TL;DR
- The core decision: where the staffing number for each period actually comes from.
- When doing nothing is right: when the number is close enough and somebody owns it.
- What has to be true: you can trace the figure back to something other than habit.
- How the options split: by who produced the number and what they were looking at.
- Decision rule: ask which direction it's usually wrong in, not how accurate it is.
- Outcome to expect: the same uncertainty, held by somebody who knows it's uncertain.
Four on a Tuesday
Every rota starts with a number. Four people on Tuesday morning, six on Friday evening, three on a Sunday. Those numbers arrive looking like facts, and somebody builds a week around them.
Ask where they came from and the answer gets vague quickly. Somebody worked it out. It's what we've always done. It came down from somewhere. The manager who set it left three years ago. Occasionally there's a real method behind it, and even then the person applying it usually can't say what assumptions it rests on.
This matters more than it seems, because every other decision in scheduling sits on top of that number. The rota, the cost of the week, whether you're short, whether anybody has to be sent home, whether the shift is unpleasant to work. All of it is downstream of a figure nobody examines, because it arrives in the form of a requirement rather than an estimate.
Best tools for Employee Scheduling
Here's the reframe. The useful question is not whether the forecast is accurate, which is unanswerable and unhelpful. It's which direction it's usually wrong in, because overstaffing and understaffing are not symmetrical and they're not paid for by the same people. Overstaffing costs the organisation money, visibly, in a number somebody reports. Understaffing costs the people on shift, invisibly, in a week that was harder than it needed to be.
Which means a forecast that's consistently a little high looks expensive and is survivable, and one that's consistently a little low looks efficient and grinds people down. Both are wrong. Only one gets noticed.
Everything below is about tracing your own number back to its origin and working out which way it leans, because that's a question you can actually answer about your own operation.
When You Genuinely Do Not Need to Act Yet
The number is close and somebody owns it. It's roughly right, somebody can explain where it comes from, and it gets revisited. That's a working arrangement.
Nobody can say where it came from. Extremely common, and worth tracing regardless of whether the number turns out to be good. An unowned figure can't be improved because nobody is responsible for it.
The same period is always wrong. Fridays are always short, or Tuesdays always have people standing around. A repeated error is information rather than noise, and it's the cheapest thing to fix here.
The edge case that forces it. Demand changed, the number didn't, and nobody noticed for months. That happens when a figure is inherited rather than produced, and it's the argument for knowing where yours comes from.
Five Questions This Reader Asks at 11pm
Where should the number come from? There isn't one right source, and the useful thing is that somebody can say what yours is. A figure from an experienced person's judgement can be excellent. A figure from history can be excellent. A figure nobody can account for is the problem, whatever its origin was.
How accurate are these forecasts? Not a question with a useful answer, and asking it tends to end the conversation. Nothing anybody quotes is grounded in your operation, your demand or your definition of adequate cover. What's worth knowing is which direction yours errs in and by roughly how much, and that's observable in your own data.
Should we staff for the average or the peak? Neither, usually, and putting it that way hides the real question, which is what happens when you're wrong. Staffing for the peak means paying for cover you frequently don't need. Staffing for the average means being short regularly, and short is felt by the people on shift rather than showing up in a report.
Who should own the forecast? Somebody named, and ideally somebody who experiences the consequence of it being wrong. A number produced by a person who never works a shift and never sees the floor will drift towards whatever is cheapest, because that's the only feedback they receive.
What if demand really is unpredictable? Some of it is, and a portion of what feels unpredictable turns out to be predictable once anybody looks: a weekly pattern, a seasonal one, an effect nobody had connected. Worth separating the genuinely random part from the part that's simply unexamined.
Wrong in Which Direction
| The error | Who absorbs it | What it looks like on the day |
|---|---|---|
| Consistently a little high | The organisation, visibly | People with not enough to do, and a cost somebody queries |
| Consistently a little low | The people on shift, invisibly | A harder week that nobody reports |
| Right on average, wrong at peaks | Whoever is on during the peak | A bad hour repeated regularly |
| Right on average, wrong at troughs | The organisation, and morale | Standing around, which people dislike more than being busy |
| Wrong for one period every week | Whoever works that period | A known problem everybody has stopped mentioning |
| Wrong since something changed | Everybody, gradually | A number that no longer matches the work |
| Right for the hours, wrong for the mix | Whoever lacks the skill on shift | Enough people, none of them able to do the thing |
| Nobody knows which way it errs | Nobody can tell | The problem is invisible, which is why it persists |
The asymmetry in the first two rows is the whole argument. An over-forecast produces a number in a report that somebody asks about, so it gets corrected. An under-forecast produces a difficult shift that the people working it absorb, and nothing generates a figure anybody reviews. The result is that forecasts drift downward over time, not by decision but because only one kind of error has a feedback loop attached.
The fourth row is worth noticing because it's counterintuitive. People generally tolerate being busy better than they tolerate standing around, so over-staffing a quiet period costs you money and some goodwill rather than money alone.
The seventh row catches operations that count heads. The right number of people is not the same as the right number of people who can do the work, and a forecast expressed purely in hours will regularly produce a shift that's technically covered and practically not.
Five Diagnostic Questions You Can Self-Assess Against
Where did this week's numbers come from? Trace one period back as far as it goes. The exercise usually ends with a person who left or a decision nobody remembers, and that's the finding.
Which way does it err? Look at periods where you sent somebody home or called somebody in. The balance between those two tells you the direction, and the direction matters more than the size.
Which period is always wrong? Every operation has one. It's known to everybody on shift and it's usually never been raised, because it's been that way long enough to seem like weather.
Does the number account for the mix? Hours versus capability. A forecast that says four is silent about whether any of the four can do the thing that actually needs doing.
When did it last change? If the answer is years, check whether the work has. Numbers inherited across a change in the operation are the most reliably wrong ones you'll have.
Run these five with somebody who works the shifts rather than only with whoever produces the numbers. The person on the floor knows which periods are wrong and in which direction, and that's precisely the information the forecast has no way to generate.
Six Ways the Number Gets Produced, Reviewed
Somebody experienced deciding from memory
A person who knows the operation says how many are needed. It earns its place because experienced judgement is genuinely good at this: it incorporates things no data captures, and somebody who has worked the floor knows what four people on a Tuesday actually feels like.
Where it falls short is transfer and drift. The reasoning is in one head, it can't be examined, and it leaves when they do. It also anchors on recent memory, so an unusual period a few weeks ago carries more weight than it should.
Excellent while you have that person. Get them to say the reasoning out loud occasionally, because the reasoning is more valuable than the number and it's the part that disappears.
The recording doesn't have to be elaborate. A line against each period saying why it's that number, written once in their own words, captures most of what would otherwise leave with them.
Watch for the anchoring problem too. Judgement of this kind weights recent experience heavily, so a fortnight that was unusually busy will pull the numbers up for months afterwards, and a quiet spell will pull them down, neither for any reason anybody would endorse if it were stated.
Last year's equivalent period
You look at what the same week required a year ago. It earns its place on seasonality, since annual patterns are real in most operations and recent averages miss them entirely.
Where it falls short is that last year contained its own errors and its own circumstances. If you were short that week, the record shows the hours you actually ran rather than the hours you needed, so the shortage gets copied forward as a target.
Useful for shape rather than level. Worth knowing whether last year's figure was what you scheduled or what you actually needed, because those are different numbers and only one is in the system.
That distinction is the trap in this method and it's easy to miss. Historic records show the hours that were worked, which are the hours that were scheduled, which came from a forecast that may have been wrong. Copying it forward copies the error with it, year after year, with increasing authority because it now has history behind it.
It's also worth checking what else was different. A period a year ago may have had a different layout, a different mix of work or a different set of people, and the number carries none of that context.
A rolling recent average
Recent periods averaged forward. It earns its place on responsiveness: it adapts to changes in demand without anybody deciding, which means it won't be years out of date.
Where it falls short is at anything unusual. An average smooths over exactly the peaks where being short hurts most, and it carries recent one-off events forward as though they were normal.
Reasonable as a baseline. Worth pairing with somebody who knows whether the recent period was typical, because the average has no way to know that a fortnight was distorted by something.
The smoothing is the property to watch. An average is by construction wrong at both ends, and the end where it's wrong low is the one that produces a difficult shift, which means a rolling average quietly guarantees a handful of bad periods you could have anticipated.
How far back it reaches matters more than people expect. A short window follows noise and a long one ignores real change, and neither setting is obviously right, which is a reason to know what yours is rather than to accept a default.
A figure the system produces from history
The scheduling tool generates a requirement from accumulated data. It earns its place on consistency and effort, and on catching patterns a person wouldn't notice.
Where it falls short is opacity and inherited error. Most people using such a figure can't say what it's built from or what it assumes, and it learns from the hours you ran, which are the hours you scheduled, which came from the previous forecast. That loop can reinforce a level nobody chose.
Worth having and worth interrogating. Ask what it's learning from and whether it knows the difference between hours scheduled and hours needed.
The loop is the thing to establish clearly. If the figure is derived from hours you ran, and those hours came from the previous figure, then the system is largely reproducing your own past decisions with increasing confidence, and a level nobody chose becomes a level nobody can argue with.
Ask as well what happens when something changes. A generated number built on a settled pattern will adapt slowly to a genuine shift in demand, which means the period after any real change is the period it's most likely to be wrong.
A target handed down from elsewhere in the business
A number arrives as a requirement rather than an estimate, usually expressed as a limit. It earns its place as a legitimate constraint: the business has a position and scheduling has to work within it.
Where it falls short is that it's a budget wearing the clothes of a forecast. Presented as how many people are needed, it's actually how many are affordable, and the difference matters because only one of those is a claim about the work.
Treat it as a constraint and say so. If the number is a limit rather than an estimate, naming it as one lets you have the real conversation about what the limit means for coverage.
The distinction is not a way of resisting the budget. Limits are legitimate and most operations have them. What the confusion prevents is anybody being able to say plainly what the limit costs in coverage terms, because arguing with a stated requirement about the work sounds like arguing that the work is different from what it is.
Where a target arrives, it's worth writing down what you'd have scheduled without it. That difference is the honest measure of what the constraint is asking people to absorb, and it's the only version of the conversation that goes anywhere.
A number nobody can account for
It's simply what's always been used. It earns its place nowhere, and it's more common than anybody admits.
Where it falls short is that it can't be improved. Nobody owns it, nobody can explain it, and nobody can say whether it was ever right, so every conversation about staffing levels stalls on a figure with no provenance.
Worth tracing once. The exercise takes an afternoon and typically ends either with a reasonable original rationale that has since expired, or with nothing at all, and both outcomes are useful.
The expired rationale is the more common finding and the more interesting one. A number set for a layout you no longer have, a service you stopped offering, or a pattern of demand that moved is not an arbitrary figure, it's a correct answer to a question nobody asks any more.
Once traced, somebody should own it going forward. The reason these figures persist unexamined is not that they're hard to examine, it's that nobody's name is against them and therefore nobody's job includes noticing when they stop being right.
The Decision Table
| Situation | Scale | Setup | Primary Pain | Recommended Starting Point |
|---|---|---|---|---|
| Number close, somebody owns it | Any | Any | None | Change nothing |
| Nobody can account for the figure | Any | Any | Cannot be improved | Trace one period back |
| One period always short | Any | Any | Known and unraised | Fix that one period |
| Sending people home regularly | Any | Any | Errs high, which gets noticed | Look at the trough forecast |
| Calling people in regularly | Any | Any | Errs low, which does not | Count these, they are invisible |
| Figure inherited from an experienced leaver | Any | Any | Reasoning gone, number remains | Rebuild the reasoning |
| System generates it from history | Any | Has a system | Learning from your own errors | Ask what it learns from |
| A budget presented as a forecast | Any | Any | Two claims confused | Name it as a constraint |
| Enough people, wrong skills | Any | Any | Covered on paper only | Forecast the mix, not the hours |
The fifth row is the one to start counting, because it's the error nobody records. Sending somebody home produces a conversation and possibly a cost line. Calling somebody in, or simply running short and coping, produces neither, so an operation that's consistently under-forecast has no data saying so.
The eighth row is worth being direct about internally. A number that's actually a budget is a legitimate thing to have, and calling it a forecast prevents the honest conversation about what that budget means for the experience of working a shift. Separating the two lets both be discussed properly.
The sixth row is the one with a deadline attached. Reasoning held by somebody who is leaving can be captured while they're still there, cheaply, and cannot be recovered afterwards at any price.
The Number Nobody Questions
It arrives as a requirement, not an estimate. That's the whole reason it goes unexamined. A figure presented as how many people are needed reads as a fact about the work, and facts don't get questioned the way estimates do.
Only one kind of error has a feedback loop. Over-forecasting generates a cost somebody queries. Under-forecasting generates a harder shift nobody reports. Over time that asymmetry pulls the number down, without anybody choosing to lower it.
Trace it once and you'll learn something. Following a single period's number back to its origin takes an afternoon and typically ends with a person who left, a decision nobody remembers, or an original rationale that stopped applying some time ago.
An expired rationale is not an arbitrary number. It was a correct answer to a question your operation no longer asks, which is a different and more fixable problem than having no basis at all.
Hours are not the same as capability. A forecast expressed in headcount is silent about whether the heads can do the work. Operations that have ever had a technically covered shift where nobody could do the necessary thing know this already.
Systems learn from what you scheduled, not what you needed. A figure generated from history is built on hours you actually ran, which came from your previous forecast. If that was low, the system will confidently reproduce it.
Which makes a generated number harder to argue with than a person's. Nobody challenges a figure that came out of a system, so an inherited error acquires authority it never had when somebody was setting it by hand.
The unpredictable part is smaller than it feels. Some demand genuinely is random. A meaningful share of what gets called unpredictable turns out to have a pattern nobody has looked for, and finding it is cheaper than staffing for chaos.
And the people on shift can usually name the pattern. They've been absorbing it for years and can tell you which days go wrong and why, which is a faster route to it than any analysis.
The conclusion is not that you need a better forecasting method. It's that you need to know which direction your existing one leans and who is absorbing the error, because that single piece of knowledge changes more than any improvement in technique.
It also makes any later improvement measurable. Changing method without knowing your current direction of error means you can't tell afterwards whether anything got better or merely different.
Where These Arrangements Go Wrong
| The failure | How it shows up | What would have to change |
|---|---|---|
| Under-forecasting has no feedback | The number drifts down over years | Count short shifts deliberately |
| Budget presented as forecast | An unwinnable argument about coverage | Name the constraint as a constraint |
| Inherited from somebody who left | Reasoning gone, number remains | Rebuild it, or replace it |
| System learns from scheduled hours | Yesterday's error, reproduced confidently | Ask what it is learning from |
| Forecast in hours, not skills | Covered on paper, stuck in practice | Include the mix |
| Known bad period never raised | Everybody accepts it as weather | Ask the people who work it |
The first row is the failure that produces the others. Because being short generates no record, the operation has no evidence that it happens, so every conversation about staffing levels is conducted with data that only shows one kind of error.
The sixth row is worth a specific conversation. There's always a period everybody on shift knows is under-resourced, it's been that way long enough to stop being mentioned, and the people who could tell you have concluded that raising it goes nowhere. Asking directly usually gets an immediate and specific answer.
The fifth row is the quiet one, because a shift can be fully staffed by the numbers and still stuck. Counting heads without counting capability produces a rota that satisfies every check and fails on the day, and nothing in the forecast will have indicated it.
What to Put in Writing
| Artefact | Who owns it | When it is written | What it prevents |
|---|---|---|---|
| Where each period's number comes from | Whoever builds the rota | Now | A figure with no provenance |
| Times you called somebody in | Whoever builds the rota | Ongoing | An invisible error direction |
| Times you sent somebody home | Whoever builds the rota | Ongoing | Only seeing one kind of error |
| Which periods are known to be short | Whoever works them | Now, by asking | A problem accepted as weather |
| Whether a number is a target or an estimate | Whoever issues it | Whenever it is issued | Two claims confused |
| The reasoning, not just the number | Whoever produced it | Before they leave | Inheriting a figure with no basis |
The second and third rows together are the instrument this whole piece is about. Two counts, kept for a couple of months, tell you which direction your forecast errs in, which is the thing nobody currently knows and the thing that determines who is absorbing your uncertainty.
The fourth row is the cheapest of the six and needs no data at all. Asking the people who work your worst period whether it's adequately staffed takes a conversation, and the answer is available today.
Questions to Ask Before You Commit
On provenance. Where does this number come from? A bad answer is the system.
On direction. Which way is it usually wrong? A bad answer is it's about right.
On evidence. How often were we short last month? A bad answer is not often.
On learning. What does the generated figure learn from? A bad answer is historical data.
On mix. Does it account for who can do what? A bad answer is headcount.
On status. Is this a target or an estimate? A bad answer is both.
What Getting This Wrong Costs
The first cost is a forecast that drifts downward without anybody deciding it should. Because over-staffing generates a visible number and under-staffing generates a difficult shift, only one of those errors gets corrected, and the result over several years is a staffing level nobody chose and nobody can defend. Each individual adjustment was reasonable. The direction of travel was never a decision.
By the time it's noticed, the current level has history behind it and looks like the norm, which makes arguing for more people sound like asking for something extra rather than like restoring something that was quietly removed.
The second cost lands on whoever works the period everybody knows is short. It's a known problem, it's been known for a long time, and it's no longer raised because raising it went nowhere previously. That's a specific group of people having a worse job than their colleagues, permanently, because of a number that was set once and has never been examined.
Those are usually also the shifts hardest to fill, which gets read as a recruitment difficulty rather than as a consequence. The two are the same problem seen from different ends.
The third cost is confusing a budget with an estimate. When a limit arrives described as a requirement, the conversation that needs to happen, about what that limit means for coverage and for the people delivering it, becomes impossible, because you can't argue with a fact. Naming it as a constraint doesn't change the constraint and does let everybody discuss the consequence honestly.
It also puts whoever builds the rota in an impossible position. Asked to produce adequate coverage from a number that was never about adequacy, they can only fail quietly, and the failure lands on the shift rather than on the conversation where it belonged.
So do three things, and two of them are just counting. Trace one period's number back to its origin. Start recording how often you call somebody in, and how often you send somebody home. And ask the people who work your worst period whether it's adequately staffed, because they know and they have probably stopped saying.
None of the three requires a system, a budget or anybody's permission, and the counting is a tally rather than a project. What they produce between them is the first honest picture most operations have had of their own staffing levels.
When You Are Ready to Go Further
Start with the two counts, because they're nearly free and they answer the question nobody can currently answer. How often you were short, how often you were over, kept for a couple of months. That balance tells you which direction your forecast leans, and therefore who has been absorbing the error, which reframes every subsequent conversation about staffing.
Keep them as a tally on whatever is nearest rather than as a system. The value is entirely in having the two numbers at the end, not in how carefully they were captured.
Then trace a single number back. Pick the period you're least confident about and follow the figure to its origin. You'll find a rationale that made sense and has expired, a person who left, or nothing, and each of those tells you something about how much of your forecasting is actually inheritance.
Write down whatever you find, even if it's nothing. A note saying this figure has no traceable origin is more useful than the silence it replaces, because it makes the number available to be changed rather than treated as given.
Finally, ask the people on your known bad period. Every operation has one, everybody who works it knows, and it's usually been unmentioned for long enough to feel permanent. That conversation costs nothing, produces a specific and immediate answer, and is the one piece of forecasting information you can obtain today without any data at all.
HROpsLab publishes independent comparison work across HR tooling and workforce systems. We sell nothing, we take no vendor money, and we publish no paid placements. If the next step is understanding what your current tooling can forecast, our comparison work is one place to start.
Frequently Asked Questions
What does labour forecasting mean in scheduling?
Producing the number of people needed for a given period, which every rota then gets built around. It sounds technical and in most operations it isn't: the figure comes from somebody's judgement, from what the same week required last year, from a rolling average, from a system, from a budget handed down, or from habit nobody can account for. The important property is that it arrives looking like a requirement rather than an estimate, which is exactly why it tends to go unexamined while everything downstream depends on it.
Where do staffing numbers actually come from?
Trace one and you'll find out, and the exercise usually takes an afternoon. Common endpoints are an experienced person's judgement, which is frequently excellent but lives in one head, a figure inherited from somebody who left, a system-generated number built from your own historic hours, or a target from elsewhere in the business that's really a budget. A meaningful share of operations end the trace with nothing: the number is simply what has always been used, which means nobody owns it and nobody can improve it.
How accurate are demand forecasts for scheduling?
It's the wrong question and asking it tends to close the conversation, because no figure anybody quotes is grounded in your demand or your definition of adequate cover. The useful question is which direction yours errs in, because the two errors aren't symmetrical. Over-forecasting costs the organisation money in a number somebody queries. Under-forecasting costs the people on shift a harder week that produces no record at all. That asymmetry is observable in your own operation and matters more than any accuracy measure.
Should you staff for the average or the peak?
Framing it that way hides the real question, which is what happens when you're wrong and who bears it. Staffing to the peak means regularly paying for cover you don't need, which is visible and gets challenged. Staffing to the average means being short frequently, which is absorbed by whoever is on shift and generates nothing anybody reviews. Neither is simply right. What's worth knowing is which of those two costs your operation is currently choosing, usually without having decided.
What happens when you consistently overstaff?
It costs money, visibly, in a figure somebody asks about, which means it gets corrected fairly quickly. Worth knowing that it also costs some goodwill, because people generally tolerate being busy better than standing around, and a quiet shift with too many people on it is unsatisfying to work. The more important point is structural: because over-staffing has a feedback loop attached and under-staffing doesn't, an operation left alone will drift towards being under-staffed rather than over, without anybody deciding to.
What happens when you consistently understaff?
The people on shift absorb it, and almost nothing records that it happened. There's no cost line for a week that was harder than it should have been, no report showing that three people did the work of four, and no figure anybody queries. It shows up eventually as turnover, as difficulty recruiting for particular patterns, or as a known bad period everybody has stopped mentioning. This is why counting the times you called somebody in is worth doing deliberately: otherwise the error has no evidence.
How do you check whether a staffing number is any good?
Two counts, kept for a couple of months. How often you called somebody in or ran short, and how often you sent somebody home or had people idle. The balance tells you the direction of your error, which is the thing that actually matters and the thing nobody currently knows. Alongside that, ask whoever works your least-liked period whether it's adequately staffed. They know, the answer is specific and immediate, and it usually hasn't been raised for years because raising it previously went nowhere.
Who should own the staffing forecast?
Somebody named, and ideally somebody who experiences the consequences when it's wrong. A number produced by a person who never sees a shift receives only one kind of feedback, the cost kind, and will drift accordingly. The other thing worth establishing is whether the figure is an estimate of what the work requires or a limit on what's affordable, because those are different claims and confusing them makes the conversation about coverage impossible. Both are legitimate; only one is a statement about the work.
Overstaffing shows up in a report. Understaffing shows up in somebody's week.
