Industry: Consumer Lending Capabilities: Randomized Control Trials
Optimizing Collections Performance Through Experimentation
How we tested our segmentation and right-time-to-call recommendations in a collections call centre to realize 20% better outcome with 15% reduced effort
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 client had outperformed its cumulative gross loss forecast through most of 2025, but late-year trends put that outperformance at risk, with the potential to push 2026 charge-offs meaningfully above plan. A diagnostic of the collections operation showed why.
Rather than rolling these changes out as the new default strategy, we determined gathering evidence would actually improve outcomes before committing operational capacity to them. Our recommendation, accepted by our client, was to design and run a rigorous, statistically sound test that could isolate the impact of each intervention against business-as-usual (BAU) performance.
The test had two specific objectives:
1. Demonstrate that calling customers at their optimal time slot outperforms random time-slot assignment.
2. Confirm that calling low ability-to-pay customers less frequently does not worsen collections outcomes.
The Solution
We designed a three-arm randomized control trial (RCT), a methodology chosen specifically because it isolates causal impact — removing the ambiguity that comes from simply comparing before-and-after performance or relying on self-selected pilot groups.
Scoring and assignment. Every customer in the portfolio was scored on two dimensions
1. A right-time-to-call score, allocating each customer to one of six two-hour calling windows (8–10am through to 6–8pm), based on a predictive model of contact likelihood.
2. An ability-to-pay score, derived from the customer’s latest credit bureau data.
Each customer was then assigned a random number from 1–100, which determined their trial arm — ensuring group allocation was independent of any customer characteristic and free from selection bias.
Trial design. The portfolio was split into three groups:
| Group | Random number range | Treatment |
|---|---|---|
| A — Control | 1–20 | BAU: calls placed at a random time slot |
| B — Right-time-to-call | 21–40 | Calls placed using the customer’s assigned optimal slot |
| C — Right-time-to-call + Ability-to-Pay | 41–100 | Calls placed at optimal slot, with low ability-to-pay customers suppressed from calls two days a week |
Group C was deliberately sized larger than A and B, reflecting the client’s intent to gather sufficient evidence on the combined strategy — the configuration expected to go into production if both hypotheses were confirmed — while keeping enough customers in the smaller arms to detect a clear effect of right-time-to-call in isolation.
Implementation
Dialer technology could not route leads to agents by precise time-of-day. We adapted the design to a workable proxy — pacing dialing volume per hour (10–12 calls/hour) rather than assigning literal calling windows — to approximate the intended effect within the existing infrastructure.
Ability-to-pay scoring used a simplified rules-based version of the scorecard rather than the full predictive model, given data and timeline constraints at the point of test launch.
Both adaptations were treated as interim measures: we flagged that a full predictive-model-based version of the ability-to-pay segmentation would likely yield stronger results than the simplified version tested.
Results
We tracked what proportion of accounts that were 60–120 days delinquent at the start of the test rolled forward into worse delinquency buckets:
| Test Group | Rolled back: 0 to 60 days | Stayed constant: 61 to 120 days | Rolled Forward: 120+ days |
|---|---|---|---|
| BAU (no change) | 13% | 28% | 57% |
| Right-time-to-call only | 19% | 33% | 46% |
| Right-time-to-call + Ability-to-Pay | 18% | 32% | 49% |
Key findings:
Right-time-to-call delivered a clear, measurable improvement. Accounts receiving calls at their optimal time slot were substantially less likely to roll into the most severe delinquency bucket (120+ days) — 46% versus 57% under BAU — with a corresponding shift into earlier-stage, more recoverable buckets. This confirmed the first test hypothesis decisively.
Suppressing low-ability-to-pay customers did not materially worsen outcomes. Group C’s results tracked closely with Group B despite roughly 40% fewer contact attempts on the lowest ability-to-pay segment, supporting the hypothesis that reduced-frequency calling can be applied to this segment without a meaningful cost to collections performance.
Together, the results validated a strategy that improves recovery rates while also reducing total call volume and cost-to-collect on segments least likely to pay.
Implications:
Together, the results validated a strategy that improves recovery rates while also reducing total call volume and cost-to-collect on segments least likely to pay.
More broadly, the engagement illustrates a principle we bring to every optimisation initiative: before committing operational scale to an algorithmic change, test it. A well-designed control trial turns “we think this will work” into “we know this works, and by how much” — giving stakeholders a defensible basis for the investment.