Igor Ivitskiy, PhDDoctor Ads

Igor Ivitskiy, PhD in Mathematical Modelling

I use past payments to estimate what you can afford to bid.

Your report counts enquiries. Payments can take months.

You can order a model on its own.

  1. Bid today
  2. Forecast from past payments
  3. Check later receiptsCash kept after refunds
Six-month repeat purchases, group 9 of 10
Forecast
92.93%
Bought again
84.07%

I match records from the bid date to later payments.

Agree how far ahead to look. Margin forecasts need costs.

See the forecast check
At the bidLater outcome
Enquiry
Still in sales cycle: unknownDeal, invoice and payment dates. Include lost deals.
Customer
Repeat orders, payments and refunds
Sign-up
Renewal or cancellation

My six-month forecast predicted repeat purchases more accurately than the recent-purchase rule.

I scored 4,966 previous customers in UCI Online Retail II using earlier records alone. I checked purchases from 9 June 2011 up to, but excluding, 9 December 2011.

Group 9 of 10 by forecast probability
0%100%
Forecast 92.93%
Bought again 84.07%

Largest gap: 8.86 percentage pointsTolerance: 5 percentage points

Four of ten groups fell within tolerance. One period, without confidence intervals.

All ten groups and test method

I chose the model before the check. I used isotonic calibration to align its scores with three-month purchase rates. To forecast six months, I assumed the purchase rate stayed steady.

I ranked customers by their forecast chance of buying again and split them into ten groups. Each row compares the group's average forecast with the share who bought again.

UCI Online Retail II · CC BY 4.0
Group by forecast Forecast Bought again
1 (lowest) 9.38% 13.68%
2 26.21% 27.36%
3 29.53% 24.35%
4 34.27% 40.04%
5 50.76% 49.30%
6 59.52% 53.92%
7 65.71% 59.68%
8 80.51% 72.78%
9 92.93% 84.07%
10 (highest) 99.07% 94.56%

Past revenue picked the buyers who brought in more money.

Read the evidence
Probability error (Brier score, lower wins) Model 0.1902 (winner) Rule: bought in previous three months, 0.3101
Share of later net revenue from top 10% (higher wins) Purchase probability 55.39% Past revenue 61.45% (winner)

Test one bid change against a comparable campaign.

The shop test covered returning customers. First-time leads and subscribers need their own test.

  1. Beat a simple rule on your records
  2. Changed bidUnchanged bid
  3. Compare cash kept
What to agree for this model

I agree code handover, integration and support with you for each task.

Tell me what you must decide before the money arrives.

Required unless marked optional

What can I bid for an open enquiry?
Describe your CRM, order or billing records. Include the dates you keep.
Privacy