While everyone is quick to jump onto the Machine Learning trend, is it really safe to implement within the financial services sector with so many issues surrounding the regulatory and ethical side of utilizing machines to make human decisions?
Overcoming the issues faced when explaining outcomes that may be discriminatory which can damage a company’s reputation
Is Machine Learning really needed to automate financial processes or does the negativity around ethical considerations enough to reconsider?
Can regulatory bodies ever be confident enough in the decisions made by the machines to allow ML to really progress in financial services?
Looking towards ensuring transparency in the models decision making process to determine if it is suitable for deployment in financial decisions
1. Debate
Is Machine Learning Mature
Enough to Successfully
Implement in Financial
Institutions
Hernan Huwyler
2. Prerequisites for implementing machine learning
• Business case with individual processes to automatize
• Hardware to support extensive computations
• Known business rules to develop the learning
• Abundant normalized quality data with correct labels and
noise-free dataset to train
• Structured documents and forms to extract data
• Technical staff capabilities to create and maintain models
• Security control protocols and data governance policies
• Decisions are based on data
• Culture to promote innovation and experimentation
3. Prerequisites for implementing machine learning
Input
• Collection
• Validation
• Cleansing
• Validation
Training
• Labelling
• Training
• Tuning
• Scoring
Deployment
• Scaling
• Testing
• Tuning
• Versioning
Execution
• Setting
• Training
• Validation
• Monitoring
Code
• Scripts
• Artifacts
Data sources
• Training
• Metadata
Configuration
• Automatic retraining
• Continuous delivery
4. Tips before implementing machine learning
• Clear strategy and cases to monetize data
• Set measurable goals to reduce costs or increase revenue
• Align the requirements with the business and IT
• Involve data owners and subject matter experts in sales,
marketing, finance, human resources and operations
• Use an agile approach with pilots
• Have a data cleansing project before testing
• Communicate to users how to use the insights provided by
machine learning
• Invest in technical skills
• Learn from deviations between model predictions versus
actual outputs
5. Requirements for planning
• Data integrity of the input data > acceptable cost and
quality of data by internal and external providers
• Model accuracy and performance > acceptable level of
noise by developers
• Quality evaluation > acceptable validations of outputs by
testers and assurance specialists
• Process flexibility > acceptable level of interactions,
updates, skills, and scalability by stakeholders
• Customer expectations > acceptable adoption and
decision-making by users
6. Potential risk events
• Missing or inaccurate data to develop training or scale
• Unacceptable false positives and negatives ratios
• Slow or partial adoption of machine learning developments
• Insights not actionable for users
• Behaviors and decisions are not impacted
• Constant model adjustments
• Compliance breaches (particularly in using personal data)
• Potential customers discrimination
• Data invisibles exclusion
• Unpredictable requirements changes