Call center forecasting is the practice of predicting how much contact volume will arrive, in what intervals, and how long each contact will take, so you can staff to it. The output is workload, not headcount: forecast volume multiplied by average handle time. Mature voice operations typically hit 5 to 8 percent error at the daily level, digital channels 10 to 12 percent, and roughly half of contact centers still do this in a spreadsheet.
Forecasting is the first step in a chain, and it is the one that gets the least scrutiny. A staffing calculator will happily turn a bad forecast into a precise, confident, wrong answer. So will your budget. If the volume number is 20 percent light, every downstream calculation inherits that error, and no amount of Erlang math will rescue it.
The uncomfortable part is that most teams cannot say how accurate their forecast is. They produce one, they staff to it, and they never score it against what actually happened. That single missing step is why forecasting feels like guesswork in so many operations: not because the methods are hard, but because nobody is measuring the result.
What is call center forecasting?
Call center forecasting predicts future contact demand from historical data plus known future events. It answers three separate questions, and confusing them is where a lot of forecasts go wrong.
The first is how many contacts will arrive in a period. The second is when they arrive, broken into intervals of 15 or 30 minutes, because 1,000 calls spread evenly across a day and 1,000 calls arriving between 9am and 11am need completely different schedules. The third is how long each one takes, which is your average handle time forecast, and it moves independently of volume.
Only after all three do you convert demand into people. That conversion is a queuing calculation, not a forecast, and it belongs to the call center staffing calculator.
The call center forecasting formula
The core formula is short:
Workload (in hours) = forecast contact volume x average handle time
If you expect 900 calls in an interval and your AHT is 6 minutes, that is 5,400 minutes, or 90 hours of work. Workload is the honest unit of demand because it survives channel changes. A shift from phone to chat can cut nothing from your workload even while call volume falls.
Building the volume side usually looks like this:
Forecast volume = base trend x seasonal index x day-of-week index x intraday distribution, adjusted for known events
Each multiplier answers a different question. The base trend captures growth or decline. The seasonal index captures the month or week of year. The day-of-week index captures the fact that Monday is not Thursday. The intraday distribution slices the day into intervals. Known events (a product launch, a billing run, a price change, a planned outage) are added by hand, because no statistical model can see them coming.
Call center forecasting methods compared
Five approaches cover almost every contact center. They are not competitors so much as a ladder: you move up it as your data and your volume justify the complexity.
| Method | How it works | Best for | History needed |
|---|---|---|---|
| Moving average | Averages the last N periods, optionally weighted toward recent ones | Stable, low-volume queues with no real seasonality | 3 to 6 months |
| Exponential smoothing | Weights recent observations more heavily on a decay curve | Queues with a trend but weak seasonality | 6 months |
| Holt-Winters (triple exponential smoothing) | Models level, trend and seasonality together | The default for most contact centers with clear weekly and annual patterns | 2 full seasonal cycles, so 2 years for annual seasonality |
| ARIMA | Models the series against its own lagged values and errors | Long, clean, high-volume series where you can validate the fit | 2 years or more |
| Regression with drivers | Predicts volume from external variables: marketing spend, shipments, active accounts, billing cycles | Operations where volume is caused by something you can see in advance | 2 years plus the driver data |
| Machine learning | Fits non-linear patterns across many inputs at once | Large multi-channel operations with messy drivers and enough data to train on | 2 years or more, plus tooling |
Two practical notes. Holt-Winters is where most teams should start and where a surprising number should stop: it handles the level, trend and seasonality that describe the overwhelming majority of contact patterns, and it runs in a spreadsheet. And regression with drivers is the method most often skipped despite being the most useful, because in a lot of businesses volume is not mysterious at all. It is caused by invoices going out, orders shipping, or a campaign going live, all of which you know about in advance.
Erlang C and Erlang A appear on many forecasting lists, but they are not forecasting methods. They are queuing models that convert a finished forecast into required agents. Erlang C assumes nobody abandons the queue; Erlang A models customer impatience and so usually returns a leaner, more realistic staffing number.
How do you measure forecast accuracy?
Score the forecast against actuals every single week, at the same grain you staff to. Three measures do the job. MAPE is the average absolute percentage error across intervals. WAPE weights that error by volume, which stops tiny intervals from dominating the score. Bias is the signed average, and it tells you whether you are consistently over or under.
Bias is the one people forget, and it is the most operationally expensive. A forecast with 8 percent MAPE and near-zero bias is noisy but honest. A forecast with 8 percent MAPE that is always low is a structural understaffing machine, and it will show up in your first response time long before anybody blames the forecast.
| Measure | Formula | What it tells you |
|---|---|---|
| MAPE | Average of |actual - forecast| / actual, as a percentage | Typical error size, treating every interval equally |
| WAPE | Sum of |actual - forecast| / sum of actual | Error weighted by volume, the fairer score for a contact center |
| Bias | Average of (forecast - actual) / actual, keeping the sign | Whether you systematically over or under forecast |
What is a good forecast accuracy for a call center?
For mature voice queues, 5 to 8 percent error at the day level is a reasonable target, and digital channels sit closer to 10 to 12 percent because they are lumpier and easier to defer. Interval-level accuracy is always worse than daily, and daily is always worse than weekly. Published guidance commonly frames the goal as 85 to 95 percent accuracy overall.
| Horizon | Realistic accuracy | What it drives |
|---|---|---|
| Annual or quarterly | Directional, within 10 to 15 percent | Budget, hiring plan, outsourcing commitments |
| Monthly | Roughly 5 to 10 percent error | Recruitment and training pipeline |
| Weekly and daily | 5 to 8 percent voice, 10 to 12 percent digital | Schedules and shift bids |
| Intraday (15 to 30 min) | Materially worse, often 15 to 25 percent | Real-time reassignment, breaks, overtime calls |
Do not chase interval accuracy that the data cannot support. A small queue with 20 calls in a half hour has irreducible randomness, and the correct response is scheduling flexibility, not a better model.
What data do you need to forecast call volume?
At minimum: two years of contact volume by interval and channel, handle time over the same period, and a record of what was abnormal. That last one is what separates a usable history from a misleading one. An outage, a recall, a botched email send or a pricing change all leave spikes in your data, and if you feed those spikes back into the model unlabeled, you will forecast them forever.
Keep a simple event log alongside the volume data with the date, what happened, and roughly what it did to volume. Cleaning known anomalies out of history, a step usually called outlier scrubbing, is often worth more accuracy than upgrading the model.
Getting at that history is its own obstacle when volume data lives in a warehouse and handle time lives in the ACD. Teams without an analyst to hand increasingly query the warehouse in plain English rather than waiting in a reporting queue, which matters here because a forecast you cannot refresh weekly is a forecast you will stop scoring.
Call center forecasting software and WFM tools
Forecasting tools fall into three tiers, and the honest answer for many teams is that the cheapest tier is fine.
| Tier | Examples | Indicative price | Best for |
|---|---|---|---|
| Spreadsheet plus an Erlang calculator | Excel or Google Sheets, free online Erlang C calculators | Effectively zero | Under roughly 25 agents, one or two channels |
| Modern standalone WFM | Assembled, injixo, Community WFM | Assembled lists from about $250 per month; per-agent plans are common in this tier | 25 to 200 agents wanting automated forecasting and scheduling without a suite |
| Enterprise WFM suites | NICE CXone Mpower, Verint, Calabrio | NICE CXone Mpower listed from about $71 per agent per month; Calabrio from about $75 per agent per month with volume discounts reported far lower; Verint from about $1,500 per month | Large multi-site, multi-skill operations needing intraday management and compliance |
Published list prices in this category are a starting point, not a quote. Enterprise WFM is almost always negotiated on agent count, contract length and module mix, and implementation, training and integration are billed separately. Confirm current pricing directly with the vendor before you build a business case on any of these figures.
When is a spreadsheet no longer enough?
Manual spreadsheet forecasting remains common, and it is one of the clearest single sources of avoidable error. That does not make it wrong for everybody. It becomes the constraint at a fairly predictable set of thresholds.
- You are running more than two or three channels, and blending them by hand has become a weekly project.
- You have multi-skill agents, where a single Erlang calculation stops describing reality.
- The forecast takes a person more than half a day a week to produce and refresh.
- You need intraday re-forecasting, not just a Monday morning plan.
- Nobody has scored forecast accuracy in months, because the process is too manual to leave room for it.
The business case is straightforward once you know your cost per ticket and your loaded agent cost. Over-forecasting means paying for idle capacity. Under-forecasting means service level misses, overtime, and attrition from a floor that is permanently underwater. Price both against the licence cost rather than arguing about features.
Common call center forecasting mistakes
A handful of errors account for most bad forecasts, and none of them are about model choice.
- Forecasting volume but not handle time. AHT drift is invisible and quietly rewrites your workload.
- Never scoring the forecast. If you do not compute MAPE and bias weekly, you have an opinion, not a forecast.
- Leaving anomalies in history. Last year's outage becomes this year's phantom peak.
- Forecasting at the wrong grain. A monthly number cannot produce a schedule.
- Adding shrinkage to the forecast. Shrinkage belongs in the staffing calculation, applied as a divisor, not bolted onto volume.
- Ignoring the business calendar. Billing runs, dunning cycles, renewals and marketing sends are all knowable in advance and all generate contacts.
The last one deserves emphasis for anyone running customer experience operations in a back office. A large share of inbound contact volume in billing-heavy businesses is self-inflicted and scheduled. If invoices go out on the first of the month, the questions arrive on the second. That is not a forecasting problem, it is a calendar you already have.
Frequently asked questions
What is the formula for call center forecasting?
Workload equals forecast contact volume multiplied by average handle time. Volume itself is usually built as a base trend multiplied by seasonal, day-of-week and intraday indices, then adjusted for known events. Workload is then converted to agents with a queuing model such as Erlang C, and finally divided by one minus shrinkage.
How far in advance should you forecast call volume?
Run three horizons in parallel. An annual or quarterly forecast drives budget and hiring, a four to six week forecast drives schedules and shift bids, and an intraday forecast drives same-day reassignment. They use the same history but serve different decisions, and accuracy expectations should fall as the horizon shortens.
What is a good MAPE for call center forecasting?
At the daily level, 5 to 8 percent MAPE is a strong result for an established voice queue and 10 to 12 percent is normal for digital channels. Interval-level MAPE is routinely two to three times higher. Track bias alongside MAPE, because a small average error that always leans one direction is worse operationally than a larger, unbiased one.
Is Erlang C a forecasting method?
No. Erlang C is a queuing formula that converts an existing volume and handle time forecast into the number of agents needed to hit a service level. It assumes no caller abandons. Erlang A extends it with an abandonment rate and generally recommends slightly fewer agents, which is usually closer to real behavior.
Can you forecast chat and email the same way as calls?
The volume forecast works the same way, but the staffing conversion does not. Email is deferrable work sized against a backlog and a turnaround target, and chat involves concurrency, so an agent handles several conversations at once. Erlang C describes neither well. Forecast the workload identically, then staff each channel with its own model.
How often should the forecast be updated?
Refresh weekly for the scheduling horizon and re-score last week's accuracy in the same sitting, so the two never drift apart. Rebuild seasonal indices quarterly, and revisit the long-range forecast whenever the business changes something that drives contacts: a launch, a migration, a pricing change, or a shift in ticket deflection.