sell-through was external and modellable. The buy was not: it is a commitment a category owner makes on a date, and working stock down in one segment while another scaled was a sequence of those commitments.

The model never got more sophisticated. The number simply became somebody’s.

Polymath does not touch a client’s model until every assumption inside it has an owner, which is the whole of what follows.

Why it happens

Divide the forecast estate in two, because the halves obey different rules.

The first half describes a world that does not know it is being forecast. Consumer sell-through, input costs, currency, the weather over a bank-holiday weekend. Your decisions do not enter the series. Statistics is the right instrument, the error is honest noise, and a better model buys real money.

Most of what an enterprise funds as predictive analytics points at the second half. Workforce cost, hiring pace, operating spend, discount realization, inventory commitment, attrition, collections. These series are not observed. They are produced: each period is the accumulated result of decisions identifiable people have not yet taken. That changes what accuracy rests on. To forecast a produced number is to predict a colleague’s conduct, so it can only be one of two things: a plan that person is inside, or an estimate about them made by somebody else.

An estimate made about someone has a predictable reception, and it is not obstruction. Having agreed nothing, they hold no position the number can contradict, and the cheapest response available is to question the method. That assumes nine weeks to fill; ours runs longer in that market. The objection may be correct. It is also unanswerable in the room, because the person who could resolve it built the model.

An assumption nobody agreed to cannot be missed. It can only be disputed.

That dispute happens in the review, the last moment the number could change a decision. The decision then proceeds on whatever figure the participants already carried.

Then the standard remedy makes it worse. The analytics team, watching its output get argued with, concludes the output was not good enough: more features, longer history, a tighter error metric. Precision rises; standing does not. A sharper estimate of someone who never consented to it is more contestable than a rough one.

The inverse shows where accuracy already lives in your business. The numbers you reliably hit are the ones somebody is judged against, and that accuracy is partly manufactured: the owner adjusts what they do to land it.

The strongest objection

The strongest case against this argument: better models close the gap, and the evidence is in your own business.

We concede it twice over. For the external half of the estate this paper barely applies — automating demand planning produced a better estimate than the manual process it replaced, and the improvement was real. Less comfortably, a badly built model gives a sceptical executive a legitimate reason to reject the number. Model quality is a necessary condition. It is not the one that is missing. The second objection is better: let the people being forecast set their own assumptions and you have invented sandbagging. A negotiated number carries bias, a modelled number carries noise, and bias is the harder error. True, and common: participation without consequence produces conservatism.

What we propose is not to remove the owner but to remove the asymmetry and keep a check. Make being under as expensive as being over — an unfilled position is unused capacity somebody explains, not a saving — and the conservative assumption stops being free. Then keep the statistical estimate alive as a challenger and track the sign of its gap. A one-sided gap is a finding about a person, not about mathematics.

The method

We run six steps, in this order. We apply process redesign to the planning routine rather than to the model, and step one decides which of the rest apply to which numbers.

1. We split your Forecast-to-Plan register into observed and produced. We take the twelve to twenty forecasts your operating cadence actually consumes, and we classify each: generated outside the company, or the sum of internal decisions? The trade-off we accept: several are genuinely mixed, and we still assign each one to a class. Skip it and you spend model budget on numbers no model governs, and attention on numbers no meeting can influence.
2. We name the person whose decisions produce each number. Not the owner of the forecast. The owner of the behaviour. Time to fill belongs to the hiring manager and the recruiter, not to workforce planning; discount realization belongs to the sales leader, not to commercial finance. The trade-off we accept: the honest answer is often three people, so we pick the one with most control. Skip it and your planning team defends assumptions about work it does not do.
3. We write the assumption in the units the owner works in. Weeks to fill, requisitions opened per month, units committed, discount floor, days sales outstanding. Your finance team converts to money afterwards; we never ask the owner to. The trade-off we accept: your planning calendar now carries each assumption as a dated submission. Skip it and the owner cannot tell whether they agree, so they agree to all of it and act on none.
4. We attach a consequence, and we make it symmetric. Landing under the assumption must cost something visible — capacity not built, a position unfilled against a commitment, stock unavailable for a promotion — or the rational assumption is always the low one. The trade-off we accept: symmetry removes the comfort of conservatism, and in our experience those who benefited say so. Skip it and every assumption drifts one way, and you have traded noise for bias.
5. We keep the statistical baseline as a challenger, never as the authority. We publish the owned assumption and the model’s estimate side by side, and we track the gap and its sign across cycles. The owned number governs the plan; the model’s number governs the conversation about the owner. The trade-off we accept: two numbers in one pack invites the question of which is correct. Skip it and participation becomes negotiation with nothing to test it.
6. We run the post-mortem on assumptions, not on variance. When the number misses, we resolve the miss to the assumption that moved and the decision underneath it before we let anybody touch the model. The trade-off we accept: the review gets uncomfortable, because the miss now carries a name rather than a method. Skip it and the model is retuned every cycle, the behaviour is untouched, and the variance returns wearing a new interval.

None of this makes modelling less important; we change what modelling is for. We stop the model being the source of the number and make it the instrument that tests one — the only role in which it can be argued with productively. Everything else we do here is Process Redesign applied to a routine few companies treat as one: your Forecast-to-Plan cycle has submissions, owners, units and dates like any other process, and we treat accuracy as a property of that routine rather than of the estimator at the end of it.

Evidence and measures

The test is not whether a forecast has been accurate. It is who set the assumption:

Five measures worth tracking over ninety days. None requires a new system:

— Share of your top assumptions whose owner sits outside the team producing the forecast.

— How many of those owners can state their own assumption without looking it up. The least flattering item here.

— The sign of the gap between owned assumptions and the statistical baseline over three cycles. Consistently one-sided is a finding about people.

— Number of assumptions changed during the review meeting itself, and by whom. One revised in the room was never agreed before it. — Elapsed time between a forecast being published and the date the decision it predicts becomes irreversible: the requisition approved, the buy committed, the offer made.

If those five move, forecasting has become a planning activity. If the only thing that improved is the error metric, it has not.

How Polymath solves this

We do not start with the model. We start inside one live reforecast cycle, listing every assumption that moves the number as it is actually produced — who submits each, in what unit, and which decision each describes — and setting each one against the person who takes that decision. In most Forecast-to-Plan cycles, a large share of the material assumptions belong to the team that also has to defend them. We deliver that register as Process Redesign, not as an analytics finding.

We work inside your next cycle rather than in a document about it. We rewrite the assumption sheets in operating units with the leaders who own the behaviour — the hiring manager, the sales leader, the category owner — and with your planning lead; we change the submission sequence and its dates; we build the challenger baseline into the pack you already publish; and we chair the variance review once in the new format, as operators who have run a forecast ourselves. Then we hand the cycle after that back to the owners and leave the room.

What compounds is the register. Once every material assumption has an owner, a unit and a date, your scenario work stops being rebuilt from scratch, variance resolves to a decision instead of a coefficient, and commitment releases can be delegated against a named assumption. It is also what makes later automation safe: a model or an agent can be handed the assumption set and its owners rather than inferring both. Every cycle after this one inherits the register the last one left behind, so what your planners spend the month on shifts from assembling assumptions to arguing about the decisions underneath them.

What it costs to do nothing

The visible cost is rework: a reforecast run twice because the first version was argued with rather than acted on.

The larger cost is that the decision still gets taken, on the last figure anybody agreed to, which is older and cruder than the one under dispute. Across a workforce P&L above $200 million and payroll in roughly 21 countries, a hiring plan held for a month while its assumptions are relitigated is capacity arriving in the wrong quarter and cash committed on the wrong side of a season.

The third cost compounds and is rarely counted. Your strongest analytical people spend their quarters defending produced numbers instead of working the external series, where the model is the constraint and the return is unambiguous. That work is never deferred visibly; it simply never starts.

Accuracy nobody agreed to is not a forecast. It is a scorecard delivered early, and nobody plans against a scorecard. Your next reforecast cycle is where we begin.