SPB Git forge
spboucher.ai ↗

spb/wp21_uqo

Public
0commits 0branches 0releases
0 Bsize
maindefault branch
—last push
9.6 KB
View source Raw Blame History

# CLAUDE.md — UQO Working Paper No. 21

PROJECT IDENTITY. UQO Working Paper Series — WP21. Provisional title: The Global Evidence on Housing Price Determinants: A Comprehensive Meta-Analysis of Hedonic Pricing Models, 1970–2026. Fields: real estate economics, housing economics, applied econometrics, meta-analysis, machine learning. Language: English. Primary deliverable: a fully reproducible, publication-quality empirical paper in LaTeX with a structured research database, extraction pipeline, statistical scripts, tables, figures, supplementary materials and a complete bibliography. Ambition: the most comprehensive verifiable quantitative synthesis of housing hedonic pricing research feasible with the accessible literature. Autonomous scientific research assignment, not a narrative review. Never invent studies, citations, coefficients, standard errors, observations, statistical results, or claims of comprehensiveness.

  1. SCIENTIFIC FOUNDATION. Start from Sirmans, Macpherson & Zietz (2005, JREL), Sirmans, MacDonald, Macpherson & Zietz (2006, JREL) and Rosen (1974, JPE); identify subsequent systematic reviews, meta-analyses, replications and methodological developments through October 2026; read in full when legally accessible; reconstruct the Sirmans programme, its limitations, and what a modern global synthesis adds. Distinguish systematic review / descriptive synthesis / statistically valid meta-analysis of comparable effect sizes / meta-regression of heterogeneity. Do not confuse the number of coefficients with the number of independent studies.

  2. RESEARCH QUESTIONS. RQ1 characteristics that consistently affect prices across countries and periods; RQ2 size and uncertainty of marginal contributions; RQ3 variation with geography, market conditions, property types, socioeconomics; RQ4 change over time (crises, monetary policy, COVID-19); RQ5 traditional vs spatial vs machine-learning estimates; RQ6 incremental information from text, computer vision and AI-derived characteristics; RQ7 genuine heterogeneity vs specification, definitions, measurement, publication selection; RQ8 understudied characteristics and gaps.

  3. LITERATURE DISCOVERY. Reproducible search 1970–2026-10-20 across OpenAlex, Crossref, Semantic Scholar, RePEc/IDEAS, EconPapers, SSRN, arXiv, Google Scholar (compliant), CORE, DOAJ, repositories, publisher sites, working-paper series, backward/forward citation networks; multilingual queries (hedonic housing prices, hedonic property valuation, residential property price determinants, housing characteristics and prices, implicit prices of housing attributes, real estate hedonic regression, spatial hedonic models, housing price meta-analysis, automated valuation models, text-based real estate valuation, computer vision property valuation, machine learning and hedonic pricing); snowballing; deduplication (DOI, title similarity, authors, year, content); full records of queries, dates, sources, screening decisions, retrieval failures. Aspirational targets: ≥5,000 unique records screened, ≥1,000 full texts, several hundred eligible studies, every usable traceable estimate; never relax criteria to reach targets. Compare with competing reviews before any "largest" claim.

  4. SYSTEMATIC REVIEW PROTOCOL. PRISMA 2020 adapted to economics; protocol before synthesis; inclusion/exclusion criteria; include empirical studies estimating the relation between residential property prices and identifiable property, neighbourhood, environmental or location characteristics; separate single-family / condominiums & apartments / multifamily / rental / mixed; distinguish transaction prices, listing prices, appraisals, rents; do not combine without justification; exclude studies without sufficient quantitative evidence from coefficient synthesis but keep them qualitatively; record exclusion reasons; PRISMA flow diagram from actual records; flag ambiguous studies for human review.

  5. AUTOMATED FULL-TEXT EXTRACTION. Python pipeline (PyMuPDF, pdfplumber, structured parsing, table extraction, vision/LLM-assisted classification). Extract study metadata (DOI, reference, authors, year, country/region/city, sample period, outlet, document type, observations, property type, transaction vs listing, frequency, scope), econometric information (dependent variable and transformation, estimation method, spatial controls, time and location fixed effects, covariates, functional form, SE estimator, sample restrictions, fit), coefficient-level information (exact variable name, coefficient, SE, t, p, CI, units, transformation, model and table identifiers, page and table, extraction confidence). Store raw and harmonised values; never infer missing SEs except transparently from reported statistics; preserve every conversion formula; every number traceable to source table.

  6. GLOBAL HEDONIC VARIABLE TAXONOMY. Hierarchical: A structural; B interior amenities; C exterior amenities; D location and accessibility; E environmental; F neighbourhood and socioeconomic; G market and transaction; H AI-derived. Standard conceptual definitions, cross-study terminology mapping, conceptually distinct variables kept separate.

  7. EFFECT-SIZE HARMONISATION. No naive averaging. Log-linear binary: 100[exp(β)−1] with retransformation and uncertainty; log-log: elasticities; level-level: monetary vs percentage; unit conversions (sq ft / m²); no monetary-to-percentage conversion without a defensible reference price; partial correlations from t and df when justified; standardised coefficients only with variance information. Families: percentage premiums, elasticities, semi-elasticities, standardised effects, partial correlations, monetary marginal prices. Pool only within comparable groups; document and propagate uncertainty.

  8. META-ANALYTICAL ECONOMETRICS. 8.1 random-effects (REML) with pooled estimate, SE, 95% CI, 95% PI, τ², I², k studies, n estimates; 8.2 multilevel meta-regression (estimate → model → sample → study); 8.3 robust variance estimation / cluster-robust inference; 8.4 moderators (country, continent, publication year, data period, property type, price vs rent, transaction vs listing, specification, spatial method, income level, urbanisation, market cycle, publication type); 8.5 temporal evolution (data period, not publication year); 8.6 publication selection (funnel, small-study effects, PET-PEESE where appropriate, selection models, published vs WP) without mechanical corrections; 8.7 Bayesian hierarchical models with convergence diagnostics; 8.8 ML meta-regression (random forests, boosting, explainable) with study-grouped CV, exploratory.

  9. AI AND HEDONIC VALUATION. Dedicated synthesis comparing traditional, spatial, regularised, random forests, boosting, neural networks, NLP-, CV-enhanced and multimodal models; separate predictive performance from interpretable marginal effects; never equate feature importance, SHAP, coefficients and causal effects; harmonise metrics where feasible; incremental information of AI-derived characteristics; no manufactured comparisons across incompatible datasets.

  10. ROBUSTNESS AND VALIDATION. Leave-one-study-out, leave-one-country-out, published vs WP, quality tiers, alternative transformations, heterogeneity estimators, study-level vs multilevel pooling, clustering, extreme estimates, period subsamples, geographic groupings, missing SEs. Extraction validation: random sample of tables independently verified; report precision, recall, numerical accuracy, error rates; unverified or low-confidence extractions excluded from the primary synthesis.

  11. REQUIRED TABLES AND FIGURES. Tables: coverage/selection; geography; study characteristics; taxonomy; most studied characteristics; pooled structural effects; pooled neighbourhood/environmental effects; geographic, temporal and specification meta-regressions; publication-selection diagnostics; AI evidence; robustness; comparison with Sirmans et al.; gaps and quality. Figures: PRISMA; annual trends; world map; variable-frequency heatmap; forest plots; geographic heterogeneity; temporal evolution; funnel plots; moderator plots; evidence density; AI methodology evolution; summary. Never synthetic results.

  12. MANUSCRIPT. Introduction; theoretical foundations; historical development; RQs and contributions; search and selection; extraction and harmonisation; meta-analytical framework; global evidence; geographic heterogeneity; temporal variation; methodological heterogeneity; AI; publication selection and robustness; implications; limitations; conclusion; bibliography; appendices. Explain precisely how the paper extends Sirmans et al. (2005, 2006); distinguish robust from tentative findings; no promotional language.

  13. LATEX / UQO STYLE. Follow the series conventions; main.tex, references.bib, sections, tables, figures, appendices, reproducibility statement; numbered equations; all tables/figures from the pipeline; verify every reference; compile to PDF.

  14. ARCHITECTURE. wp21_uqo/{CLAUDE.md, README.md, protocol/{preregistration,eligibility,search_strategy}.md, data/{raw,metadata,extracted,validated,final}, scripts/{discovery,screening,extraction,harmonization,validation,econometrics, visualization}, paper/{main.tex,sections,tables,figures,appendices,references.bib}, reports/{search_audit,extraction_audit, methodological_decisions,limitations}.md, tests, requirements.txt}. Python for discovery/extraction/processing/visualisation; R (metafor, clubSandwich) where superior; reproducible environments, seeds, executable scripts.

  15. EXECUTION ORDER. A research design → B literature acquisition → C data construction → D empirical analysis → E scientific interpretation → F manuscript → G final audit. No results written before analysis; no fabricated evidence; document obstacles and continue with feasible work.