Two metrics, two questions
Forecast accuracy discussions go wrong when everyone quotes a different number. There are two measures that matter in workforce management, and they answer different questions.
MAPE, Mean Absolute Percentage Error, answers "how wrong were we, on average, interval by interval?" For each period take the absolute gap between forecast and actual, divide by the actual, then average those percentages across all periods.
MAPE = average( |actual - forecast| / actual ) x 100
WFA, Weighted Forecast Accuracy, answers "how wrong were we across the whole volume?" It weights each period by its size, so a busy Monday morning counts more than a quiet Sunday evening.
WFA = 100 - ( sum|actual - forecast| / sum(actual) ) x 100
A worked example makes the difference obvious. Two intervals: interval one forecast 1,000 against actual 1,100, interval two forecast 20 against actual 10. MAPE is the average of 9.1% and 100%, which is 54.5%, a catastrophic looking number. WFA is 100 minus (100 plus 10) divided by 1,110, which is roughly 90.1%. The forecast was good where it mattered and terrible where it did not.
Neither number is lying. MAPE is telling you the small intervals are unpredictable. WFA is telling you the staffing implication was modest. Report both, and never let anyone compare your MAPE to another site's WFA.
How to measure honestly
Always hold data back. Fit the model on older periods, then score it on periods it never saw. Scoring a model on the data it was fitted to always flatters it, which is how teams end up trusting a forecast that collapses in production.
Match the grain to the decision. Monthly accuracy justifies hiring. Weekly accuracy drives scheduling. Interval accuracy drives intraday. A site with 4% monthly MAPE can still be unschedulable if its interval distribution is wrong, because the total is right and the shape is not.
Never forecast handled volume. Handled volume is capped by your own staffing, so a forecast fitted to it learns your past under staffing and quietly repeats it. Forecast offered contacts.
Report bias separately. Take the signed error, not the absolute one, and average it. MAPE of 8% with a plus 7% bias means you are systematically under forecasting, which is a fixable problem. MAPE of 8% with near zero bias is just noise.
Rough expectations
Accuracy targets are context specific, but as a starting point: monthly volume within 5%, weekly within 8%, daily within 10%, and interval level MAPE of 15 to 25% is normal even in a well run operation. Anyone claiming 3% interval accuracy is either measuring something else or has extraordinarily stable demand.
What actually improves accuracy
Clean the history first. One outage week or a mis coded queue drags a trend line off course for months. Outlier correction is usually worth more than any change of model.
Separate the drivers from the baseline. Marketing campaigns, billing cycles, product launches and public holidays are not noise, they are known events. Model the baseline statistically and add the events explicitly, with a size taken from the last comparable event.
Blend models. Averaging your best two or three methods reliably beats picking one. The errors of different models tend to cancel out. This is the least glamorous and most dependable improvement available.
Rebuild the intraday curve by day type. Monday morning does not look like Saturday morning. Separate curves by day type, and refresh them quarterly, because arrival patterns move as customers move channels.
Feed intraday learnings back. Every large interval miss should carry a one line cause. After a month those lines will point at the same two or three structural issues, and fixing those beats any amount of parameter tuning.
Make the review a habit
Run a monthly accuracy review with the same three slides every time: accuracy and bias by grain, the largest five misses with causes, and the actions carried into next month. Change one thing at a time so you can tell what worked. Forecast accuracy is not won by a clever model, it is won by a boring routine.