Webinar on call center forecasting: Why a forecast accuracy of 100% is sheer luck March 20, 2020 Ger Koole, PhD
Webinar on call center forecasting:Why a forecast accuracy of 100%
is sheer luckMarch 20, 2020Ger Koole, PhD
Why do we forecast?
• All WFM processes start withforecasting• Bad FC leads to:• More ad hoc decisions• (Structural) under or overstaffing• Higher costs• Bad service levels• Abandonments• Less sales• …
• Garbage in = garbage out Shor
t-ter
m
lo
ng-te
rm
Forecasting
Capa
city
plan
ning
Agen
t sc
hedu
ling
Real
-tim
e pe
rform
ance
m
anag
emen
t
budget
hiring decisions
What drives actuals?
• To be forecasted: actuals, handling times, sick leave, ….• Actuals are impacted by:• predictable intra-day, intra-week and intra-year fluctuations (seasonality)• (long-term) trend• events (holidays, marketing, etc.)• unpredictable noise
• Impact seasonality & repeating events can be accurately predicted• Short-term trend as well• How about noise? What is it??
Drilling down in granularityfrom intra-year … to intra-week … to intra-day … to intra-hour
Repeats itself: predictable
Repeats itself: predictable
Repeats itself: predictable
Unpredictable?
The origin of noise
• Flip a coin 20 times…• How many times heads?• Excel demo
• Flip a biased coin (success rate 0.0005%) 2M times…• How many times heads?
Binomial distribution Poisson distribution
Poisson noise
• Number of arrivals for many customers behaving independently• Unexplained part of volume• Explainable part depends on seasonality,
trend, events: can be forecasted• “Discovered” by S.D. Poisson• Quantifies noise• Which part of error comes from noise?
SD Poisson (1781-1840)
5% 15% 22% 22% 17% 10% 5% 2% 1% 0.3% 0.1%
Mean = 3
Measuring accuracy
• Error = Forecast – actual, measured over multiple periods• Period can be of different length: quarter, day, week, …
• Goal: reduce errors to a single number• “the overall daily FC error of March was 5%”
• One period should not compensate the other • Sum of errors = 0, accuracy = 100%? NO• Solution: use squares or absolute values • Measures based on squares are hard to interpret:
“the RMSE is 12.9”• Measure needs to be relative to size: “5%” instead of
“12.9”• Solution = WAPE = Weighted Absolute Percentage Error• WAPE = sum of absolute errors / sum of actuals =
(20+0+10) / (80+100+110) = 10.3%
actuals
forecast
error 20 0 -10
Squares 400 0 100
Abs values 20 0 10
Quantifying noise
• WAPE can be split in part due to FC error and noise• Which part is due to noise?• What is the noise of a perfect FC?• Answer: minimal APE = √ (2/(FC π))• E.g., √ (2/(100 π)) = 8%• √ (2/(1000 π)) = 2.5%• Smaller for big volumes
• Minimal WAPE can be estimated using actuals
WAPE = 23%
WAPE = 36%
Minimal WAPE = 25%
Consequences
• Small volumes cannot be predicted accurately• Always FC errors, minWAPE gives minimal error• Most relevant for interval-level forecasts• Smart method needed for intra-day FC = splines
• SL is equally hard to predict• Also SL ”errors” and/or real-time performance
management• But: Erlang C/X* or simulations take Poisson
fluctuations as input
“Strive for five” is often infeasible at interval-level
* see https://www.ccmath.com/online-calculators/
Learning more
• www.wfmfellowship.com• Online trainings for
workforce management• At you own time & place• Videos, exercises,
assignments • Supported by CCmath
professionals• Subscribe online or at
Forecasting in practice
• CCforecast by CCmath• Cloud-based call center
forecasting tool• Own dedicated CCmath
algorithms• Automatic parameter setting• Any volume, any number of lines• Consistently outperforms Excel• Reduces time up to 90%
Next week
• Why is Erlang C unsuitable for call center safety staffing?• Friday March 27, 16:00 AMS time
• Questions? Chat or voice