A key challenge for call centres remains the forecasting of high frequency call arrivals collected in hourly or shorter time buckets. These forecasts are required for decisions concerning the scheduling, hiring and training of staff. In addition to the high frequency nature of call arrival series and the complex seasonal patterns, including the multiple seasonal cycles, call arrival data often contain a large number of anomalies, driven by holidays, special events, promotional activities and system failures. This study presents an approach based on artificial neural networks (ANNs) for forecasting intraday call arrivals. In so doing, we empirically evaluate alternative methodologies for modelling and forecasting outliers in high frequency data, which span over several periods, addressing a gap in research of practical significance considering the difficulty and the cost associated with manual exploration and treatment of such data. We assess the performance of different ANN modelling methodologies in terms of the accuracy with which normal and outlying periods are modelled. Multi-period outliers are modelled using alternative encodings ranging from binary dummy variables to functional profiles, as well as segmenting the series to separate it into outlying and normal days. Results show that ANNs outperform conventional benchmarks and are capable of modelling high frequency outliers using relatively simple outlier modelling approaches.
Conf42-LLM_Adding Generative AI to Real-Time Streaming Pipelines
Forecasting Intraday Arrivals at a Call Centre using Neural Networks: Forecasting Anomalous Days
1. Forecasting Intraday Call
Arrivals
Modelling Special Days
Devon K. Barrow
Nikolaos Kourentzes
27th European Conference on Operational Research
12-15 July 2015
University of Strathclyde
2. 1. Research Questions
2. Call Arrival Data and
3. Experimental Design
4. Results
5. Conclusions and future work
Outline
Forecasting Intraday Call Arrivals Outline 2
3. • How do different forecasting methods
perform when forecasting (high frequency)
call centre arrivals?
• What is the impact of coding or not coding
of (functional) outlying periods on
forecasting performance?
Research Questions
Research Questions 3Forecasting Intraday Call Arrivals
4. 1. Research Questions
2. Call Arrival Data and
3. Experimental Design
4. Results
5. Conclusions and future work
Outline
Forecasting Intraday Call Arrivals Call Centre Data 4
5. • High dimensional and sampled at a high
frequency (Kourentzes and Crone, 2010).
• Complex seasonal patterns
– Intraday
– Intraweek
– interyear dependencies(De Livera et al., 2011)
Call Arrival Data and Challenges
Challenges
5Call Centre DataForecasting Intraday Call Arrivals
6. Call Arrival Data and Challenges
Some call centre data
6
• Two series from a large UK service provider call
centre
• A leading entertainment company in Europe
• Data is sampled at half-hourly intervals
• Consists of 103 weeks and 3 days from 29 June 2012
to 23 June 2014 inclusive including bank holidays
and weekends.
Call Centre DataForecasting Intraday Call Arrivals
7. Call Arrival Data and Challenges
Complex seasonal patterns
7
Intraday Intraweek
Mean
Median
Monday
Intrayear
Call Centre DataForecasting Intraday Call Arrivals
8. • Development of new methods
– Double seasonal exponential smoothing (Taylor, 2008)
– Multiplicative double seasonal ARMA model (Taylor, 2003)
– Exponential weighting (Taylor, 2008, 2010)
– Regression (Tych et al., 2002; Taylor, 2010)
– Singular vector decomposition (Shen and Huang, 2005, 2008;
Shen, 2009)
– Gaussian linear mixed-effects models (Aldor-Noiman et al.,
2009; Ibrahim and L’Ecuyer, 2013)
• Intraweek seasonal moving average performs well
at medium to long horizons (Tandberg et al. 1995; Taylor, 2008; Taylor,
2010; Ibrahim and L’Ecuyer, 2013)
Call Arrival Data and Challenges
Handling high frequency complex seasonality
8Call Centre DataForecasting Intraday Call Arrivals
!
9. • Data is context sensitive
– Effects of holidays, special events and
promotional activities
• Prone to quite sizeable unexplained
variations (outliers)
– E.g. due to system failures and data
processing
Call Arrival Data and Challenges
Challenges
9Call Centre DataForecasting Intraday Call Arrivals
10. Call Arrival Data and Challenges
Anomalies (outliers)
10
Series 1
Series 1I
Call Centre DataForecasting Intraday Call Arrivals
13. • Approaches to handling ‘special days’:
– Information is either available and/or data is pre-
cleansed
– The forecaster has an external methodology for
tackling ‘special days’ (Jongbloed and Koole, 2001; Avramidis et al.,
2004; Taylor, 2008a; Pacheco et al., 2009; Taylor, 2010b)
– Removing such days altogether (Taylor et al. 2006)
– Singular vector decomposition for automatic outlier
detection (Shen and Huang, 2005)
• Kourentzes (2011) demonstrates that there are
substantial accuracy benefits to be had from
modelling irregular load patterns.
Call Arrival Data and Challenges
Handling ‘special days’
14Call Centre DataForecasting Intraday Call Arrivals
!
14. 1. Research Questions
2. Call Arrival Data and
3. Experimental Design
4. Results
5. Conclusions and future work
Outline
Forecasting Intraday Call Arrivals Experimental Design 15
15. Experimental Design
Functional outlier modelling
16
• We evaluate seven alternative methodologies for
modelling functional outliers
• These are based on extensions of conventional
outlier modelling and novel approaches
• We assume that the data generating process of
normal observations is captured adequately and
the outliers are already labelled
Control
Single Binary
Dummy
Multiple Binary
Dummy
Single Integer
Profile Dummy
Trigonometric
Dummy
Model
Separately
Experimental DesignForecasting Intraday Call Arrivals
16. Experimental Design
Functional outlier modelling
17
• Control/benchmark
– A set of autoregressive lagged inputs (past values) are
identified using stepwise regression
– Outliers are not modelled
• Single Binary Dummy Variable
– Indicator variable s.t. one = outlier; zero otherwise
• Multiple Binary Dummy Variable
– S is the seasonal length (S=48)
– 48 (S) dummy variables to code each observation
– 47 (S – 1) dummy variables to code each observation
– Stepwise selection, s = {1,…,S=48} s.t. s is significantly
different from normal observations
Experimental DesignForecasting Intraday Call Arrivals
17. Experimental Design
Functional outlier modelling
18
• Single Integer Dummy Variable
– Monotonically increasing variable from 1 to 48 if outlier; zero
otherwise
• Profile Dummy Variable
– This variable is equal to the profile if there is an outlier; zero
otherwise
• Trigonometric Dummy Variables
– Sine and cosine if there is an outlier; zero otherwise
• Model Separately
– Create a new series containing only outliers
– Replace outliers in original series with normal observations
Experimental DesignForecasting Intraday Call Arrivals
18. Experimental Design
Experimental setup
Experimental design 19
• Neural network setup
– Inputs based on backwards regression + dummies +
seasonal dummies.
– Mode ensemble of 50 networks trained by scaled
conjugate gradient descent
– Hidden nodes identified experimentally for each set of
inputs
• Forecast creation
– Forecast horizon is set to 1 day ahead (1-48 half hourly
steps ahead)
– Test set of 100 days (4800 data points x 48 forecasted
horizons)
• Forecast evaluation
– Mean Absolute Error (MAE)
– Relative Mean Absolute Error (RMAE) i.e. MAE_{Method} /
MAE_{Control}
AE = |Yt - Ft|
Forecasting Intraday Call Arrivals
19. 1. Research Questions
2. Call Arrival Data and
3. Experimental Design
4. Results
5. Conclusions and future work
Outline
Forecasting Intraday Call Arrivals Results 20
20. Results
Neural networks versus benchmarks
21
Time Series 1 Overall Outlier Normal
Naïve 1 3.930 2.711 4.269
Naïve Day 1.417 1.211 1.475
Naïve Week 1.333 1.057 1.410
MA Day 1.238 1.278 1.226
MA Week 1.233 0.845 1.341
ETS Day 1.887 1.438 2.013
ETS Week 1.529 1.165 1.630
ETS Double 1.086 0.857 1.149
MAPA Day 1.522 1.362 1.566
MAPA Week 1.484 1.274 1.542
NN Control 1.000 1.000 1.000
ResultsForecasting Intraday Call Arrivals
21. Results
Neural networks versus benchmarks
Time Series II Overall Outlier Normal
Naïve 1 2.236 1.770 2.500
Naïve Day 1.415 1.119 1.582
Naïve Week 1.433 1.157 1.590
MA Day 1.415 1.119 1.582
MA Week 1.129 0.973 1.217
ETS Day 1.600 1.368 1.731
ETS Week 1.479 1.288 1.587
ETS Double 1.151 0.990 1.242
MAPA Day 1.347 1.242 1.406
MAPA Week 1.280 1.201 1.325
NN Control 1.000 1.000 1.000
22ResultsForecasting Intraday Call Arrivals
22. Results
Outlier modelling versus control
Time Series 1 Overall Outlier Normal
NN Control 1.000 1.000 1.000
NN Bin1 0.996 0.933 1.013
NN BinS 0.884 0.822 0.901
NN BinS-1 0.873 0.831 0.885
NN Bin Step 0.897 0.858 0.908
NN Bin Back 0.895 0.850 0.907
NN Int 0.935 0.913 0.941
NN SinCos 0.980 0.865 1.012
NN Profile 0.986 0.953 0.995
NN Replace 1.002 1.019 0.997
23ResultsForecasting Intraday Call Arrivals
23. Results
Outlier modelling versus control
Time Series II Overall Outlier Normal
NN Control 1.000 1.000 1.000
NN Bin1 0.956 0.930 0.970
NN BinS 0.922 0.866 0.954
NN BinS-1 0.923 0.870 0.954
NN Bin Step 0.929 0.876 0.958
NN Bin Back 0.931 0.875 0.963
NN Int 0.957 0.940 0.966
NN SinCos 0.951 0.912 0.972
NN Profile 0.963 0.936 0.979
NN Replace 1.030 1.090 0.996
Results 24Forecasting Intraday Call Arrivals
24. 1. Research Questions
2. Call Arrival Data and
3. Experimental Design
4. Results
5. Conclusions and future work
Outline
Forecasting Intraday Call Arrivals Conclusion and future work 25
25. Conclusion and future work
Conclusion and future work 26
• Conclusion
– Neural networks are good for call centre
data for two reasons:
• They can do complex structures
• They can do complex outliers with relatively
simple modelling
• Future work
– Automatic functional outlier detection
and modelling for call centre arrival data
Forecasting Intraday Call Arrivals
26. Devon K. Barrow
Coventry Business School
Coventry University, Priory Street, Coventry, CV1
5FB
Direct line: + 44 024 7765 7413
Skype: devon.k.barrow
Email: devon.barrow@coventry.ac.uk