Land Value Tax Calculator

Compare the taxes you currently pay to a Land Value Tax system

More detail on methodology

Step	Step Name	Simple Description	Technical Description
1	Data Loading and Initial Cleaning	Download and tidy up the four raw datasets — Land Registry sales, property addresses, energy certificates, and plot boundaries — ready for processing.	Load LR PPD, OS Open UPRN, EPC (348 LA files), and INSPIRE GML batches; standardise columns, deduplicate EPC on UPRN, batch-reproject INSPIRE to WGS84, write clean parquets and GeoPackages.
2	Address Normalisation	Reformat addresses in both the sales data and the energy certificates so they can be reliably compared and matched against each other.	Uppercase, strip punctuation, expand abbreviations (ST→STREET etc.), and concatenate component fields into a single normalised string per dataset; postcode excluded from string and used as match key.
3	Land Registry to EPC Fuzzy Match	Find the energy certificate that belongs to each property sale, so that floor area and building details can be linked to the sale price.	Two-pass match within postcode: exact string match first, then token-ratio fuzzy match on residuals; confidence threshold applied; unmatched records flagged for postcode centroid fallback.
4	Assign Coordinates	Give each property sale a precise location — either the exact address point if matched, or the centre of its postcode otherwise.	Key-join EPC UPRN to OS Open UPRN for precise lat/lon; postcode centroid from ONS Postcode Directory as fallback; coordinate type flag added to all records.
5	Attach EPC Attributes	Add the physical building details — floor area, property type, age, and energy rating — to each matched sale record.	Carry forward floor_area_m2, property_type, built_form, construction_age_band, and energy_rating from matched EPC certificates; unmatched records flagged with imputed floor area by property type.
6	Spatial Join to INSPIRE Polygons	Work out which registered land parcel each property sits on, and calculate the size of that plot.	GeoPandas spatial join across 7 INSPIRE batch GeoPackages; assigns inspire_id and plot_area_m2 from polygon geometry; records categorised as polygon-matched, UPRN-only, or postcode-centroid.
7	Plot Area Filter	Remove implausibly tiny or very large plots from the training data, as they give unreliable signals about land value per square metre.	Lower bound ~10 m² removes artefact slivers; upper bound ~10,000 m² excludes very large sites from model training only; large polygons retained for display; thresholds set from Step 6 size distribution.
8	Building Value Model	Estimate what the building alone is worth, using prices paid in cheap areas where land adds little value. The remainder of the sale price is treated as the land’s value.	Gradient boosting model trained on North East + Wales transactions below national P30 price/m²; features: log floor area, property type, built form, age band, energy rating; no location inputs. Applied nationally: land_value_est = price − building_value_est. Flats and leaseholds excluded throughout.
9	Grid Construction	Divide England and Wales into a regular grid of 100 m × 100 m squares, clipped to the coastline.	100 m raster constructed in British National Grid, clipped to OS land boundary; each cell assigned centroid lat/lon and a data-presence category (INSPIRE-matched, postcode fallback, or empty).
10	Grid Surface Assembly	Spread each property’s estimated land value across the grid using a distance-decay formula, blending towards a low agricultural baseline in areas with sparse sales data.	FFT convolution on 100 m raster; exponential decay h = 0.3 km, radius 3 km; Bayesian prior (W_PRIOR = 2.0, floor £2/m²); sqrt plot-area normalisation (alpha ≈ 0.5); negative LV clamped to zero but retained in denominator.
11	Grid Surface Validation	Check how accurately the finished grid predicts actual sale prices by comparing the grid’s values back against the transactions used to build it.	Reconstruct predicted_price = grid_lv × sqrt(REF_AREA × effective_plot_area) + building_value_est per transaction; report R², RMSE, MAE, MdAPE, and bias by quality band, region, and price decile.
12	PMTiles Generation	Package the grid and all land parcel boundaries into a single map tile file for the web app, covering the whole country at every zoom level.	Three-pass tippecanoe: Pass A — background grid z6–13 (no-drop); Pass B — grid + parcels z6–16 (drop-densest); Pass C — tile-join merge to land_values.pmtiles (7.56 GB). Parcels assigned LV/m² by sampling grid at polygon centroid.