Real estate sales × Research note
Why flat bookings cancel: payment timing, home loans and what a call list can do
A cancelled booking costs a developer the sale, the marketing spent to win it and months of carrying the flat again. The warning signs usually appear in the first weeks, if someone is watching the right ones.
Key takeaways
- Under RERA, a promoter cannot take more than 10% of the price as an advance without first registering an agreement for sale, so the first payment demands after agreement are a key moment.
- In my synthetic data, the strongest early signals were lateness on the first demand, a high EMI-to-income ratio, investor buyers and a loan not yet sanctioned.
- Scored 45 days after booking, a model found 65% of cancellations by calling the riskiest 20% of buyers.
- With under 900 training bookings, simple models were as good as complex ones, so the explainable one was used.
- A risk score says who to call, not what the call achieves. Measuring retention needs an A/B test.
01The regulatory frame
Housing regulation shapes the booking journey. India’s Real Estate (Regulation and Development) Act, 2016 (RERA) is a good example: under Section 13, a promoter may not accept more than 10% of the cost of an apartment as an advance or application fee without first entering into a written agreement for sale and registering it[1]. The agreement must set out the payment schedule, the possession date and the interest payable by either side on default.
That makes the weeks after agreement critical. The buyer now faces the first construction-linked demands, the home loan must be sanctioned and disbursed, and the gap between enthusiasm at booking and affordability in practice becomes visible.
02Case study: Real Estate Sales
From my portfolio · project 13 of 15
Synthetic data- Problem
- Flat bookings cancel, unsold stock is priced by list, and buyers are segmented only by type.
- Decision supported
- Who the CRM team should call, what each unsold flat will fetch, and who the buyers are.
- Data
- 2,159 bookings across 8 launches in 7 Indian cities, with price lists and a city price index.
- Method
- Seven classifiers scored 45 days after booking; five price models; K-means segments.
- Result
- Calling the riskiest 20% finds 65% of cancellations; pricing error 2.0% vs 2.3% for the price list.
- Limits
- Synthetic data and small samples; retention impact needs an A/B test.
03What drives cancellations
My Real Estate Sales project models 2,159 bookings across eight launches in seven Indian cities, with realistic patterns: broker deals carrying bigger discounts and thinner booking amounts, subvention plans attracting investors, festival and quarter-end offers, and loan rejections driving cancellations. Across bookings at least a year old, 12.9% cancelled[2].
Each booking is scored 45 days after booking, once the first instalment is due, using only what is known by then. The logistic model’s odds ratios show what moves the risk:
Figure
Change in odds of cancelling, per standard deviation or when present
Timing is the feature. How late the first payment is says more than anything known on booking day.
04From a score to a call list
Seven classifiers were compared on 465 later bookings, 60 of which cancelled. The practical test is how many cancellations a CRM team finds by calling the riskiest fifth of buyers:
Figure
Cancellations found by calling the riskiest 20% of bookings
With under 900 training bookings, logistic regression, Naive Bayes and random forest were within noise of each other. The app uses the logistic model because every score can be explained to the person making the call: “the first demand is 20 days late and the loan isn’t sanctioned” is something a relationship manager can act on.
05Pricing unsold flats and knowing the buyers
The same project asks two more questions of the data[2]:
- What will each unsold flat actually fetch? An XGBoost model priced flats with a 2.0% error against 2.3% for the price list less the average discount, and removed the list’s +1.2% bias. A good price list is hard to beat; the gain comes from knowing which deals need a bigger discount.
- Who are the buyers? K-means on buying behaviour (income, age, ticket size, loan-to-value, EMI burden, booking amount, visits and time to decide) found five segments, with buyer type deliberately left out.
| Segment | Cancellation rate |
|---|---|
| Stretched buyers (high EMI) | 28% |
| Budget first-home buyers | 14% |
| Careful researchers | 11% |
| Affluent upgraders | 11% |
| Cash buyers | 6% |
06How strong is the evidence?
Not every finding in this note rests on the same kind of evidence. This is how I would weigh each one before acting on it.
| Finding | Evidence | Strength | Main caveat |
|---|---|---|---|
| Advance capped at 10% before a registered agreement | RERA 2016, s.13 | Strong | State rules add detail |
| Late first payment is the strongest early signal | Odds ratios on my synthetic bookings | Indicative | Synthetic data built with these patterns |
| Simple models match complex ones at this size | Seven classifiers on 465 later bookings | Moderate | Small sample; ranking is fragile |
07Using this in a sales team
- Watch the first demand closely. Lateness on the first payment is the strongest early signal.
- Track loan sanction status for every booking and follow up before the first demand, not after.
- Check affordability at booking: a high EMI-to-income ratio is a risk worth discussing openly.
- Call from a ranked list with the reasons shown, so each call addresses the actual problem.
- Test the calls. Hold back a random group to measure whether outreach actually reduces cancellations.
NotesSources
- Real Estate (Regulation and Development) Act, 2016, section 13. Text in IBC Laws: Section 13 of RERA.
- Iwale, A. (2026). Real Estate Sales, including the cancellation, pricing and segmentation upgrade. GitHub.
Figures are quoted from the sources above as published; where a source reports a range or a survey estimate, it is described that way. Results from my own projects say whether they use real public data or synthetic data.