stephen curiskis - actionable insights through customer analytics and data science - futuredata 2017
TRANSCRIPT
Actionable insightsthrough customer analytics and data science
Stephan Curiskis, Lead Data ScientistTyro Payments
Tyro Payments
• Founded in 2003
• Grown to 19,000 customers, $42 billion in transactions, 371 staff
• Granted a banking license in 2015, becoming an Authorised Deposit‐taking Institution (ADI)
Three products
Payments
Deposits
Lending
• Pricing• Transaction forecasting• Fraud• Marketing optimisation
• Fraud• Liquidity forecasting• Marketing optimisation
• Credit risk• Liquidity forecasting• Marketing optimisation
Product insights and new feature development
Three products
Payments
Deposits
Lending
• Pricing• Transaction forecasting• Fraud• Marketing optimisation
• Fraud• Liquidity forecasting• Marketing optimisation
• Credit risk• Liquidity forecasting• Marketing optimisation
Product insights and new feature development
Lending product
Loan AmountLoan
Amount
Interest RateInterest RateDurationDuration
Lending product
Loan AmountLoan
Amount
One‐off feeOne‐off feeDuration
Deduct a % of transaction value periodically until the loan is paid customer specifies %
Calculate a one‐off fee for a target annual percentage rate
Modelling transaction stability and default risk is critical
Customer Example
Data to model to decision
Transactions
Transaction Forecast
Credit Risk Model
Address
Industry
Business details
Strong seasonality for each customer
Seasonality Modelling
ABS + Census Data
BehaviouralClustering
Decision making based on model outputs
Address not useful on its own
Large data becomes unwieldly
“Data is our strategic asset”
Only when we can extract actionable insights. Data first needs to be cleaned, structured, enriched, then modelled.
Location data
• Address matched to other sources is powerful. E.g. Census – Challenge: how to make sense of so much data
• Clustering is useful to find patterns in location for a customer base– Take a large number of variables (Census and internal data)– Find what data points occur together frequently, or are correlated– Assign them to a group– We now have a manageable and rich data set
Seasonality modelling
Requires two years of data for each customer
Behavioural clustering
Seasonality
LocationIndustry
1. Train seasonality models where we can (2 years of data)
2. Cluster customers by seasonality
3. Predict seasonality cluster by Industry and Location groups
4. Infer the trend
We now know the amount to lend
Additional data sources
• How do consumers choose which café to go to for lunch?
Reviews
Ratings
Popular times
More reviewsratings
Articles and blogs
Social media
Signal and noise
We can get a lot of data this way, but how much is really useful?
• Need to ensure it is reliable• Must be collected over time (can take many months)• Machine learning is key to determine useful variables• Constructing reliable outcomes is critical. E.g.
– Rapid growth– Fraud– Bankruptcy
Summing up
• Majority of work in analytics and data science is in creating useful and predictive data sets
• Data enrichment is key – create long term plans for data assets from many sources
• Machine learning is critical to separate the signal from the noise– Ensure that outcomes of interest are recorded accurately