collections-data-analyst-assignment

Executive Memo

To: Leadership Team From: Collections Analytics Date: September 2026 (revised; see CORRECTIONS.md) Re: Is recovery improving, and where should ₹10 Cr go?

---

1. What happened?

The reported “11% month-on-month improvement” is one month, not a trend.

• March recovery was 11.0% above February. That is real. But February was 9.1% below January, and the months since alternate up and down.
• Across January to July, recovery is flat: ₹18.72 Cr in January and ₹18.72 Cr in July. Month-on-month changes from February to July average +0.3%.
• The data needed one correction: 325 duplicated successful payments (₹2.45 Cr, about 1.9%) were removed, keyed on payment_id. They are spread evenly across months and do not explain the March number.

Month	Recovery (₹ Cr)	MoM	Accounts paying	Paying / portfolio
Jan	18.72	—	2,374	7.9%
Feb	17.01	−9.1%	2,173	7.2%
Mar	18.89	+11.0%	2,419	8.1%
Apr	17.51	−7.3%	2,304	7.7%
May	18.43	+5.2%	2,344	7.8%
Jun	17.56	−4.7%	2,286	7.6%
Jul	18.72	+6.7%	2,335	7.8%

2. Why?

Nothing in operations changed:

Driver	Finding	Evidence level
Contact	Answer rate 19–20%, right-party-contact 76–79% of dispositioned calls, flat	Fact
Promises	Promise-kept rate 24–26%, flat	Fact
Productivity	₹16,100–17,000 recovered per agent-hour, flat	Fact
Portfolio mix	Share of paying accounts in early DPD 44–47%, flat	Fact
April strategy shift	Field vs digital difference-in-differences +1.3 pp, 95% CI −0.2 to +2.9	No detectable effect
Month-to-month swings	Alternating ±5–11% with no driver above	Most consistent with normal variation

Reporting problems that would mislead a dashboard (fixed in the pipeline):

• PTP and PROMISE_TO_PAY both mean a promise, in all three schema versions; counting one misses about half of promises.
• Calls are logged in three time zones; hourly analysis needs UTC conversion.
• Agent master data changes employee code and name from row to row, so per-agent league tables cannot be trusted until it is fixed at source.

3. How confident are we?

Claim	Confidence
Recovery is flat Jan–Jul	High: same result with or without the duplicate correction
The +11% is a single rebound month	High
No channel shows measurable lift	Moderate: observational data; every 95% interval includes zero
The April shift had no detectable effect	Moderate: parallel-trends assumption, small Field sample

Limits: synthetic dataset, 7 full months (August has 8 days), no randomised control group, so causal claims about channels are not possible from this data.

4. What should we do with ₹10 Cr?

Do not commit ₹10 Cr to one channel on this evidence. WhatsApp is credited with the most recovery (₹16.95 Cr under last-touch attribution) only because it sends the most messages. Accounts reached by any channel pay within 30 days at 7.8–8.0%, the same as the 7.7% baseline for all accounts.

Recommended split:

1. Up to ₹1.5 Cr (proposed): an 8-week randomised holdout test. Randomly hold out 20% of the accounts each channel would contact and compare 30-day payment rates. With about 20,000 accounts per channel and a 7.8% base rate, this detects a lift of about 1.3 percentage points (80% power, 5% level). This is what turns “WhatsApp gets the most credit” into “WhatsApp causes recovery”.
2. Up to ₹1.5 Cr (proposed): data fixes. Payment IDs enforced at ingestion, one disposition taxonomy, UTC timestamps, a clean agent master. Without these, the next report will repeat this review’s problems.
3. The remaining ₹7 Cr or more: held, then deployed to whichever channel the test shows works, in proportion to measured lift per rupee.

Why not hire more agents? Productivity per agent-hour is flat and contact shows no measurable lift, so more calls are unlikely to add recovery until the test says otherwise.