Volume Forecasting and Interval Distribution

Forecasting in workforce management is a three step job. First you predict how much work arrives in a period, a month, a week or a day. Then you split that total across the days of the week. Then you split each day across the intervals people actually work. Get any one of the three wrong and the staffing numbers that follow are wrong too, even if the total looks fine.

The nine methods, in plain English

• Naive simply repeats the last period. It is the honest baseline: if a fancy model cannot beat it, the model is not earning its keep.
• Simple Moving Average averages the last few periods. Good for noisy, flat demand, always a step behind a real trend.
• Weighted Moving Average uses the same window but leans on recent periods, so it turns the corner faster after a step change.
• Simple Exponential Smoothing gives every past period a weight that fades geometrically. One dial, alpha, controls how reactive it is.
• Holt's Linear Trend adds a second dial for trend, so a steadily growing queue keeps growing in the forecast instead of flattening out.
• Holt-Winters adds a third dial for seasonality. It needs at least two full cycles of history, twenty four months for a monthly seasonal pattern.
• Linear Regression fits a straight line through everything you have. Simple, stable, and blind to seasonality.
• Year-over-Year Growth takes the same period last cycle and applies the growth rate you have actually seen. Popular with finance for a reason.
• Seasonal Index removes the seasonal shape, fits a trend on the clean series, then reapplies the seasonal factor. Often the best of both worlds.

The Weighted Blend then averages the three methods that scored best on your own data. Blending is not a gimmick, it is one of the most reliable ways to reduce error, because the mistakes of different models often cancel each other out.

What MAPE means and why we hold data back

MAPE is Mean Absolute Percentage Error. For each period you take the gap between what was forecast and what actually happened, express it as a percentage of the actual, and average those percentages. A MAPE of 8% means the forecast was typically 8% off, in either direction.

To make that number honest, this tool hides the most recent slice of your history, fits every method on the earlier data only, then scores each method against the periods it never saw. That is a holdout backtest. Scoring a model on data it was fitted to always flatters it, which is how teams end up trusting a forecast that falls apart in production. As a rough guide, under 10% MAPE is strong at monthly level, 10 to 20% is workable, and above 20% means you should be looking at the drivers behind the volume rather than tuning the maths.

Day of week and intraday breakdown

A monthly or weekly total is useless to a scheduler on its own. The day of week split turns it into seven daily totals, either from percentages you enter or from a pattern derived automatically out of your daily history. Most contact centres see a heavy Monday, a steady midweek, and a much lighter weekend, so the seven numbers rarely sit near an even 14.3% each.

The intraday curve then splits each day into intervals. The default shape here is the classic contact centre profile: a morning ramp as people start their day, a peak late morning, a dip around lunch, a smaller afternoon hump, and a taper into the evening. Every interval is editable, so you can paste in your own arrival pattern from your ACD. The rule that matters is simple: the percentages must add up to 100, or the daily total will silently drift.

Practical advice

• Clean your history first. One outage week can drag a trend line off course for months.
• Forecast contacts, not calls handled. Handled volume is capped by your own staffing.
• Re-run the backtest every month. The best method for your data changes over time.
• Keep the intraday curve separate by day type. Monday morning rarely looks like Saturday morning.