A small learning system picks the channel, send time, and tone for each group of overdue accounts. It trains, freezes its choices, then is tested on new simulated customers against a fixed default.
Loading picker results...
Synthetic customers with made-up response and payment odds. One train/test split; illustrates the method, not real-world lift or statistical significance.
| Days past due | Learned choice |
|---|---|
| 1-30 | SMS, morning (10:00 local), friendly |
| 31-60 | SMS, evening (18:00 local), friendly |
| 61-90 | SMS, morning (10:00 local), formal |
| 91-120 | SMS, morning (10:00 local), friendly |
| Days past due | Cases each | Paid: fixed | Paid: learned | Difference | Replied: fixed / learned |
|---|---|---|---|---|---|
| 1-30 | 200 | 33.0% | 33.0% | +0.0 pts | 53.5% / 53.5% (+0.0 pts) |
| 31-60 | 200 | 20.5% | 22.5% | +2.0 pts | 39.0% / 43.0% (+4.0 pts) |
| 61-90 | 200 | 9.5% | 12.5% | +3.0 pts | 27.0% / 27.0% (+0.0 pts) |
| 91-120 | 200 | 6.0% | 6.0% | +0.0 pts | 20.0% / 20.0% (+0.0 pts) |
Differences are in percentage points (learned minus fixed). With payment rates as low as 6.0% in some groups, a few hundred cases are not enough to be sure the learner found the best option; the result is reported as it came out.
| All groups |
|---|
| 800 |
| 17.3% |
| 18.5% |
| +1.3 pts |
| 34.9% / 35.9% (+1.0 pts) |