Asking rent research on Isambard AI
Can AI spot asking rent cuts before they happen?
Bottom line
The historical cut rate gave the lower seven day forecast error overall. The apparent fourteen day advantage from nearby adverts in Wales did not repeat on two additional dates.
- Published
- Saved captures
- 18 September 2026 to 15 November 2026
- Adverts in the research archive
- 179,218
- Comparable starting prices at 7 days
- 88,073
- Forecast windows
- 7, 14 and 30 days
- Publication code
- UI-RENT-FORECAST-2026-01
Wales follow up
30 October 2026
Of 290 Welsh adverts we found again around fourteen days later, 22 had a lower asking rent. The model using adverts within 15 km found 5 cuts among its 29 highest ranked adverts.
31 October 2026 to 1 November 2026
We found 369 Welsh adverts again exactly fourteen days later. 51 had a lower asking rent. Among the 37 highest ranked adverts, the 15 km model found 2 cuts and the 5 km model found 3. A random selection of the same size would have found about five on average.
The historical cut rate gave the most accurate probabilities on the extra dates. Its forecast error was 0.1192, compared with 0.1197 for the 5 km model and 0.1208 for the 15 km model. Lower scores are better. The extra dates came from the same test week.
New areas and later weeks
Where the forecasts held up
Pick a time window and a region to see whether AI beat the historical rate on the same adverts.
Lower forecast error on the adverts we tested
7 days · 30 October 2026 to 7 November 2026
East Midlands
Advantage uncertain across weeksThese are scores for the adverts we tested. More test weeks are needed to see whether the difference holds.
- Eligible starting prices
- 9,922
- Comparable follow ups
- 7,732
- Advertised rents fell
- 735
- Distinct adverts tested
- 5,261
9.5% of comparable follow ups showed a lower asking rent.
Among the cuts we saw, the median drop was £50.00 a month, with a median percentage cut of 4.5%.
Forecast error Lower is better
The historical rate scored 0.0012 lower.
There were too few test weeks to estimate a reliable uncertainty range.
Read the figures for every region
| Region | Window | Comparison | Paired prices | Cuts | AI error | Historical error |
|---|---|---|---|---|---|---|
| North East | 7 days | Historical rate had lower error | 2,585 | 168 | 0.0695 | 0.0610 |
| North West | 7 days | Historical rate had lower error | 7,417 | 668 | 0.0840 | 0.0820 |
| Yorkshire and The Humber | 7 days | Historical rate had lower error | 5,242 | 442 | 0.0826 | 0.0772 |
| East Midlands | 7 days | Historical rate had lower error | 7,732 | 735 | 0.0874 | 0.0862 |
| West Midlands | 7 days | Historical rate had lower error | 6,266 | 558 | 0.0840 | 0.0812 |
| East of England | 7 days | Historical rate had lower error | 7,392 | 628 | 0.0783 | 0.0778 |
| London | 7 days | Historical rate had lower error | 26,298 | 2,826 | 0.1020 | 0.0969 |
| South East | 7 days | Historical rate had lower error | 14,511 | 1,313 | 0.0870 | 0.0825 |
| South West | 7 days | Historical rate had lower error | 7,016 | 627 | 0.0839 | 0.0815 |
| Wales | 7 days | Historical rate had lower error | 1,973 | 98 | 0.0495 | 0.0480 |
| Scotland | 7 days | Historical rate had lower error | 1,641 | 155 | 0.0868 | 0.0858 |
| Northern Ireland | 7 days | No comparison | — | — | — | — |
| North East | 14 days | Too little evidence | 70 | 6 | — | — |
| North West | 14 days | Historical rate had lower error | 545 | 112 | 0.1812 | 0.1679 |
| Yorkshire and The Humber | 14 days | Historical rate had lower error | 169 | 22 | 0.1210 | 0.1133 |
| East Midlands | 14 days | Historical rate had lower error | 412 | 67 | 0.1503 | 0.1369 |
| West Midlands | 14 days | Historical rate had lower error | 365 | 62 | 0.1551 | 0.1421 |
| East of England | 14 days | Historical rate had lower error | 461 | 100 | 0.1887 | 0.1763 |
| London | 14 days | Historical rate had lower error | 1,334 | 285 | 0.1983 | 0.1757 |
| South East | 14 days | Historical rate had lower error | 958 | 193 | 0.1806 | 0.1655 |
| South West | 14 days | Historical rate had lower error | 417 | 65 | 0.1386 | 0.1320 |
| Wales | 14 days | AI had lower error | 290 | 22 | 0.0712 | 0.0732 |
| Scotland | 14 days | Too little evidence | 42 | 10 | — | — |
| Northern Ireland | 14 days | No comparison | — | — | — | — |
| North East | 30 days | Too little evidence | 224 | 55 | — | — |
| North West | 30 days | Too little evidence | 715 | 240 | — | — |
| Yorkshire and The Humber | 30 days | Too little evidence | 459 | 168 | — | — |
| East Midlands | 30 days | Too little evidence | 581 | 196 | — | — |
| West Midlands | 30 days | Too little evidence | 497 | 173 | — | — |
| East of England | 30 days | Too little evidence | 669 | 229 | — | — |
| London | 30 days | Too little evidence | 1,341 | 490 | — | — |
| South East | 30 days | Too little evidence | 1,136 | 408 | — | — |
| South West | 30 days | Too little evidence | 509 | 183 | — | — |
| Wales | 30 days | Too little evidence | 175 | 44 | — | — |
| Scotland | 30 days | Too little evidence | 165 | 70 | — | — |
| Northern Ireland | 30 days | No comparison | — | — | — | — |
What the tests found
We used 1,032,653 dated asking prices across 179,218 adverts in captures saved from 18 September 2026 to 15 November 2026. The figures below follow starting prices in places and weeks set aside for testing.
Seven days
Of 139,102 starting prices old enough to check, 88,073 (63.3%) had a comparable price saved near the target date. The asking rent was lower in 8,218 of those comparisons (9.3%). Among the cuts, the median drop was £50 a month. The median percentage cut was 4.3%.
The historical rate had the lower forecast error, scoring 0.0848 against 0.0885 for the model using nearby adverts. The model using only the advert’s own history scored 0.0894. Both forecasts used the same 88,073 price comparisons, with starting prices from 30 October 2026 to 7 November 2026. There were too few separate test weeks to tell whether this difference holds across weeks.
The historical rate scored lower in 11 regions, including North East, Yorkshire and The Humber and London.
Fourteen days
Of 28,783 starting prices old enough to check, 5,246 (18.2%) had a comparable price saved near the target date. The asking rent was lower in 977 of those comparisons (18.6%). Among the cuts, the median drop was £55 a month. The median percentage cut was 4.5%.
The historical rate had the lower forecast error, scoring 0.1554 against 0.1698 for the model using nearby adverts. The model using only the advert’s own history scored 0.1664. Both forecasts used the same 4,951 price comparisons, with starting prices from 30 October 2026. There were too few separate test weeks to tell whether this difference holds across weeks.
The nearby advert model scored lower in Wales. The historical rate scored lower in 8 regions, including London, South East and East Midlands.
Thirty days
Of 15,646 starting prices old enough to check, 6,471 (41.4%) had a comparable price saved near the target date. The asking rent was lower in 2,256 of those comparisons (34.9%). Among the cuts, the median drop was £65 a month. The median percentage cut was 5.3%.
A 30 day forecast needs an earlier training group whose prices were already known a month later, and a separate test group with enough cuts and other outcomes. The dates we captured left too few qualifying examples to make that comparison.
How far should the model look?
We repeated the test with a model that could look up to 15 km away.
At 7 days the 15 km forecast had higher error on the same 88,073 price comparisons. It scored 0.0895, compared with 0.0885 for the 5 km forecast.
At 14 days the 15 km forecast had lower error on the same 4,951 price comparisons. It scored 0.1672, compared with 0.1698 for the 5 km forecast.
The share of asking prices that fell uses only the prices we found again in each window. An advert can contribute a starting price in more than one week. Where the follow up is missing, the asking rent at that point is unknown. The figures describe advertised prices. What tenants agreed to pay is a separate question.
How we tested the forecasts
At each starting price, the model estimated the chance that the same advert would be cheaper when it next appeared around the target date. The test withheld entire regions and later weeks from training. Each comparison used the same starting prices and follow ups.
The historical approach applied the rate of asking rent cuts seen in the training sample to every advert in the test. Forecast error uses the Brier score, which compares each predicted chance with what happened. A lower score means more accurate probabilities. The difference is positive when AI had the lower error.
The main model could use up to 16 comparable adverts within 5 km, drawn from the previous seven days. They were in the same region and asked between 80% and 125% of the starting rent. When both adverts gave a bedroom count, it had to match. We also tested a model using only each advert’s own price history.
We checked for comparable captures near days 7, 14 and 30. An advert could contribute more than one starting date. The uncertainty range accounts for repeated observations of the same adverts and for variation between test weeks where enough weeks were available.