1. What We Observed
905
Loans Postponed
29d
Avg Extension
650
Avg Fido Score
LN 28
Avg Loan Number
Feature Eligibility (from codebase)
| Rule | Requirement |
|---|---|
| Fido Score | ≥ 500 |
| Loan state | ACTIVE (borrower is current, not in arrears) |
| Installments | Single installment only |
| Product | Not BNPL / Goods |
| Timing | Second half of loan term, up to 1 day before due date |
| Frequency | Max 1 postponement per client (lifetime) |
2. Aggregate DR: Postponed vs. Eligible Population
| Group | N | DR@DPD0 | DR@DPD7 | DR@DPD15 | DR@DPD30 |
|---|---|---|---|---|---|
| Postponed | 905 | 11.83% | 8.12% | 7.26% | 6.99% |
| Eligible, not postponed | 85,565 | 8.29% | 4.45% | 4.07% | 3.65% |
| Delta | +3.54pp | +3.67pp | +3.19pp | +3.34pp |
Selection bias warning: Borrowers who choose to postpone are self-selecting — they're more likely to be under cash-flow pressure than the average eligible borrower. Comparing these groups directly overstates postponement's negative impact.
3. Bounding the True Impact
We can't observe the counterfactual (what postponers would have done without the option), but we can bound it using two extreme assumptions:
| Assumption | Benchmark | Postponed DR@DPD30 | Benchmark DR@DPD30 | Implication |
|---|---|---|---|---|
| Best case All postponers would have defaulted |
DPD0 missers (same LN/FS bucket) |
5–7% | 35–56% | Feature saves ~30pp of defaults |
| Worst case All postponers would have paid on time |
Full eligible population (same LN/FS bucket) |
5–7% | 3–4% | Feature adds ~3pp of defaults |
The truth sits somewhere in this range. The key unknown: what share of postponers would have missed DPD0 without the option? If it's >10%, the feature is net positive on credit performance. Only an A/B test can answer this.
4. Segment-Level Detail
DR@DPD30 by LN and Fido Score bucket. Three benchmarks for each segment.
| Segment | Postponed N | Postponed DR | Eligible Pop DR | DPD0 Missers DR |
|---|---|---|---|---|
| LN21+ / FS 700+ | 350 | 5.3% | 2.6% | 37.4% |
| LN21+ / FS 600–699 | 160 | 4.6% | 3.9% | 34.7% |
| LN21+ / FS 500–599 | 86 | 7.1% | 5.9% | 37.2% |
| LN11–20 / FS 600–699 | 90 | 5.5% | 4.8% | 48.7% |
| LN11–20 / FS 500–599 | 142 | 14.0% | 5.9% | 55.6% |
| LN1–10 / FS 500–599 | 47 | 16.5% | 6.3% | 46.7% |
Pattern: High-tenure / high-score borrowers (LN21+, FS 600+) sit very close to the eligible-population benchmark, suggesting postponement does little harm in these segments. Lower-tenure segments show a wider gap, but also have the most to gain if the feature prevents defaults.
5. A/B Test Design
An experiment to measure the causal impact of postponement on default rate, repayment timing, and retention.
Hypothesis
Offering in-app loan postponement to eligible borrowers reduces net default rate (measured at DPD30) without materially increasing time-to-repayment for borrowers who would have paid on time.
Test & Control
Control (50%)
Postponement feature is hidden. Eligible borrowers see the standard loan page with no postponement option.
- No UI change
- No backend eligibility check surfaced
- Borrower experience identical to pre-launch
Treatment (50%)
Postponement feature is shown. Eligible borrowers see the postponement offer in the loan page (current behavior).
- Full postponement flow available
- All current eligibility rules apply
- No other changes
Randomization
| Parameter | Value | Rationale |
|---|---|---|
| Unit | Client ID | Ensures consistent experience across loans. Client-level avoids contamination if same person takes multiple loans. |
| Split | 50/50 | Maximizes statistical power. Given current adoption is ~1% of eligible loans, we need a large pool. |
| Stratification | LN bucket + Fido Score bucket | Ensures balanced segments — the observational data shows these are the strongest confounders. |
| Eligibility | All clients meeting current postponement criteria | Keep the same FS ≥ 500, single installment, ACTIVE state rules. |
Primary Metric
Default Rate at DPD30 (installment-level FDR):
Measured on all loans disbursed during the test period that have reached DPD30 maturity. Compares treatment vs control at the population level (intent-to-treat).
1 - SUM(REPAID_AMOUNT_DPD30) / SUM(DISBURSED_AMOUNT_DPD30)Measured on all loans disbursed during the test period that have reached DPD30 maturity. Compares treatment vs control at the population level (intent-to-treat).
Secondary Metrics
| Metric | Definition | Why It Matters |
|---|---|---|
| DR@DPD0 | Default rate at original due date | Measures on-time repayment. If treatment DR@DPD0 is higher, postponement may reduce urgency. |
| DR@DPD7, DPD15 | Default rate at 7 and 15 days past due | Repayment curve shape — are treatment borrowers catching up? |
| Time to repayment | Days from disbursement to full repayment | Even if DR is similar, postponement may delay cash collection. |
| Next-loan take rate | % who take another loan within 30d of repayment | Retention/engagement signal — does postponement improve loyalty? |
| Postponement adoption | % of treatment group who actually postpone | Feature usage rate. Expected ~1–2% based on current data. |
| Collection cost | Calls/SMS per loan in 0–30 DPD window | If postponement reduces delinquency, it should reduce outreach costs. |
Sample Size & Duration
| Parameter | Value | Notes |
|---|---|---|
| Baseline DR@DPD30 | 3.65% | From eligible population (Nov – Feb) |
| MDE (minimum detectable effect) | 1.0pp | Detect if treatment DR is ≥ 4.65% or ≤ 2.65% |
| Power | 80% | Standard |
| Significance | 5% (two-sided) | Standard |
| Required N per arm | ~7,000 loans reaching DPD30 | For proportions test on 3.65% base rate with 1pp MDE |
| Expected weekly eligible volume | ~5,000 loans/week | Based on ~85K eligible loans in ~17 weeks |
| DPD30 maturity lag | ~60–75 days from disbursement | Loan term (~30d) + 30 DPD observation |
| Enrollment period | 3 weeks | ~15K loans per arm |
| Total duration | ~13 weeks | 3 weeks enrollment + 10 weeks for last cohort to reach DPD30 |
Timeline
Week 0–1
Implementation: add client-level flag to
PostponementEligibilityService, stratified randomization, logging
Week 2–4
Enrollment: eligible clients randomly assigned. Monitor for balance (SRM check)
Week 5–8
Observation: early DPD0/DPD7 reads available. Check guardrails
Week 9–13
Maturation: DPD30 data fills in. Final readout at week 13
Guardrails (stop the test if breached)
| Guardrail | Threshold | Action |
|---|---|---|
| Treatment DR@DPD7 vs control | > 3pp worse at week 6 | Stop test, revert to treatment-for-all |
| SRM (sample ratio mismatch) | Chi-square p < 0.01 | Investigate randomization bug |
| Control group complaints | Support tickets mentioning missing feature | Monitor; tolerable if < 5 tickets/week |
Implementation Notes
- Randomization point: Add a check in
PostponementEligibilityService.isEligible(). Before returningtrue, check if the client is in the treatment group. If not, returnfalse. This keeps all other eligibility logic intact. - Assignment persistence: Hash
client_id + experiment_saltto deterministically assign groups. No need for a lookup table — same client always gets the same group. - Logging: Log the assignment (
postponement_ab_group: control/treatment) on every eligibility check, even for control. This lets us measure intent-to-treat. - Stratification: Pre-stratify by LN bucket (1–10, 11–20, 21+) and FS bucket (500–599, 600–699, 700+). Hash-based assignment naturally balances if volume is sufficient, but verify with a balance check at week 1.
6. Decision Framework
How to interpret results and what to do with them.
| Outcome | Treatment DR vs Control | Action |
|---|---|---|
| Clear win | Treatment DR significantly lower | Roll out to 100%. Postponement prevents defaults. Consider relaxing the 1-per-lifetime cap. |
| Neutral | No significant difference | Keep the feature — it's not hurting credit and likely improves retention/engagement. Measure next-loan take rate as the tiebreaker. |
| Mild negative | Treatment DR 1–2pp higher | Segment analysis: keep for LN21+ (low risk, high value), tighten for lower-tenure segments. |
| Clear negative | Treatment DR >2pp higher | Disable feature or restrict to FS ≥ 650 / LN ≥ 20. Redesign the UX to reduce moral hazard. |
7. Methodology
Data source: GHANA_PROD.ML.LOAN_INFO_TBL — installment-level FDR columns.
Population: Disbursed single-installment loans, Fido Score > 500, Nov 1 2025 – Feb 23 2026.
Postponed group: LOAN_POSTPONEMENT_DATE IS NOT NULL (N=905). DR measured from the new (extended) due date.
DPD0 missers: Non-postponed loans where REPAID_AMOUNT_DPD0 < 95% of DISBURSED_AMOUNT_DPD0.
Eligibility rules: Verified from PostponementEligibilityService.java in the fido-loan codebase.
Key limitation: All observational comparisons suffer from selection bias. The A/B test is the only way to establish causation.