constructioncapital/england-housing-supply-2000-2025

England housing supply: a 25-year panel and retrospective baselines

This dataset provides 250 annual observations of net additional dwellings for England and its nine regions, from 2000/01 to 2024/25. It includes the original government workbook, source-cell references, revision flags, a data dictionary and a Python script that rebuilds the tables without network access or additional packages.

The accompanying exercise compares two simple one-year-ahead baselines: the previous year's value and the mean of the previous three years. It supplies every prediction and error, making it suitable for teaching time-series evaluation, testing data-loading pipelines and checking more elaborate models against simple alternatives.

This is a small, fixed-vintage research dataset. It is not a real-time forecasting backtest. All observations come from the MHCLG edition published on 20 November 2025, which includes retrospective revisions. Values for an earlier year may differ from those available to a forecaster at that time. No forecast for an unobserved future year is provided.

Version 1.0.0, prepared 29 September 2026. Research reference B001-R30-V08. Curated and analysed by Matt Lenzie, Construction Capital.

Data and scope

The source is MHCLG Live Table 118, worksheet LT118_unrounded, in the 20 November 2025 workbook.

Each observation is one geography and one financial year, covering April to March. The measure is annual net additions to the dwelling stock, in dwellings. It includes new build, conversions, changes of use, other gains and losses, demolitions and census adjustments as specified by the source. It does not measure the total housing stock, planning approvals, transactions or loans. Census adjustments are already included in the selected net-additions row and must not be added again.

England is an aggregate of the regions. Do not add England and the regional rows together or treat all ten series as independent samples. No personal, address-level or borrower data are included. Scotland, Wales and Northern Ireland are outside this dataset.

File	Contents
observations.csv	250 official observations with geography, financial year, source cell and flags
baseline_predictions.csv	200 retrospective predictions: 10 geographies, 10 target years, two baselines
baseline_metrics.csv	MAE, RMSE, mean signed error and MAPE for each geography and baseline
regional_reconciliation.csv	Regional sums and differences from the published England totals
source_notes.csv	The official workbook's explanatory notes and flag definitions
data_dictionary.json	Field types, units and definitions for the five CSVs
Live_Table_118.ods	Unmodified original source workbook
reproduce.py	Standard-library extraction and calculation script
validation.json	Record counts, source hash and calculation diagnostics
source_manifest.json	Source provenance, edition and transformations
SHA256SUMS.txt	File-integrity checksums

The Hub's split named data is a complete table, not a model training split. The three configurations separate observations, predictions and metrics. No random train/test split is supplied.

Reproduce and use

After downloading the repository files, run:

python reproduce.py

The script checks the original workbook's SHA-256 before extraction. It writes the five CSVs and validation.json. Use --output another-directory to keep the downloaded outputs unchanged.

Read the data with the Python standard library:

import csv
with open("observations.csv", encoding="utf-8", newline="") as f:
    rows = list(csv.DictReader(f))
england = sorted(
    (r for r in rows if r["geography"] == "England"),
    key=lambda r: int(r["year_start"]),
)
values = [int(r["net_additions_dwellings"]) for r in england]

Baseline method and results

For each geography, predict the target years 2015/16 through 2024/25. The previous-value baseline uses only the preceding year's value. The three-year-mean baseline uses the three immediately preceding values. Neither includes the target observation in its input window. Both are evaluated on the same ten targets per geography.

An error is prediction minus actual. MAE is the mean absolute error. RMSE is the square root of mean squared error. Mean signed error indicates average overprediction or underprediction. MAPE is the mean of absolute error divided by the positive observed value, multiplied by 100. Prediction values are not rounded to whole dwellings before evaluation.

England, targets 2015/16 to 2024/25	MAE (dwellings)	RMSE (dwellings)	MAPE
Previous year's value	14,541.6	17,594.5	6.58%
Previous three-year mean	23,161.1	28,292.7	10.51%

The previous-value baseline has lower MAE in England and each of the nine regions in this particular window. This is a descriptive result from ten evaluated years per series. It does not establish statistical significance, superiority in other periods or reliability for investment decisions. National and regional errors are reported separately; they are not pooled into a misleading combined national score.

Quality and limitations

• The 2024/25 figures are provisional. The source flags 2023/24 as revised. is_revised=false means the row lacks the explicit [r] marker, not that its history has never been revised.
• The source records estimated component breakdowns for some joint regional returns in 2006/07 and 2007/08. The original notes are included.
• Published regional sums can differ from the England total. The maximum absolute difference in the extracted period is 10 dwellings. These differences are retained rather than forcing a reconciliation.
• There are only 25 observations per series and ten evaluation targets. The series are related, and the evaluation spans substantial economic and policy changes. Randomly splitting rows would obscure the temporal and geographic structure.
• All baseline input values use the 2025 edition. This prevents future target rows from entering the input window, but cannot eliminate look-ahead from historical revisions. A genuine real-time test requires archived release vintages.
• No missing observations were imputed, no synthetic housing counts were generated, and no causal effects were estimated.

Related analysis and attribution

Construction Capital's explanation of regional housing output and changing national shares discusses why a region's share can rise while its output falls. The companion housing peak-gap analysis uses a narrower 2015/16 to 2024/25 window. These publications share the same official source family; they are not independent validation of this dataset.

Suggested attribution: MHCLG, Live Table 118, edition published 20 November 2025; dataset preparation and retrospective baseline analysis by Matt Lenzie / Construction Capital, England housing supply: a 25-year panel and retrospective baselines, version 1.0.0, 29 September 2026. Identify any changes you make.

Contains public sector information licensed under the Open Government Licence v3.0. See LICENSE.md for the separate data, documentation and code licences. Neither MHCLG nor Hugging Face endorses this analysis.