Budapest rental price prediction

[§U]Use case

Suggest a rent in seconds, list faster

A landlord opens the upload flow and enters the characteristics of the flat they want to let — area, rooms, floor, condition, balcony, and so on. Before they finish, the page shows a suggested monthly rent. Because the number is credible and arrives instantly, they price competitively from day one, list sooner, and start earning rent earlier.

The chain is simple: faster pricing → quicker listing → happier landlord → higher engagement → more supply on the platform. For ingatlan.com, which leads the Hungarian for-sale market but faces stiffer competition in rentals, this is a direct lever to grow the rental user base.

[§1]The decision this supports

Can we suggest a credible rent at upload?

The capstone stays offline — modelling and recommendations, not deployment or A/B testing. The question is whether a model can hit the accuracy target given the data, scope, and product constraints.

• Scope: Budapest rental flats.
• Target: monthly rent per square metre (HUF/sqm).
• Metric: MdAPE — typical absolute % miss.
• Success criterion: ≤10% MdAPE, set by the client. Met at ~9.8% in time-ordered CV.

[§2]Data & temporal validation

Why the newest 20% is reserved

Budapest rental prices move quickly, so a random split would leak market-regime information from the future. Candidate selection uses rolling time-series CV on older listings; a frozen holdout on the newest 20% is kept for the eventual production go/no-go.

[§3]Pipeline · medallion architecture

Bronze → silver → gold

The medallion pattern layers a warehouse into three progressively curated tiers. Bronze is raw ingested data as it arrived, with lineage but no cleaning. Silver is typed, deduplicated, and quality-checked — the same shape as source but trustworthy. Gold is modelling-ready feature marts: business rules applied, scope filters baked in, joins with enrichment tables done.

This capstone follows the same shape: raw listing dumps land in a bronze layer, the data-prep pipeline produces a cleaned silver extract, and the modelling table plus geospatial features form the gold layer that CV and the holdout consume.

  bronze                silver                 gold
  ──────                ──────                 ────
  raw dumps    ─────►   typed / cleaned  ────► modelling extract
  ingatlan.com          scope filters          + WorldPop lag features
  scrape                dedup + QA             + NDVI H3 features
                                               + 90-day comparable prior
                                                     │
                                                     ▼
                                          time-series CV  +  holdout

[§4]Model leaderboard

Boosting clears the 10% bar; linear stays close

Same target, same rolling time-series CV, same metric. Boosting leads at ~9.8–9.9% MdAPE, but well-specified linear models are only ~0.2 pp behind — and both are ~7 pp better than a district-median baseline. Stability and product fit matter more than the marginal CV gain.

[§5]What the model attends to

Size, comparables, condition, location

Filling core structured fields matters more than adding NLP at this stage. The neighbourhood comparable prior and location features carry strong signal, confirming the client team's interim feedback. Importance is not causality — use it for field prioritization and sanity checks.

[size & layout] [local comparables] [condition] [balcony ratio] [air conditioning] [centrality]

Interaction terms kept in the model

A small set of numeric interactions, chosen to represent plausible housing-market effects without exploding the one-hot feature space.

[§6]Geospatial enrichment

Population and greenness, lagged and historized

Geospatial features stay lagged and source-dated to avoid leakage. Static land cover is useful for EDA but is not aligned with historized validation.

A quick primer on H3

H3 is Uber's open-source discrete global grid: the earth is tiled with hexagons, each with a stable string ID (e.g. 8830953b13fffff). Hexagons are convenient because every neighbour is the same distance away, so aggregations and lookups behave uniformly. Resolution controls cell size — res 6 covers a metro area, res 7 fits a district, and res 8 (used below) is about the size of a few city blocks, matching how buyers think about local rent.

Fig. summer Sentinel-2 NDVI, aggregated to H3 res-8 cells over Budapest. Darker green = more vegetation, useful as a proxy for green-space amenity around a listing.

[§7]Implications for the upload flow

[§8]Risks & recommended next steps

Bottom line: a credible, reproducible offline artifact and a practical roadmap. The remaining decision is product governance, not more exploratory modelling.

[§A]Appendix

Diagnostics, feature policy, data quality

Feature policy

Data-quality summary

Source repo: github.com/balintdecsi/ceu-public-thesis · Original slide deck: html deck ↗