Industry: Consumer Lending Capabilities: Machine/Reinforcement Learning
Improving Collections through Intelligent Call Time Prediction
Our use of sophisticated self-learning algorithm to determine right-time-to-call customers caused a dramatic increase in contact rates leading to lower delinquencies and credit losses
Client Profile
Our client is a point-of-sale consumer loans fintech that experienced rapid, sustained growth in loan originations. At the time of engagement, the client was originating $100M in new loans monthly, managing a $1.8B loan portfolio, and absorbing roughly $30M in annual credit losses.
The Challenge
The lender’s collections teams operated under a deceptively simple strategy: randomize customer call times to prevent predictability. However, this approach came with hidden costs. Call center data revealed that customers have distinct preferences for when they’re available to speak—preferences driven by work schedules, family responsibilities, and personal routines. The current random-call model systematically ignored these patterns. More critically, the system failed to learn from contact outcomes. When a customer declined a call at 2pm, that signal was treated the same as successful connections at 9am: it was simply randomized again. The organization was essentially throwing away valuable behavioral data that could predict future contact success. The business impact was stark: despite substantial collections investments, contact rates stagnated, leading to upstream delinquencies and accelerated charge-offs.
The Solution
We designed a self-learning algorithm grounded in Thomson sampling—a Bayesian approach well-established in contextual bandits literature. Rather than treating each call time slot as equally viable, the system learns from every interaction.
How it works: For each customer, the algorithm maintains success-to-failure ratios for each call time slot (customizable; typically five 2-hour windows across business hours). These ratios are converted into probability distributions and corresponding cumulative scores. If a customer historically answers 60% of morning calls but only 20% of afternoon calls, the morning slot receives proportionally higher probability weight.
Consider a concrete example: Customer X has a 50% success rate for 8-10am calls. This slot receives 50% probability, with the remaining 50% distributed evenly across four alternative slots (12.5% each). The cumulative scores become: 8-10am (50%), 10am-12pm (62.5%), 12-2pm (75%), 2-4pm (87.5%), 4-6pm (100%). A random draw of 0.45 triggers an 8-10am call. If answered, the 8-10am probability increases; if not, it decreases.
Results and Business Impact
Pilot testing across a controlled subset of the portfolio delivered striking outcomes:
Contact Rate: 85% improvement in first-contact success versus randomized calling
Delinquency Reduction: 20% fewer customers progressing to higher delinquency buckets
Charge-off Impact: Full deployment is projected to reduce annual charge-offs by 8%—translating to $2.4 million in direct profit preservation
The success rate improvement compounded across the portfolio: higher first-contact rates meant fewer required retry attempts per customer, reducing operational costs while simultaneously improving customer satisfaction and collection velocity.
Key Takeaway
This engagement demonstrates that even within mature operations, sophisticated analytics can unlock hidden value. By respecting customer behavioral patterns and systematically learning from every interaction, organizations can dramatically improve contact efficiency—without relying on intrusive predictive modeling or customer friction. The algorithm improved business outcomes while enhancing the customer experience.