ImmoDash
ENDE

How ImmoDash works

Eleven public datasets become one tested picture of the German rental market, rebuilt automatically every week.

Scroll

From raw files to this page

Each step is automated and tested. If any step fails, nothing is published and the site keeps the last good data.

  1. 01

    Collect

    Eleven open datasets: Zensus 2022, GREIX listings, Bundesbank loan rates, regional incomes, district and postcode boundaries.

    Python ingestion, raw files kept as-is (bronze)
  2. 02

    Clean and test

    Tables are joined to Germany's 400 districts and checked automatically: no gaps, no impossible values.

    dbt on DuckDB / PostgreSQL, 75 data tests
  3. 03

    Model

    A forecasting model predicts each city's rent 3, 6 and 12 months ahead, and unusual months are flagged.

    LightGBM quantile regression, conformal intervals, robust z-scores
  4. 04

    Serve

    Every number on this site comes from one typed interface, so the website and analyst tools never disagree.

    FastAPI, 25 endpoints, Pydantic contracts
  5. 05

    Explain

    An AI model writes each city summary from a fixed list of facts. If it uses a number that is not in the list, the text is thrown away.

    Groq LLM + numeric grounding check
  6. 06

    Publish

    Every Monday the whole chain reruns and this site is rebuilt from the fresh data.

    GitHub Actions, Next.js static export on Vercel

We tested the model on data it had never seen

We repeatedly cut the data at an earlier month, trained only on what was known then, and compared the forecast with what happened. Twelve starting points, all 37 cities.[15]

Over 12 months the model was off by 1.74% on average. Assuming rents stay the same was off by 4.59%, and extending last year's trend by 2.33%. That is about 62% smaller errors than the no-change guess.

82% of real outcomes fell inside the model's 80% range, close to the 80% it promises.

Average forecast error, lower is better

Mean absolute percentage error, out of sample.

ImmoDash model (LightGBM)Assume last year's trend continuesAssume no change
Source: ImmoDash back-test.
Show the numbers
HorizonModelError80% range hit rate
3 monthsImmoDash model1.06%81%
3 monthsTrend continues1.32%n/a
3 monthsNo change1.51%n/a
6 monthsImmoDash model1.22%77%
6 monthsTrend continues1.58%n/a
6 monthsNo change2.43%n/a
12 monthsImmoDash model1.74%82%
12 monthsTrend continues2.33%n/a
12 monthsNo change4.59%n/a

What each number means

Asking rent

The middle price per m² of flats advertised for rent in a month. Half of the listings cost more, half cost less.

Median of listing prices, GREIX hedonic index (adjusts for flat size and quality)

Based on [3][11]

New-lease premium

How much more someone signing a lease today pays than the average existing tenant paid in 2022.

Asking rent ÷ Zensus 2022 average contract rent − 1

Based on [4]

Rent burden

How much of a typical household's income the rent for a 60 m² flat would take.

Rent × 60 m² × 12 ÷ (2 × disposable income per resident)

Affordability index

The rent burden compared with the typical German district: 100 is average, higher is more affordable.

Median district burden ÷ district burden × 100

Estimated asking rent (districts)

For districts without listing data: the census rent raised by the gap between census and asking rents seen in nearby tracked cities.

Zensus rent × (1 + premium of the city, else state median, else national median)

Rent range within a district

Rents differ from street to street. The range shows the cheaper and dearer ends of a district, leaving out the cheapest and dearest 10% of its area.

District estimate × (10th and 90th percentile ÷ median) of Zensus grid cells: 100 m in the 37 cities, 1 km elsewhere

Postcode estimate

For the 37 cities: how a postcode compares with the rest of its city, applied to the city-wide estimate.

District estimate × median grid rent of the postcode's 100 m cells ÷ city median (postcode areas: OpenStreetMap)

Supply-demand index

High when few flats are empty and the population is growing.

z(population growth 2018 to 2023) − z(empty and available flats, 2022); ≥ 1 tight, ≤ −1 slack

Based on [9][8]

Live market pressure

How quickly flats are let in practice: days online and the share gone within a week.

Mean z-score of (inverted) days on market and share let within a week, per quarter

Unusual month

A month where one city moved far more than the others, compared with its own last two years.

City change minus all-city median change; robust z vs. previous 24 months; |z| ≥ 3.5

Based on [16]

References

Every explanation on this site that goes beyond our own numbers links to one of these publications. The numbers in square brackets, such as [1], point here. The datasets themselves are listed under Sources below.

  1. Lebuhn, H., Holm, A., Junker, S. & Neitzel, K. (2017). Wohnverhältnisse in Deutschland: eine Analyse der sozialen Lage in 77 Großstädten. Humboldt-Universität zu Berlin, funded by the Hans-Böckler-Stiftung (press release, 13 September 2017).

    Used for: Rent above 30% of household net income is treated as problematic, and many landlords use the same limit.

  2. Deutsche Bundesbank (2024). Preise für Wohnimmobilien sanken 2023 deutlich; Mieten stärker unter Druck. Summary of the Bundesbank Monthly Report on the housing market.

    Used for: Higher interest rates made buying unaffordable for many, and housing demand shifted towards the rental market.

  3. Amaral, F., Jungwirth, P., Zdrzalek, J. & Zetzmann, S. (2025). Dokumentation für den Mietpreisindex (GREIX). Kiel Institut für Weltwirtschaft, 12 June 2025.

    Used for: Asking rents in 37 cities from VALUE listing data, adjusted by hedonic regression for 21 characteristics including floor area, building age, rooms, condition, fittings and postcode.

  4. Bundesinstitut für Bau-, Stadt- und Raumforschung (BBSR) (2025). Wohnungsmieten. BBSR Fachbeitrag, Wohnungs- und Immobilienmarktbeobachtung.

    Used for: New-lease rents are well above existing rents because existing rents are adjusted to the market less often and rises for sitting tenants are capped by law.

  5. Sagner, P. & Voigtländer, M. (2026). In Großstädten: Mietangebote seit 2022 um bis zu 57 Prozent eingebrochen. Institut der deutschen Wirtschaft (IW), press release, July 2026.

    Used for: When new-lease rents are high, households stay longer, including in flats that no longer suit them.

  6. Gohl, N. (2019). House prices and spatial mobility: Lock-in effects on the German rental market. Beiträge zur Jahrestagung des Vereins für Socialpolitik 2019, ZBW.

    Used for: Academic study of the lock-in effect: the gap between sitting-tenant and new-lease rents reduces renters' mobility.

  7. Deutscher Mieterbund (2025). Betriebskostenspiegel für Deutschland, Abrechnungsjahr 2024. Deutscher Mieterbund, 18 December 2025.

    Used for: Running costs averaged 2.67 € per m² and month in 2024, on top of the cold rent.

  8. Sächsische Staatsregierung (2022). Sächsische Mietpreisbegrenzungsverordnung (SächsMPBVO), Begründung zu Indikator 5. REVOSax, 31 May 2022.

    Used for: Citing the BBSR, the normal turnover vacancy a market needs is mostly given as 2 to 3% in the literature, sometimes 3 to 5%.

  9. Bürgerliches Gesetzbuch (2025). § 556d BGB, Abs. 2: Gebiete mit angespannten Wohnungsmärkten. German Civil Code.

    Used for: German law names low vacancy with high demand, and population growth without enough new building, as signs of a tight housing market.

  10. Statistisches Bundesamt (Destatis) (2024). Zensus 2022: 4,3 % aller Wohnungen stehen leer. Press release No. 46, 4 July 2024.

    Used for: Census vacancy figures for May 2022, including flats that could be occupied within three months.

  11. Eurostat, ILO, IMF, OECD, UNECE & World Bank (2013). Handbook on Residential Property Prices Indices (RPPIs). Publications Office of the European Union.

    Used for: Standard reference for hedonic quality adjustment in housing price indices.

  12. Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q. & Liu, T.-Y. (2017). LightGBM: A Highly Efficient Gradient Boosting Decision Tree. Advances in Neural Information Processing Systems 30 (NeurIPS 2017).

    Used for: The gradient boosting library used for the forecasts.

  13. Koenker, R. & Bassett, G. (1978). Regression Quantiles. Econometrica 46(1), 33–50.

    Used for: Quantile regression: predicting a range (10th to 90th percentile) instead of a single value.

  14. Romano, Y., Patterson, E. & Candès, E. J. (2019). Conformalized Quantile Regression. Advances in Neural Information Processing Systems 32 (NeurIPS 2019).

    Used for: Calibrating quantile ranges so that they contain the stated share of real outcomes.

  15. Hyndman, R. J. & Athanasopoulos, G. (2021). Forecasting: Principles and Practice, 3rd edition, section 5.10 Time series cross-validation. OTexts, Melbourne.

    Used for: Rolling-origin back-testing: train only on data before each cut-off and score the forecast on what came after.

  16. Iglewicz, B. & Hoaglin, D. C. (1993). How to Detect and Handle Outliers (as summarised in the NIST/SEMATECH e-Handbook of Statistical Methods, 1.3.5.17). ASQC Quality Press.

    Used for: Modified z-scores based on the median absolute deviation, with |z| above 3.5 flagged as a likely outlier.

  17. United Nations, Department of Economic and Social Affairs, Population Division (2018). World Urbanization Prospects: The 2018 Revision. Key facts. United Nations, New York.

    Used for: Mumbai is among the world's largest cities, with close to 20 million inhabitants.

  18. Bhagat, R. B. & Jones, G. W. (2016). Demographic Dynamics of Mega-Urban Regions: The Case of Mumbai. Asia Research Institute Working Paper No. 201, National University of Singapore.

    Used for: Census 2011: Greater Mumbai had 12.47 million people, and 41.8% (5.2 million) lived in slums.

  19. Statistisches Bundesamt (Destatis) (2025). Anteil der Wohnkosten am verfügbaren Haushaltseinkommen (EU-SILC). Destatis table, Mikrozensus / EU-SILC.

    Used for: In 2025 housing took 24.1% of disposable income on average, but 43.7% for people at risk of poverty.

  20. Statistisches Bundesamt (Destatis) (2025). Wohnkostenanteil und Wohnkostenüberbelastung (EU-SILC). Destatis table, Mikrozensus / EU-SILC.

    Used for: In 2025, 11.2% of the population spent more than 40% of their income on housing; among people at risk of poverty it was 35.0%.

  21. Holm, A., Regnault, V., Sprengholz, M. & Stephan, M. (2021). Muster sozialer Ungleichheit der Wohnversorgung in deutschen Großstädten. Hans-Böckler-Stiftung, Working Paper Forschungsförderung No. 222.

    Used for: Almost 13% of tenant households in Germany's 77 big cities have less than the subsistence minimum left after paying rent.

About

ImmoDash is an independent portfolio project by Zoeb Ali Khan: a full data platform from raw government files to tested models, an API and this website, maintained automatically. It is not affiliated with any of the data providers.

  • Data engineering: medallion warehouse with dbt, DuckDB and PostgreSQL, Prefect orchestration
  • Geo analytics: 380,000+ census grid cells mapped to all 400 districts
  • Machine learning: LightGBM quantile forecasts, conformal calibration, anomaly detection
  • Applied AI: grounded LLM summaries with automatic fact checking
  • Product: FastAPI, Next.js, Plotly Dash, CI/CD with GitHub Actions
View the sourcePortfolio

Sources

DatasetLicence
Statistische Ämter des Bundes und der Länder, Zensus 2022 (table 4000W-0011)Datenlizenz Deutschland - Namensnennung - Version 2.0
Statistische Ämter des Bundes und der Länder, Zensus 2022 (table 4000W-0002)Datenlizenz Deutschland - Namensnennung - Version 2.0
Statistische Ämter des Bundes und der Länder, Zensus 2022 (table 4000W-0009)Datenlizenz Deutschland - Namensnennung - Version 2.0
Statistische Ämter des Bundes und der Länder, Zensus 2022 (table 4000W-0004, Bezirke)Datenlizenz Deutschland - Namensnennung - Version 2.0
OpenStreetMap postcode boundaries via yetzt/postleitzahlen (release 2026.02)ODbL 1.0 (© OpenStreetMap contributors); the derived postcode rent table is shared under ODbL
Statistische Ämter des Bundes und der Länder, Zensus 2022 Gitterzellen (100 m / 1 km)Datenlizenz Deutschland - Namensnennung - Version 2.0
Arbeitskreis VGR der Länder, Reihe 2 Band 3 (Berechnungsstand 2024)Datenlizenz Deutschland - Namensnennung - Version 2.0
GREIX Mietpreisindex, Kiel Institut für WeltwirtschaftFree use with attribution (GREIX, Kiel Institut / VALUE Marktdaten)
GREIX Miet- und Transaktionspreisindex, Kiel Institut für WeltwirtschaftFree use with attribution (GREIX, Kiel Institut)
Deutsche Bundesbank, MFI interest rate statistics (SUD131Z)Deutsche Bundesbank, free use with attribution
Bundesamt für Kartographie und Geodäsie via Praesklepios/geoGermany© GeoBasis-DE / BKG (VG5000, Stand 01.01.2021), dl-de/by-2-0