Forecasting the wrong thing — granularity, metric, and channel
Technically right, operationally useless
A forecast can be technically right and still useless if it’s answering the wrong question. This is a different failure mode to bad data or a wrong algorithm. The model is sound; the data is clean; the methodology is defensible. But what the planner forecast and what the operation needs to schedule against are different things, and the gap costs the operation in SL, cost, or both.
Four versions of this failure recur across operations. Each is a category error that’s easy to slip into and surprisingly hard to spot once it’s embedded.
The wrong channel
The operation handled 80% voice and 20% chat two years ago. It now handles 60% voice and 40% chat. The forecasting team still forecasts voice, then projects chat as a fixed proportion, on the assumption that the underlying customer demand is what’s being modelled. It isn’t. Customers are shifting channel, the chat AHT is half the voice AHT but concurrency means agent-time-per-contact is similar, and the staffing implications of getting the channel split right are larger than getting the total volume right.
The fix: forecast each channel separately as a demand stream, and forecast the channel split as its own series. Most operations find this exposes an under-modelled trend toward chat that the voice-only forecast was hiding.
The wrong metric
Gross dialled volume includes everyone who phoned. Net contacts offered is what reaches the agent queue after IVR self-service and abandonment. They’re different numbers and they trend differently. The IVR containment rate isn’t stable — it depends on customer mix, IVR design, and the most recent customer-comms decision. A forecast of gross volume that’s right won’t produce the right staffing answer if the IVR containment rate has shifted.
The fix: forecast at the level the operation actually staffs against, which for voice is almost always net contacts offered (NCO). Track gross volume separately as an input to the IVR containment conversation. Don’t conflate them in the forecasting pipeline.
The wrong granularity
The capacity plan operates monthly. The schedule operates weekly. The intraday operates at 15- or 30-minute intervals. A forecast of monthly total is fine for capacity planning and useless for scheduling. A forecast of daily total is fine for scheduling and useless for intraday. A forecast that produces only a daily number and then assumes a fixed intraday curve is hiding the variance that actually matters for SL.
The fix: forecast at the granularity the downstream decision needs, not the granularity that’s easiest. The intraday curve isn’t a constant — it shifts with channel mix, seasonality, and operational changes, and forecasting it explicitly produces a 1–3 point SL gain over assuming a fixed pattern.
The wrong segmentation
The forecast is per-queue: sales, service, billing. The routing pools them through a skills-based rule that sends contacts wherever there’s capacity. The queue-level forecast becomes irrelevant once the router starts borrowing across queues, and the planner is left with a forecast that doesn’t match how the operation actually distributes work. Either segment the forecast to match the routing logic, or forecast at the pool level and let the operation handle the within-pool distribution.
The reverse trap is also real: an operation that runs single-skill queues but forecasts pooled volume produces an answer that over-counts capacity. The diagnostic question: does the forecast you produce match the level at which contacts are actually routed and staffed?
The conversation that prevents this
The single most useful question to ask at the start of any forecast assignment is: “what decision is this forecast going to support?” The answer determines the channel, metric, granularity, and segmentation the forecast needs to produce. Most operations skip this question and inherit the forecast structure from the previous planner, often without checking that the structure still fits the operation. The conversation takes fifteen minutes and prevents a year of producing the wrong answer.
How to test whether your current forecast is right
Four diagnostic questions. Channel: Does your forecast carry chat, email, social, and any other channel the operation handles, each as its own series? Metric: Does your forecast use the same volume number the operation staffs against (typically net contacts offered, not gross dialled)? Granularity: Do you produce forecasts at the granularity each downstream decision needs (monthly for capacity, weekly for schedule, interval for intraday)? Segmentation: Does the forecast match the level at which the operation actually routes work?
Two or more “no” answers means the forecast is solving the wrong problem and no amount of algorithm tuning will fix it.
Conclusion
The right forecast of the wrong thing is one of the most consistent ways forecasting goes wrong in contact centres. The fix is the conversation at the start — what decision is this for, at what level, in what units — rather than the methodology in the middle. Operations that ask this question on day one of every forecast assignment avoid the failure. Operations that inherit the previous planner’s structure typically discover the mismatch only when something visibly breaks.
Next in the series: The point estimate and the false-certainty trap.
Pair this with demand decomposition, hierarchical reconciliation, and capacity planning when the mix is changing.