planning workforce management for bank operation centers with neural networks
TRANSCRIPT
Planning Workforce Managament for Bank Operation Centers with Neural Networks
Sefik Ilkin Serengil
joint work with Alper Ozpinar
AIKED Conference Venice, Italy
January 29, 2016
p.3 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Talk Outline
1. Operation Centers
2. Problems
3. Optimization Objective
4. Motivation
5. Results
6. Proposed Method
7. Conclusion
p.4 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Money Transfer Orders
• Customers still tend to use bank branches
• 35% of bulk transactions tranmitted on branches
• Mostly commercial customers
• Faxing instruction, no need to be situated at branch
• Branch employees validate the signature
• Scan and deliver instruction to OC
p.5 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Money Transfer Orders #2
• Could include multiple transactions (15% bulk rate)
• Large amount (Avg 27K USD per transaction)
• 10M count money transfer order (50% of all)
• 16M count money transfer transactions
• Branch operations distribution for last 16 months
p.6 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Operation Centers
• Serve to reduce operational workload of branches
• Centralized management, expert employees
• Offering faster, high quality service
• High turnover rate (e.g. 50-300 employees)
• Digitalizing the hard copy instruction
• Commit the transaction
p.7 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Problems
• OC Managers predict workload by experience
• Planning the workforce manually
• Rescheduling when density is observed
• Deadline is strictly defined by Government (5.00 pm)
• Service Level Aggrement (90 minutes)
• Delays cause to suffer customers
p.8 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Problems #2
• Insufficient employee reservation is clearly seen
• Y-axis: Total work and reserved employee ratio
• X-axis: Work hours
p.9 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Optimization Objective
• Proper and efficient employee planning
• Preventing excess employee reservation for low transaction volume
• Avoiding insufficient employee reservation for high transaction volume
• Machine learning based workload prediction
• Workforce planning by considering employee skills
p.10 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Motivation
• Thought as machine learning problem
• A function is modeled by historical examples
• Function forecasts for un-known examples (y)
• Underfitting for simple complexity function
• Overfitting for too complex function
• Function should be derived from affecting factors (x)
Historical Data
ML Algorithm
Mathematical Functionx[] y – forecasting
p.11 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Affecting Factors vs Correlation
Factor Scale Correlation Co.
Hour [9, 17] 0.0500
Day [1, 31] -0.0557
Month [1, 12] 0.0048
Year [2012, 2016] -0.0767
Weekday [2: Monday, 6: Friday] 0.0728
Is first or last work day [0, 1] 0.1790
Is half day [0, 1] -0.0048
Transaction count (h-1) [-∞, +∞] 0.2114
Transaction count (h-2) [-∞, +∞] -0.0415
Transaction count (h-3) [-∞, +∞] 0.2666
Yearly deviation [-∞, +∞] 0.0388
• Potential Function Parameters
p.12 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Neural Networks
• Ability to learn, remember and predict
• Multiple inputs and an output
• Inputs (x) are involved in network through own weight
• Weight (w) specifies the strength of input on output
• Adjusting weight values implement learning
• Assembly function (∑) calculates net input (o)
• Activation function (f) computes the net output (y)
p.13 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Neural Network Model
• 3 layered network with node numbers 11, 8, 1
• 8 nodes in hidden layer acc. 2/3 rule (Heaton, 2000)
• Sigmoid for activation, Back-propagation for learning
p.14 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Workload Forecast Results
• Suppose x is prediction set, y is actual set
• Evaluation metric
• One day’s result for Dec 04, 2015
p.15 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Results #2
• A sample from long term results for 100 days
• Historical data obtained for last 4 years.
EFT MO
MAE 60.95 60.99
MAE / Mean 10.29% 15.19%
Correlation Co. 96.47% 93.04%
Mean 592.40 401.42
Instances (hour) 548 548
p.16 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Workforce Planning
• Employee skill map for 2 months period
• X-axis: unit perform time in seconds
• Y-axis: Average completed work count on a hour
• PN: Expected transaction count (NN result)
• PQ: Transactions waiting on queue
p.17 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Conclusion
• An approach introduced to plan workforce
• Based on a machine learning discipline
• Simulated for EFT and Money Order
• Satisfactory results for workload forecasting
• Workforce planning by considering skills
• Future work; workforce optimization on production
• Thought to be applied in turnover requiring areas
p.18 / 18Sefik Ilkin Serengil AIKED Venice, January 2016
Acknowledgements
• Conducted by SoftTech under project number 5059.
• Supported by TEYDEB (Technology and InnovationFunding Programs Directorate ) of
• TUBITAK (The Scientific and Technological ResearchCouncil of Turkey)
• In scope of Industrial Research and DevelopmentProjects Grant Program (1501)
• Under the project number 3150070.