spb/spboucher.ai Public
spboucher.ai — personal website of Simon-Pierre Boucher.
TypeScript 93.4%
HTML 5.5%
CSS 1%
1/*2 research-details.ts3 spboucher.ai Web4 Author: Simon-Pierre Boucher5 Mail: contact@spboucher.ai6*/78/** Rich per-paper content for the /research/[slug] detail pages. */9export interface ResearchDetail {10 slug: string;11 repo: string;12 hero: { headline: string; subheadline: string };13 abstract: string[];14 findings: { value: string; label: string }[];15 contributions: { title: string; description: string }[];16 data: { title: string; description: string }[];17 methodology: { title: string; description: string }[];18 reproducibility: string[];19 keywords: string[];20}2122export const researchDetails: ResearchDetail[] = [23 {24 "slug": "wp11",25 "repo": "https://git.spboucher.ai/wp11_uqo",26 "hero": {27 "headline": "Random cross-validation is the wrong experiment for valuation models",28 "subheadline": "Twenty hedonic models raced on 514,212 Quebec sales under identical conditions — machine learning wins the random split and loses the forward-in-time one."29 },30 "abstract": [31 "Hedonic price models underpin mass appraisal, price indices, and housing research, yet specification choices — functional form, time effects, spatial controls, estimation method — are usually made ad hoc. This paper runs a controlled horse race of 20 models on 514,212 roll-matched Quebec residential sales (2021–2026) with assessor-grade structural attributes, scoring every model on the same price-level scoreboard via Duan smearing and Box–Cox inversion.",32 "The design crosses four axes: six functional forms (linear to Box–Cox and cubic splines), a time-fixed-effects ladder (none to month), a spatial-fixed-effects ladder (none to a ~1.1 km grid), and six estimation methods (OLS, ridge, random forest, gradient boosting, spatial k-NN comparables). All models face two holdouts: a random 80/20 split and a forward-in-time split trained before 2025 and tested on 2025–2026.",33 "Spatial controls are first-order, worth roughly 10 percentage points of median absolute percentage error, but overfit at ~1 km granularity; functional form is second-order, worth about ±2 points. Gradient boosting wins under random validation (MdAPE ≈ 14.4% vs 16.5% for the best linear model), but under the forward-in-time split every model degrades and the ranking reverses — splines reach 19.1% against 21.2% for boosting, at identical log R². Implied price indices and implicit-price profiles agree across forms far more than the accuracy gap suggests."34 ],35 "findings": [36 { "value": "514,212", "label": "roll-matched Quebec sales, 2021–2026" },37 { "value": "20", "label": "hedonic models raced on one scoreboard" },38 { "value": "~10 pp", "label": "MdAPE improvement from spatial fixed effects" },39 { "value": "±2 pp", "label": "MdAPE spread across functional forms" },40 { "value": "14.4%", "label": "gradient-boosting MdAPE under the random split" },41 { "value": "19.1%", "label": "spline MdAPE under the forward split — beating ML's 21.2%" }42 ],43 "contributions": [44 {45 "title": "A controlled horse race, not anecdotes",46 "description": "Six functional forms × time-FE ladder × spatial-FE ladder × six estimation methods, all on the same sales, attributes, and price-level metrics — specification advice grounded in one consistent experiment."47 },48 {49 "title": "Spatial controls are first-order",50 "description": "Location fixed effects are worth roughly 10 percentage points of MdAPE — an order of magnitude more than functional form — but a ~1.1 km grid begins to overfit."51 },52 {53 "title": "The validation split is the result",54 "description": "Gradient boosting beats every linear model under random 80/20 validation, yet under a forward-in-time split the advantage reverses: splines at 19.1% MdAPE versus boosting at 21.2%."55 },56 {57 "title": "Functional form matters less than assumed",58 "description": "Across linear, semi-log, log-log, Box–Cox, quadratic, and spline specifications, price-level accuracy moves by only about ±2 percentage points of MdAPE."59 },60 {61 "title": "Economics is stable across specifications",62 "description": "Implied price indices and implicit-price profiles agree across functional forms far more closely than the accuracy differences suggest — the economics is robust even where the forecasts are not."63 },64 {65 "title": "A fair price-level scoreboard",66 "description": "Log and transformed models are mapped back to dollar space with Duan smearing and Box–Cox inversion, so every model is judged on MdAPE, MAPE, and RMSE in levels."67 }68 ],69 "data": [70 {71 "title": "Roll-matched Quebec sales",72 "description": "514,212 residential transactions (2021–2026), filtered from 745,119 raw sales and matched to assessment rolls for assessor-grade structural attributes — floor area, lot, age, storeys, units, class, physical link."73 }74 ],75 "methodology": [76 {77 "title": "Functional-form ladder",78 "description": "Linear, semi-log, log-log, Box–Cox with λ estimated by profile likelihood, quadratic terms, and cubic splines — curvature tested systematically rather than assumed."79 },80 {81 "title": "Fixed-effects ladders",82 "description": "Time effects from none to month dummies; spatial effects from none through municipality to ~5.5 km and ~1.1 km grid cells, isolating where granularity pays and where it overfits."83 },84 {85 "title": "Estimation methods",86 "description": "OLS, cross-validated ridge, random forest, gradient boosting with and without coordinates, and spatial k-NN comparables — the machine-learning contenders face the same attributes as the linear models."87 },88 {89 "title": "Two holdout designs",90 "description": "A random 80/20 split and a forward-in-time split (train before 2025, test 2025–2026 with carry-forward time effects) — the contrast between them is itself the headline result."91 },92 {93 "title": "Municipality-clustered evaluation",94 "description": "Metrics computed on price levels (MdAPE, MAPE, RMSE_ln, R²_ln) with municipality-clustered design throughout."95 }96 ],97 "reproducibility": [98 "End-to-end Python pipeline: five numbered scripts from raw sample construction (745,119 → 514,212 sales) to LaTeX tables",99 "The 20-model harness lives in src/wp11/models.py with a shared config and plot style",100 "9 journal-calibre figures and 7 LaTeX tables regenerated by scripts 04 and 05",101 "Full LaTeX source of the 23-page paper in paper/ with sections and bibliography"102 ],103 "keywords": [104 "hedonic pricing",105 "model validation",106 "machine learning",107 "gradient boosting",108 "spatial fixed effects",109 "Box–Cox",110 "housing markets",111 "Quebec"112 ]113 },114 {115 "slug": "wp2",116 "repo": "https://git.spboucher.ai/wp2_uqo",117 "hero": {118 "headline": "Listing narratives carry named, priceable information",119 "subheadline": "Do agent-written property descriptions contain price-relevant information beyond structural attributes? Yes — 20 interpretable semantic scores lift adjusted R² from 0.452 to 0.511."120 },121 "abstract": [122 "Hedonic pricing models decompose a property's price into the implicit prices of structured characteristics — bedrooms, bathrooms, lot size — but the narrative of a listing carries quality information that no structured field captures. This paper embeds the free-text descriptions of 17,087 Quebec single-family house listings with a sentence transformer and projects each embedding onto 20 researcher-defined semantic reference descriptions (Luxury, Needs Renovation, Waterfront, Motivated Seller, and others) via cosine similarity.",123 "Adding the 20 named similarity scores to a log-price hedonic OLS model raises adjusted R² from 0.452 to 0.511, with the text block jointly significant (F = 99.53, p < 0.001). Each dimension carries a signed, economically meaningful implicit price: sounding Modern/Contemporary is worth +16.4% per standard deviation, Luxury +14.2%, while Motivated Seller and Needs Renovation carry discounts of 8.5% and 7.9%.",124 "The approach resolves the depth-versus-interpretability trade-off in text-based hedonics: unlike raw 384-dimensional embeddings or unstable LDA topics, every coefficient reads directly as the implicit price of a human-named concept. Quantile regressions show the Luxury premium rising monotonically from 9.7% at the 25th price percentile to 17.3% at the 75th, and a PCA benchmark explicitly quantifies the fit sacrificed for interpretability."125 ],126 "findings": [127 {128 "value": "0.452 → 0.511",129 "label": "adjusted R² gain from adding 20 semantic dimensions"130 },131 {132 "value": "17,087",133 "label": "Quebec single-family house listings analyzed"134 },135 {136 "value": "F = 99.53",137 "label": "joint significance of the 21 text variables (p < 0.001)"138 },139 {140 "value": "+16.4%",141 "label": "Modern/Contemporary implicit price per +1 SD"142 },143 {144 "value": "+14.2%",145 "label": "Luxury premium per +1 SD (9.7% → 17.3% across price quantiles)"146 },147 {148 "value": "9.3%",149 "label": "share of Model D's explained variance attributable to semantics"150 }151 ],152 "contributions": [153 {154 "title": "Reference-based cosine projection method",155 "description": "Instead of 384 anonymous embedding dimensions, listings are projected onto 20 researcher-defined reference descriptions, yielding named, interpretable regressors — deep semantics with full economic readability."156 },157 {158 "title": "Named implicit prices for narrative content",159 "description": "Each semantic dimension carries a signed, economically meaningful coefficient: Modern/Contemporary +16.4%, Luxury +14.2%, Land & Nature +13.0%, Motivated Seller −8.5%, Needs Renovation −7.9% per standard deviation."160 },161 {162 "title": "Interpretability-versus-fit trade-off quantified",163 "description": "A PCA benchmark on raw embeddings fits better (ΔR² +0.070 vs +0.044) but is economically unreadable; the paper measures exactly what interpretability costs."164 },165 {166 "title": "Heterogeneity across the price distribution",167 "description": "Quantile regressions show the Luxury premium rising monotonically from 9.7% at the 25th percentile to 17.3% at the 75th, while urgency and condition discounts attenuate at the top."168 },169 {170 "title": "Full robustness battery with disclosed limits",171 "description": "Bootstrap (1,000 replications), outlier trimming, Lasso/Elastic-Net selection, and VIF analysis; multicollinearity among similarity dimensions (mean VIF 17.2) is disclosed and its inferential consequences stated."172 },173 {174 "title": "Fully machine-generated results",175 "description": "Every number in the paper's statistical tables is written by the pipeline from results CSVs — hand-transcription errors in the original project were caught and corrected."176 }177 ],178 "data": [179 {180 "title": "Quebec residential listings database (louka.db)",181 "description": "268 MB SQLite database of 46,479 Realtor.ca/Centris listings across five categories, each with structured fields plus the agent-written free-text description."182 },183 {184 "title": "Single-family house analysis sample",185 "description": "17,087 houses with positive price and description of at least 20 characters; median listing price $589,900, descriptions averaging 510 characters."186 },187 {188 "title": "20 semantic reference descriptions",189 "description": "Synthetic French paragraphs (146–257 characters each) spanning six domains — quality, condition, physical features, location, style, and market signals — reproduced verbatim in the paper's appendix."190 }191 ],192 "methodology": [193 {194 "title": "Sentence-transformer embeddings",195 "description": "All 17,087 descriptions encoded with all-MiniLM-L6-v2 (22.7M parameters, 384 dimensions, L2-normalized); the whole corpus embeds in seconds on Apple Silicon."196 },197 {198 "title": "Cosine-similarity projection",199 "description": "A single matrix product against the 20 embedded reference descriptions yields a 17,087 × 20 similarity matrix of named, interpretable semantic scores."200 },201 {202 "title": "Nested hedonic OLS (Models A–E)",203 "description": "Log-price regressions from structural-only to the full specification, all covariates standardized, HC3 robust errors, joint F-tests, and a parsimonious 16-dimension variant."204 },205 {206 "title": "Quantile regression",207 "description": "Estimates at the 25th, 50th, and 75th price percentiles reveal how semantic premia and discounts vary across the market's price distribution."208 },209 {210 "title": "Robustness suite",211 "description": "Bootstrap standard errors (1,000 replications, seeded), winsorization at the 1st/99th percentiles, VIF diagnostics, Breusch–Pagan tests, and Lasso/Elastic-Net variable selection."212 },213 {214 "title": "PCA benchmark",215 "description": "20 principal components of the raw embeddings serve as an upper-bound fit comparison, quantifying the cost of interpretability against anonymous dimensions."216 }217 ],218 "reproducibility": [219 "End-to-end Python pipeline: six numbered scripts take the raw SQLite database to figures, results CSVs, and machine-generated LaTeX tables.",220 "All stochastic steps seeded (seed 42); embeddings are deterministic and cached; validated bit-identical against the original project's similarity matrix and model coefficients.",221 "All 11 figures regenerated from the pipeline as vector PDF and 300-dpi PNG; the 53-page LaTeX paper compiles with zero errors and zero undefined references.",222 "86 bibliography entries, every one verified against OpenAlex and DOI records.",223 "AUDIT.md and CHANGES.md document a forensic audit of the original project, including every numeric correction with old and new values."224 ],225 "keywords": [226 "hedonic pricing",227 "semantic embeddings",228 "textual analysis",229 "real estate",230 "sentence transformers",231 "interpretable machine learning",232 "Quebec housing",233 "implicit prices"234 ]235 },236 {237 "slug": "wp3",238 "repo": "https://git.spboucher.ai/wp3_uqo",239 "hero": {240 "headline": "Random splits flatter machine learning in housing",241 "subheadline": "How well do hedonic methods really generalize? XGBoost's R² of 0.833 under a random split collapses to 0.425–0.547 when ten entire states are held out."242 },243 "abstract": [244 "This paper compares three frameworks for hedonic housing valuation on a single dataset of 788,842 active Zillow listings spanning all 50 U.S. states and the District of Columbia: semi-log OLS with 62 regressors, quantile regression at five points of the price distribution, and gradient-boosting models interpreted with SHAP. Each framework answers a distinct question — average capitalization gradients, distributional heterogeneity, and predictive performance.",245 "The core methodological contribution is a systematic study of spatial leakage in hedonic model evaluation. Random train/test splits let geographically proximate listings appear on both sides of the split, inflating performance: XGBoost reaches R² = 0.833 under a random 80/20 split but only 0.425–0.547 when ten entire states (438,315 listings) are held out. Removing all geographic features actually improves geographic-holdout performance, because region dummies memorize training-set price levels.",246 "Mean effects also mask substantial heterogeneity: inter-quantile Wald tests reject coefficient equality for 11 of 13 variables, the garage gradient is ten times larger at the bottom decile than at the top, and the lot-size gradient triples once state-median-imputed values are dropped — a warning for hedonic work on scraped listing data. All estimates are presented as listing-price capitalization gradients, not causal willingness-to-pay parameters."247 ],248 "findings": [249 {250 "value": "788,842",251 "label": "Zillow listings across 50 states + DC"252 },253 {254 "value": "0.833 vs 0.425–0.547",255 "label": "XGBoost R²: random split vs 10-state geographic holdout"256 },257 {258 "value": "11 of 13",259 "label": "variables where inter-quantile Wald tests reject coefficient equality"260 },261 {262 "value": "10×",263 "label": "garage gradient at τ = 0.10 relative to τ = 0.90 (z = 28.9)"264 },265 {266 "value": "0.2745",267 "label": "mean Moran's I of OLS residuals (all p < 0.001)"268 },269 {270 "value": "0.630 → 0.725",271 "label": "OLS out-of-sample R² from region dummies to 886 ZIP3 fixed effects"272 }273 ],274 "contributions": [275 {276 "title": "Spatial leakage in hedonic evaluation quantified",277 "description": "Three complementary designs — a 10-state geographic holdout, a 6-stage feature-ablation cascade, and a lat/lon augmentation experiment — measure how much random validation inflates machine-learning performance."278 },279 {280 "title": "Geographic features can hurt generalization",281 "description": "Removing all geographic features improves geographic-holdout R² from 0.425 to 0.519, while adding raw coordinates boosts random R² but degrades out-of-region prediction — region dummies memorize price levels."282 },283 {284 "title": "Spatial granularity dominates OLS fit",285 "description": "Moving from 4 Census-region dummies to state fixed effects to 886 ZIP3 fixed effects raises out-of-sample R² from 0.630 to 0.678 to 0.725 — a 9.5 point gain from geography alone."286 },287 {288 "title": "Distributional heterogeneity in attribute gradients",289 "description": "The garage gradient is 10× larger at the bottom price decile than at the top, the pool premium only emerges above the median, and the age penalty concentrates in lower-priced homes."290 },291 {292 "title": "SHAP rankings are model-stable, not structural",293 "description": "Spearman correlations of mean |SHAP| across XGBoost, LightGBM, and Random Forest range from 0.89 to 0.99, with the same six features on top — but these remain predictive decompositions, not implicit prices."294 },295 {296 "title": "Imputation sensitivity exposed",297 "description": "The lot-size gradient triples (0.018 to 0.059) when state-median-imputed observations are dropped — a caution for hedonic research built on scraped listing data."298 }299 ],300 "data": [301 {302 "title": "Zillow active for-sale listings (2025–2026 snapshot)",303 "description": "839,313 raw residential properties with 116 variables in a 1.6 GB DuckDB database, including price history, schools, and tax-history tables."304 },305 {306 "title": "Analytical sample",307 "description": "788,842 listings after five sequential filters (price $10K–$10M, living area 200–20,000 sqft, 1–10 beds/baths, valid coordinates; 6.0% attrition), with 68 engineered columns including six interaction terms."308 }309 ],310 "methodology": [311 {312 "title": "Semi-log OLS with HC3 errors",313 "description": "62 regressors on log listing price with fixed effects escalating from Census region to state to 886 ZIP3 areas; price-to-area elasticity of 0.63."314 },315 {316 "title": "Quantile regression",317 "description": "Estimated at τ = 0.10 to 0.90 on a 150,000-observation subsample with a 10-seed stability check and inter-quantile Wald z-tests on coefficient differences."318 },319 {320 "title": "Gradient boosting benchmark suite",321 "description": "XGBoost (1,000 trees, depth 8), LightGBM, Random Forest, Ridge, Lasso, and Elastic Net, evaluated under both random 80/20 and 10-state geographic-holdout validation."322 },323 {324 "title": "TreeSHAP interpretation",325 "description": "SHAP values on 10,000 test observations, with cross-model Spearman rank-stability analysis across XGBoost, LightGBM, and Random Forest."326 },327 {328 "title": "Spatial diagnostics",329 "description": "Moran's I on OLS residuals with row-standardized 8-nearest-neighbor weights, three independent 5,000-listing subsamples, and 999 permutations."330 },331 {332 "title": "Feature-ablation cascade",333 "description": "Six nested feature sets from structural-only to the full specification isolate each block's predictive contribution; neighborhood scores alone add 17.6 R² points."334 }335 ],336 "reproducibility": [337 "Complete research compendium: analysis code, LaTeX source, 11 publication-ready figures, and 13 machine-readable CSV tables mirroring every table in the paper.",338 "Every script prints verification statistics against the stored result artifacts; the Moran's I pipeline reproduces stored values to four decimal places.",339 "A systematic audit reconciled the manuscript against stored artifacts, fixing six internal inconsistencies and filling 13 placeholder table cells — with no scientific result altered.",340 "Paths resolve relative to the repository root (WP3_ROOT-overridable); pinned dependencies validated 2026-08-05; the 60-page paper compiles with zero errors.",341 "Data artifacts (up to 1.6 GB) exceed GitHub limits and are available on request, subject to Zillow's Terms of Service."342 ],343 "keywords": [344 "hedonic pricing",345 "housing markets",346 "machine learning",347 "quantile regression",348 "spatial leakage",349 "XGBoost",350 "SHAP",351 "United States"352 ]353 },354 {355 "slug": "wp5",356 "repo": "https://git.spboucher.ai/wp5_uqo",357 "hero": {358 "headline": "Nearby Airbnb activity is priced into Quebec rents",359 "subheadline": "Is short-term rental density associated with higher residential rents? Each additional Airbnb listing within 500 m corresponds to roughly +0.4% monthly rent."360 },361 "abstract": [362 "This paper investigates the relationship between Airbnb short-term rental activity and residential rents in Quebec, Canada, using cross-sectional microdata on 3,456 cleaned Airbnb listings and 8,303 rental listings. For every rental unit, Airbnb exposure is measured within 250 m, 500 m, 1 km, and 2 km buffers via Haversine distances, and a hedonic pricing framework augmented with spatial econometric techniques quantifies the conditional association between nearby Airbnb presence and monthly rents.",363 "Baseline estimates indicate that an additional Airbnb listing within 500 m is associated with a statistically significant rent increase of approximately 0.3–0.5%, controlling for dwelling characteristics, building type, and city fixed effects — about $6–10 per month at the median rent, the same order of magnitude as Berlin's quasi-experimental evidence. The per-listing association decays monotonically with distance, from 0.93% at 250 m to 0.06% at 2 km.",364 "Quantile regressions show the association is strongest at the upper tail of the rent distribution (0.0047 at the 90th percentile, about 25% above OLS), and the coefficient survives spatial autoregressive and spatial error models estimated by GMM with only mild attenuation. The paper explicitly cautions that these cross-sectional associations should not be read as causal effects, and discusses implications for housing affordability and short-term rental regulation."365 ],366 "findings": [367 {368 "value": "+0.4%",369 "label": "monthly rent per additional Airbnb listing within 500 m (0.3–0.5% across specifications)"370 },371 {372 "value": "8,303 + 3,456",373 "label": "cleaned rental listings and Airbnb listings across Quebec"374 },375 {376 "value": "0.93% → 0.06%",377 "label": "monotonic spatial decay of the association from 250 m to 2 km"378 },379 {380 "value": "0.0047 at τ = 0.90",381 "label": "quantile coefficient at the top of the rent distribution, ~25% above OLS"382 },383 {384 "value": "ρ̂ = 0.137***",385 "label": "spatial lag parameter; coefficient survives SAR/SEM with mild attenuation (0.0034)"386 },387 {388 "value": "0.0038–0.0040",389 "label": "leave-one-city-out coefficient range, including dropping Montreal"390 }391 ],392 "contributions": [393 {394 "title": "Canadian evidence on the Airbnb–rent link",395 "description": "Fills a documented gap in Canadian and Quebec evidence on short-term rentals and housing costs, benchmarking magnitudes against Berlin, Los Angeles, and Boston quasi-experiments."396 },397 {398 "title": "Multi-radius spatial exposure measurement",399 "description": "Six exposure metrics (count, density, mean price, entire-home share, mean rating, superhost share) computed at four Haversine buffer radii yield 24 exposure variables per rental listing."400 },401 {402 "title": "Distance-decay gradient documented",403 "description": "The per-listing rent association falls monotonically from 0.93% at 250 m to 0.06% at 2 km, consistent with a genuinely local neighborhood-level channel."404 },405 {406 "title": "Robustness across estimators and samples",407 "description": "The coefficient stays positive and significant at 1% in every check: SAR/SEM spatial models, city-clustered errors, ring decomposition, trimming, and leave-one-city-out including Montreal."408 },409 {410 "title": "Complementary Airbnb pricing model",411 "description": "A hedonic model of nightly prices shows short-term rental pricing is driven by listing characteristics, with a superhost discount that contrasts with premia in prior work — discussed, not hidden."412 },413 {414 "title": "Candid identification discussion",415 "description": "Endogeneity and identification limits are addressed head-on: cross-sectional associations are explicitly framed as conditional capitalization gradients, not causal effects, with seven stated limitations."416 }417 ],418 "data": [419 {420 "title": "Airbnb listings (Quebec)",421 "description": "Roughly 5,000 scraped listings cleaned to 3,456, with nightly price, coordinates, property type, rating, review counts, and superhost status."422 },423 {424 "title": "Realtor.ca rental listings",425 "description": "8,356 scraped listings cleaned to 8,303, with monthly rent, coordinates, bedrooms, bathrooms, building type, and unit size, covering the province of Quebec."426 },427 {428 "title": "Spatially merged exposure dataset",429 "description": "For each rental, six Airbnb exposure metrics at four buffer radii (250 m to 2 km) computed via chunked vectorized Haversine distances, plus city-level aggregates for 153 cities."430 }431 ],432 "methodology": [433 {434 "title": "Hedonic rent regressions",435 "description": "Log monthly rent on Airbnb exposure plus dwelling controls, building type, and city fixed effects (Models 1a–1e), estimated by OLS with HC1 robust errors."436 },437 {438 "title": "Spatial econometrics (SAR/SEM)",439 "description": "Spatial lag and spatial error models with row-standardized KNN(5) weights, estimated by Kelejian–Prucha GMM via PySAL/spreg; ρ̂ = 0.137, λ̂ = 0.539."440 },441 {442 "title": "Quantile regression",443 "description": "Estimated at τ ∈ {0.10, 0.25, 0.50, 0.75, 0.90} plus a fine grid, tracing the exposure gradient from 0.0037 to 0.0047 across the rent distribution."444 },445 {446 "title": "Machine-learning benchmark with SHAP",447 "description": "OLS, LASSO, Elastic Net, Random Forest, and GBM under an 80/20 split (RF test R² = 0.71), with SHAP attribution confirming the exposure variable's role."448 },449 {450 "title": "Extended robustness programme",451 "description": "Nine checks: alternative radii and exposures, subsamples, trimming, city-clustered standard errors, size control, log(1+count) form, ring decomposition, and leave-one-city-out."452 }453 ],454 "reproducibility": [455 "Eleven-script numbered Python pipeline from raw data inspection to extended robustness; fully deterministic (seed 42), most steps run in under 15 seconds on Apple Silicon.",456 "Committed processed parquet files make steps 04–11 and the entire paper reproducible without the raw scraped data, which cannot be redistributed.",457 "Rebuilt under a strict no-result-changes policy: 3/3 merged parquets value-identical and 14 of 16 tables byte-identical to the originals, with residual float-level diffs documented.",458 "All 32 references added in the scholarly upgrade verified via Crossref/JMLR DOIs, with per-reference justification; 56 references in total.",459 "The 51-page LaTeX paper builds via latexmk with zero unresolved references; 13 LaTeX table fragments and 16 publication PDF figures regenerate from the pipeline."460 ],461 "keywords": [462 "Airbnb",463 "short-term rentals",464 "housing rents",465 "hedonic pricing",466 "spatial econometrics",467 "quantile regression",468 "housing affordability",469 "Quebec"470 ]471 },472 {473 "slug": "wp7",474 "repo": "https://git.spboucher.ai/wp7_uqo",475 "hero": {476 "headline": "What 3.8 Billion Option Contracts Know About Tomorrow",477 "subheadline": "Do option-implied moments predict returns and volatility across assets? Weekly returns yes, daily no — and implied volatility dominates every realized-volatility benchmark."478 },479 "abstract": [480 "This paper asks five questions about the information content of equity option markets, answered with 3.83 billion option contracts on 11,077 underlyings (2010–2025) merged with 11.5 billion intraday OHLCV observations. The estimation panel covers 264,383 ticker-days across 69 tickers.",481 "Implied moments predict cross-sectional returns at the weekly horizon (R² of 4.8–19.3%) but not daily; a long/short portfolio sorted on implied kurtosis earns a Sharpe ratio of 2.33 (t = 19.8). Adding the IV surface to a HAR model raises 1-day realized-volatility forecasting R² by +23.3%, robust across all nine subperiods, and IV Granger-causes RV for 100% of tickers.",482 "The implied-minus-realized correlation ratio predicts market stress at 5–20 day horizons (t-statistics 3.91–8.30). The popular max-open-interest \"price magnet\" hypothesis is rejected (47.0% hit rate, below the 50% coin flip), and on the SPX surface a simple HAR-RV model beats random forests, gradient boosting, and the VIX out of sample."483 ],484 "findings": [485 {486 "value": "3.83B",487 "label": "Option contracts analyzed (11,077 underlyings, 2010–2025)"488 },489 {490 "value": "2.33",491 "label": "Sharpe ratio of the implied-kurtosis long/short strategy (t = 19.8)"492 },493 {494 "value": "+23.3%",495 "label": "Gain in 1-day RV forecasting R² from adding the IV surface to HAR"496 },497 {498 "value": "100%",499 "label": "Share of tickers where IV Granger-causes realized volatility (F = 62.4)"500 },501 {502 "value": "73.8%",503 "label": "RV forecast-error variance explained by IV shocks at 20 days (FEVD)"504 },505 {506 "value": "47.0%",507 "label": "Max-OI \"price magnet\" hit rate — below 50%, hypothesis rejected"508 }509 ],510 "contributions": [511 {512 "title": "Weekly, not daily, return predictability",513 "description": "Implied moments predict cross-sectional returns at the weekly horizon (R² 4.8–19.3%) but carry no daily signal; the kurtosis long/short portfolio delivers a 2.33 Sharpe ratio."514 },515 {516 "title": "IV surface dominates HAR and GARCH",517 "description": "Augmenting HAR-RV with implied-volatility surface features improves 1-day forecasting R² by +23.3%, a gain that survives in all nine subperiods examined."518 },519 {520 "title": "Implied–realized correlation as a stress gauge",521 "description": "The ratio of implied to realized correlation predicts market stress at 5–20 day horizons, with t-statistics ranging from 3.91 to 8.30."522 },523 {524 "title": "Price-magnet hypothesis rejected",525 "description": "Prices gravitate toward maximum-open-interest strikes only 47.0% of the time — below chance — refuting a widespread piece of options-market folklore."526 },527 {528 "title": "Simple models beat ML on the SPX surface",529 "description": "For SPX realized-volatility forecasting, HAR-RV beats random forests, gradient boosting, and the VIX out of sample; two-week ATM IV alone captures 50.8% of feature importance."530 },531 {532 "title": "Placebo-validated information content",533 "description": "A placebo design confirms the results are not mechanical: shuffled predictors yield R² of 0.0008 against 0.053 for the actual implied moments."534 }535 ],536 "data": [537 {538 "title": "Equity options database",539 "description": "3.83 billion option contracts across 11,077 underlyings, 2010–2025, stored in external DuckDB stores and distilled into five derived parquet files."540 },541 {542 "title": "Intraday OHLCV database",543 "description": "11.5 billion intraday open-high-low-close-volume observations merged with the options data to construct realized-volatility measures."544 },545 {546 "title": "Estimation panel",547 "description": "264,383 ticker-days over 69 tickers, 2010–2025 — the analysis sample behind all cross-sectional and time-series results."548 }549 ],550 "methodology": [551 {552 "title": "Fama-MacBeth cross-sectional regressions",553 "description": "Panel regressions of daily and weekly returns on option-implied moments, with Newey-West, clustered-SE, and quantile robustness checks."554 },555 {556 "title": "HAR-RV and GARCH forecasting horse race",557 "description": "Realized-volatility forecasts from HAR and GARCH benchmarks compared against IV-surface-augmented models across nine subperiods and rolling windows."558 },559 {560 "title": "Granger causality, VAR, IRF and FEVD",561 "description": "Vector autoregressions establish that IV leads RV for every ticker; forecast-error variance decompositions attribute 73.8% of 20-day RV variance to IV shocks."562 },563 {564 "title": "Portfolio sorts and double sorts",565 "description": "Single and double sorts on implied moments, including decile sorts and Spearman information coefficients, generate the long/short trading strategies."566 },567 {568 "title": "Machine learning on the SPX surface",569 "description": "Random forests and gradient boosting trained on implied-volatility surface features, benchmarked against HAR-RV and the VIX out of sample."570 },571 {572 "title": "Extended robustness battery",573 "description": "Winsorization and Newey-West lag sensitivity, leave-one-year-out stability, and a placebo test with shuffled predictors validate the headline findings."574 }575 ],576 "reproducibility": [577 "Fully scripted pipeline: 13 numbered Python entry points (make pipeline | figures | paper) with a reusable src/wp7 package.",578 "All 28 regenerable result files were re-run and verified on 2026-08-05: 10 byte-identical, 15 equal to floating-point noise, 1 explained deviation, 2 newly regenerated.",579 "Every table in the 29-page paper derives from one of 41 shipped CSV result tables; raw-dependent steps skip gracefully when the DuckDB stores are absent.",580 "Reproduction evidence (rerun logs and a CSV comparator) ships in _verify/, with a full audit trail in AUDIT.md and CHANGES.md.",581 "LaTeX source is modular (one file per section, 45 BibTeX references, all cited) and builds with latexmk via make paper."582 ],583 "keywords": [584 "options-implied moments",585 "implied volatility surface",586 "realized volatility forecasting",587 "HAR-RV",588 "cross-sectional return predictability",589 "Granger causality",590 "portfolio sorts",591 "machine learning"592 ]593 },594 {595 "slug": "wp9",596 "repo": "https://git.spboucher.ai/wp9_uqo",597 "hero": {598 "headline": "Location Is Worth More Than Every Wall Combined",599 "subheadline": "What do structure and neighbourhood each contribute to Canadian home prices? Absorbing 1,153 neighbourhood fixed effects lifts explained variance from 46% to 77%."600 },601 "abstract": [602 "A dwelling is the archetypal heterogeneous good, and its most important attribute — location — cannot be observed as a scalar. This paper estimates a semi-logarithmic hedonic price equation at national scale for Canada on 140,931 MLS listings (82,334 houses, 57,857 condos, nine provinces), absorbing 1,153 Forward Sortation Area fixed effects so that structural implicit prices are identified purely from within-neighbourhood variation.",603 "The specification ladder is decisive: structural attributes alone explain 46.4% of log-price variance; adding province effects reaches 56.7%; the grand model with neighbourhood fixed effects reaches 76.7%. Location alone is worth roughly 30 percentage points of R² — more than every structural attribute combined. The living-area elasticity is 0.547, each full bathroom adds about 11%, and bedrooms conditional on area are worth approximately zero.",604 "Out of sample, the model values held-out homes with a median absolute error of 15.8% (OOS R² = 0.764), competitive with commercial AVMs while remaining fully transparent. Neighbourhood premia span a factor of nine: the most expensive FSAs net of structure are all in Vancouver (+150–200% versus the national median); the cheapest sit in rural Saskatchewan, Manitoba, and Newfoundland (−60 to −67%)."605 ],606 "findings": [607 {608 "value": "140,931",609 "label": "MLS listings in the estimation sample, across 9 provinces"610 },611 {612 "value": "46% → 77%",613 "label": "Explained log-price variance after absorbing 1,153 neighbourhood fixed effects"614 },615 {616 "value": "0.547",617 "label": "Living-area elasticity in the grand model (cluster SE 0.009)"618 },619 {620 "value": "15.8%",621 "label": "Median absolute out-of-sample valuation error (OOS R² = 0.764)"622 },623 {624 "value": "+11%",625 "label": "Price premium per full bathroom (0.109 log points)"626 },627 {628 "value": "×9",629 "label": "Span of neighbourhood premia between the most and least expensive FSAs"630 }631 ],632 "contributions": [633 {634 "title": "National-scale variance decomposition",635 "description": "Quantifies structure versus location for the entire Canadian market: neighbourhood identity contributes about 30 percentage points of R², exceeding all structural attributes combined."636 },637 {638 "title": "High-dimensional fixed-effects hedonic model",639 "description": "Absorbs 1,153 FSA intercepts via absorbing least squares, identifying implicit prices from within-neighbourhood variation with standard errors clustered by FSA."640 },641 {642 "title": "Transparent AVM-grade valuation accuracy",643 "description": "On a held-out 20% sample the model achieves a 15.8% median absolute error with 59% of homes priced within ±20% — competitive with commercial automated valuation models."644 },645 {646 "title": "The textbook bedroom result, confirmed",647 "description": "Conditional on floor area, bedroom count is worth approximately zero, while living-area elasticity of 0.547 and an 11% full-bathroom premium dominate structural pricing."648 },649 {650 "title": "Spatial diagnostics validate the decomposition",651 "description": "Moran's I of residuals falls from 0.46 to 0.08 (−82%) once neighbourhood effects are absorbed, showing FSA intercepts capture nearly all spatial price structure."652 },653 {654 "title": "Urban gradient and neighbourhood ranking",655 "description": "Location premia decline 8.5% per doubling of distance to the nine major metros; a ranked national map places Vancouver FSAs (V6S, V8E, V6T) at +150–200%."656 }657 ],658 "data": [659 {660 "title": "Canadian MLS listings snapshot",661 "description": "De-duplicated DuckDB of 172,019 for-sale listings × 81 columns with list price, geocoded coordinates, postal code, and semi-structured building and lot attributes (747 MB, not redistributed)."662 },663 {664 "title": "Estimation sample",665 "description": "140,931 filtered and trimmed listings (committed as a 6 MB parquet): median list price ≈ $639,888, median living area ≈ 135 m², 3 bedrooms, 2 full bathrooms."666 }667 ],668 "methodology": [669 {670 "title": "Specification ladder M1–M5",671 "description": "Five nested hedonic models, from structural attributes only to the grand model with FSA fixed effects; the M3-to-M5 R² gap measures the value of resolving location at neighbourhood scale."672 },673 {674 "title": "Absorbing least squares",675 "description": "linearmodels AbsorbingLS sweeps out 1,153 FSA intercepts without materializing dummies — numerically identical to full-dummy OLS — with FSA-clustered standard errors throughout."676 },677 {678 "title": "Duan smearing retransformation",679 "description": "Level predictions from the log model use Duan's (1983) smearing estimator, avoiding any log-normality assumption in out-of-sample valuation."680 },681 {682 "title": "Out-of-sample and transferability validation",683 "description": "Random 80/20 split restricted to FSAs seen in training, plus leave-one-province-out cross-validation with province-specific intercepts."684 },685 {686 "title": "Quantile and nonlinearity extensions",687 "description": "Quantile hedonic regressions across τ = 0.1–0.9, a quadratic test of diminishing returns to floor space, and an urban price gradient in metro distance."688 },689 {690 "title": "Moran's I spatial diagnostics",691 "description": "Row-standardized k-NN weights (k = 10) on a 15,000-listing sample with 199 permutations test residual spatial autocorrelation before and after absorbing FSA effects."692 }693 ],694 "reproducibility": [695 "End-to-end five-script pipeline (build sample, core estimation, extended estimation, figures, tables) running in under 10 minutes on Apple Silicon.",696 "Two results tiers keep the published record intact: results/reference/ holds the original outputs, results/reproduced/ the regenerated ones, switchable via a --results flag.",697 "All six LaTeX tables consumed by the 26-page paper are verified numerically identical to the originally published versions.",698 "The lost upstream cleaning code was reconstructed from the paper's data section, reproducing the published sample to within +0.86% and the full R² ladder to the second decimal; residual gaps are flagged in AUDIT.md, not hidden.",699 "The committed 6 MB estimation parquet lets anyone run estimation, figures, and tables (steps 02–05) without the 747 MB raw DuckDB."700 ],701 "keywords": [702 "hedonic pricing",703 "housing markets",704 "neighbourhood fixed effects",705 "Canada",706 "MLS listings",707 "automated valuation",708 "spatial econometrics",709 "Moran's I"710 ]711 },712 {713 "slug": "wp10",714 "repo": "https://git.spboucher.ai/wp10_uqo",715 "hero": {716 "headline": "Quebec's Cheapest Homes Pay 65% Too Much Tax",717 "subheadline": "Are municipal assessments equitable? Matching 522,769 sales to the roll shows systematic regressivity — 99% of municipalities fail the IAAO uniformity standard."718 },719 "abstract": [720 "Quebec taxes every dwelling in proportion to its assessed value, redrawn on a triennial roll that must by statute reflect market conditions at a single reference date. If assessments are regressive — cheap homes overvalued relative to expensive ones — the effective tax rate silently falls with wealth. Matching 522,769 residential sales (2021–2026) at the parcel level to the assessment roll in force at sale, this paper delivers the first province-wide audit of property-assessment equity in Canada.",721 "Within the same municipality × roll × sale-year cell, the elasticity of the assessment ratio with respect to price is −0.34 under Cheng fixed-effects estimation, and remains −0.08 under Clapp's measurement-error-robust rank instrument — regressivity is real, not a statistical artifact. The quantile profile shows failure concentrated at the top: β falls from 0.87 at the 10th percentile to 0.49 at the 90th.",722 "The consequences are stark. Ninety-nine percent of municipalities have a negative price-related bias, 95% fall below the IAAO vertical-equity band, and the median municipal COD of 26 far exceeds the IAAO ceiling of 15. The median dwelling in the bottom local price decile pays roughly 65% more property tax than uniform assessment would imply, while the top decile pays about 5% less. Montréal is the lone progressive large market, with PRB +0.07 in all six years."723 ],724 "findings": [725 {726 "value": "522,769",727 "label": "Sales matched to the assessment roll, across 625 municipalities"728 },729 {730 "value": "−0.34",731 "label": "Elasticity of the assessment ratio w.r.t. price (Cheng fixed effects)"732 },733 {734 "value": "−0.08",735 "label": "Elasticity under the measurement-error-robust Clapp rank IV"736 },737 {738 "value": "99%",739 "label": "Municipalities failing the IAAO uniformity standard (PRB < 0)"740 },741 {742 "value": "+65%",743 "label": "Excess property tax paid by the median bottom-decile dwelling"744 },745 {746 "value": "26",747 "label": "Median municipal COD — versus the IAAO ceiling of 15"748 }749 ],750 "contributions": [751 {752 "title": "First province-wide equity audit in Canada",753 "description": "Parcel-level matching of 522,769 sales to the triennial assessment rolls (median match distance 0.6 m) produces the first Canadian province-scale test of assessment equity."754 },755 {756 "title": "Regressivity survives measurement-error correction",757 "description": "The Clapp rank IV, designed to purge the attenuation bias that inflates naive ratio studies, still finds γ = −0.083 — regressivity is genuine, not spurious."758 },759 {760 "title": "Failure is concentrated at the top",761 "description": "Quantile regressions show β falling from 0.87 at the 10th percentile to 0.49 at the 90th: expensive homes are the ones assessed furthest below market."762 },763 {764 "title": "Near-universal IAAO standard failure",765 "description": "99% of municipalities show negative price-related bias and 95% fall outside the IAAO vertical-equity band; the median COD of 26 breaches the uniformity ceiling of 15."766 },767 {768 "title": "Exact intra-municipal tax-shift calculation",769 "description": "Under Quebec's exemption-free ad valorem rule, the bottom local price decile overpays about 65% while the top decile underpays about 5% — a silent redistribution."770 },771 {772 "title": "The Montréal exception",773 "description": "Montréal is the only progressive large market, with a positive PRB of +0.07 in all six sample years, possibly reflecting borough-level composition."774 },775 {776 "title": "Subgroup anatomy of the gap",777 "description": "Regressivity is strongest for plexes (γ = −0.49), single-family homes (−0.45), high land share (−0.50) and old dwellings (−0.43); condos are mildest at −0.12."778 }779 ],780 "data": [781 {782 "title": "Matched sales–roll snapshot",783 "description": "745,119 residential transactions (Jan 2021 – Jul 2026) matched at the parcel level to the assessment roll in force at sale; every retained sale reproduces the roll value exactly."784 },785 {786 "title": "Assessment-roll attributes",787 "description": "Total, land, and building assessed values, lot and floor areas, year built, unit count, CUBF use code, roll vintage, and the statutory market-condition date."788 },789 {790 "title": "Final estimation sample",791 "description": "522,769 sales in 625 municipalities and 2,884 municipality × roll × sale-year cells, after residential filtering, match-quality screens, 1/99% ratio trimming, and a 20-sale cell minimum."792 }793 ],794 "methodology": [795 {796 "title": "IAAO ratio-study battery",797 "description": "Median assessment ratio, COD, PRD, and PRB with percentile-bootstrap confidence intervals, computed for every municipality with at least 100 usable sales."798 },799 {800 "title": "Cheng (1974) log-log fixed effects",801 "description": "ln AV regressed on ln SP with 2,884 absorbed market-timing cells (AbsorbingLS), standard errors clustered on 625 municipalities; γ = β − 1 measures vertical inequity."802 },803 {804 "title": "Clapp (1990) rank instrument",805 "description": "A three-valued instrument built from within-cell rank agreement of ln AV and ln SP, estimated by 2SLS on demeaned data, provides the conservative measurement-error-robust bound."806 },807 {808 "title": "Quantile and heterogeneity regressions",809 "description": "Quantile regressions on within-cell demeaned data (τ = 0.10–0.90) and separate fixed-effects estimates by dwelling class, age, land share, roll lag, market size, and sale year."810 },811 {812 "title": "Horizontal-inequity regressions",813 "description": "The absolute deviation of the log ratio from its cell median is regressed on property characteristics to measure dispersion in assessment quality among comparable homes."814 },815 {816 "title": "Exact tax-shift computation",817 "description": "Ratio relative to the cell median by within-cell price decile yields exact percentage over- and under-payment under Quebec's exemption-free ad valorem property tax."818 }819 ],820 "reproducibility": [821 "Five-script end-to-end pipeline (sample, IAAO stats, regressions, figures, tables) runs in about 5 minutes on an Apple Silicon laptop with roughly 6 GB peak RAM.",822 "A reusable src/wp10 package separates IAAO diagnostics (iaao.py), econometric models (models.py), sample construction, and a validated journal-calibre plot style.",823 "Every CSV the scripts emit is committed under results/reproduced/, and 10 LaTeX tables plus 10 publication figures (300 dpi) feed the 29-page paper directly.",824 "The paper builds with latexmk from modular LaTeX sections via a Makefile."825 ],826 "keywords": [827 "property tax",828 "assessment equity",829 "vertical inequity",830 "IAAO ratio studies",831 "Quebec",832 "Clapp rank IV",833 "fixed effects",834 "housing valuation"835 ]836 },837 {838 "slug": "chapter-1",839 "repo": "https://git.spboucher.ai/phd_thesis",840 "hero": {841 "headline": "Speculators Dampen Macro News Shocks in Energy Futures",842 "subheadline": "Does speculative trading amplify macroeconomic news in commodity futures? No — greater speculation dampens price drift, volatility, and spreads while improving liquidity."843 },844 "abstract": [845 "Using five-minute futures data from 2007 to 2024 and 26 macroeconomic announcement releases, this essay asks whether speculative trading amplifies or dampens the impact of macro news on commodity futures. Speculation intensity is measured with an NLS proxy built from the CFTC disaggregated Commitments of Traders report, distinguishing money managers from swap dealers across energy (crude oil, natural gas) and metals (gold, silver, copper, palladium) contracts.",846 "The evidence points to a stabilizing role for speculators: increased speculative trading dampens the impact of standardized macro surprises on price drift, volatility, and bid-ask spreads, improving liquidity and price discovery. The damping effect is stronger for procyclical commodities such as oil and natural gas than for safe havens like gold, and the beneficial effects are driven by money managers rather than swap dealers."847 ],848 "findings": [849 {850 "value": "26",851 "label": "Macroeconomic announcement releases studied"852 },853 {854 "value": "6",855 "label": "Futures contracts (CL, NG, GC, SI, HG, PA)"856 },857 {858 "value": "5-min",859 "label": "Sampling frequency of futures data, 2007–2024"860 },861 {862 "value": "14",863 "label": "Tables of results in the article"864 },865 {866 "value": "6",867 "label": "Figures documenting speculation and news effects"868 }869 ],870 "contributions": [871 {872 "title": "Speculation conditions macro news impact",873 "description": "Shows that the intensity of speculative trading systematically conditions how commodity futures react to standardized macroeconomic surprises, rather than treating announcement effects as uniform."874 },875 {876 "title": "Evidence that speculators stabilize markets",877 "description": "Increased speculative activity dampens the impact of macro surprises on price drift, volatility, and bid-ask spreads, countering the view that speculation amplifies commodity price shocks."878 },879 {880 "title": "Trader-type decomposition of effects",881 "description": "Using the CFTC disaggregated Commitments of Traders data, the beneficial liquidity and price-discovery effects are attributed to money managers, not swap dealers."882 },883 {884 "title": "Procyclical versus safe-haven contrast",885 "description": "Documents that the damping effect is stronger for procyclical commodities such as crude oil and natural gas than for safe havens like gold."886 },887 {888 "title": "NLS speculation proxy at announcement times",889 "description": "Builds a nonlinear-least-squares speculation intensity proxy from disaggregated positioning data and interacts it with high-frequency announcement-window reactions in energy and metals futures."890 }891 ],892 "data": [893 {894 "title": "Five-minute commodity futures, 2007–2024",895 "description": "High-frequency price data for crude oil (CL), natural gas (NG), gold (GC), silver (SI), copper (HG), and palladium (PA) futures."896 },897 {898 "title": "26 macroeconomic announcement series",899 "description": "Scheduled U.S. macroeconomic releases converted into standardized surprises to measure announcement-window reactions in returns, volatility, and spreads."900 },901 {902 "title": "CFTC disaggregated Commitments of Traders",903 "description": "Weekly positioning data separating money managers from swap dealers, used to construct the NLS speculation intensity proxy."904 }905 ],906 "methodology": [907 {908 "title": "WLS-EWMA event regressions",909 "description": "Weighted least squares with exponentially weighted moving average variance estimates to measure announcement effects on high-frequency returns and volatility."910 },911 {912 "title": "GARCH volatility modelling",913 "description": "GARCH-type specifications capture conditional volatility dynamics around macroeconomic announcement releases in energy and metals futures."914 },915 {916 "title": "NLS speculation proxy from COT data",917 "description": "A nonlinear-least-squares proxy for speculation intensity built from CFTC disaggregated positions, interacted with standardized macro surprises."918 },919 {920 "title": "COVID and ZLB robustness appendices",921 "description": "Dedicated appendices test the robustness of the results across the COVID-19 period and the zero-lower-bound monetary policy regime."922 }923 ],924 "reproducibility": [925 "Full LaTeX source lives in phd_chap1_20260731/ as a frozen July 31, 2026 snapshot: a monolithic main.tex plus tables.tex (14 tables), figures.tex (6 figures), and master.bib.",926 "The folder includes COVID and ZLB robustness appendices and the compiled main.pdf alongside standalone figure files (FIG_NLS, FIG_MSCT, FIG_WT, per-commodity plots).",927 "The chapter is also integrated into the full thesis in these-ulaval/, where labels are prefixed ch1: and the bibliography is merged into a consolidated 271-key BibTeX file."928 ],929 "keywords": [930 "speculative trading",931 "energy futures",932 "macroeconomic announcements",933 "high-frequency data",934 "CFTC Commitments of Traders",935 "liquidity",936 "price discovery",937 "volatility"938 ]939 },940 {941 "slug": "chapter-2",942 "repo": "https://git.spboucher.ai/phd_thesis",943 "hero": {944 "headline": "Intraday iNAV Reveals Jump-Driven ETF Volatility Transmission",945 "subheadline": "How does volatility flow between commodity ETFs and their underlyings? The intraday indicative NAV shows transmission runs through jumps, not diffusion."946 },947 "abstract": [948 "This essay asks how volatility flows between commodity ETFs and their underlying assets, and what the intraday indicative NAV (iNAV) reveals that daily data cannot. It builds a novel minute-level iNAV dataset for four single-commodity ETFs — gold (GLD), silver (SLV), oil (USO), and natural gas (UNG) — from roughly 45 million tick observations spanning 2010 to 2023.",949 "Realized variance is decomposed into continuous and jump components and modelled with HAR-X and HAR-CJ-X regressions at 1, 5, and 30 minutes, plus a Bayesian VAR. The iNAV yields a sharper image of the ETF–underlying volatility relationship: transmission runs primarily through jumps, sampling frequency matters (1-minute estimates are up to twice as large as 30-minute ones), and precious metals show unidirectional transmission while energy is bidirectional and asymmetric."950 ],951 "findings": [952 {953 "value": "≈45M",954 "label": "Tick observations underlying the minute-level iNAV dataset, 2010–2023"955 },956 {957 "value": "4",958 "label": "Single-commodity ETFs studied (GLD, SLV, USO, UNG)"959 },960 {961 "value": "2×",962 "label": "1-minute transmission estimates up to twice as large as 30-minute ones"963 },964 {965 "value": "3",966 "label": "Sampling frequencies compared: 1, 5, and 30 minutes"967 },968 {969 "value": "13",970 "label": "Tables and 8 figures (realized volatility and IRF plots)"971 }972 ],973 "contributions": [974 {975 "title": "Novel minute-level iNAV dataset",976 "description": "Constructs an original minute-frequency indicative NAV dataset for four single-commodity ETFs from roughly 45 million tick observations covering 2010 to 2023."977 },978 {979 "title": "Sharper view of ETF–underlying dynamics",980 "description": "Shows the intraday iNAV gives a sharper image of the volatility relationship between ETFs and their underlying assets than daily data can provide."981 },982 {983 "title": "Jumps as the transmission channel",984 "description": "Decomposing realized variance shows volatility transmission runs primarily through jump components rather than the continuous diffusion component."985 },986 {987 "title": "Sampling frequency matters for inference",988 "description": "Comparing 1-, 5-, and 30-minute estimates shows 1-minute transmission coefficients can be up to twice as large as 30-minute ones."989 },990 {991 "title": "Asset-class contrast in transmission direction",992 "description": "Precious metals show unidirectional iNAV-to-ETF transmission consistent with passive arbitrage, while energy ETFs display bidirectional and asymmetric volatility flows."993 }994 ],995 "data": [996 {997 "title": "Minute-level iNAV series, 2010–2023",998 "description": "A novel intraday indicative NAV dataset for four single-commodity ETFs, built from approximately 45 million tick observations."999 },1000 {1001 "title": "Commodity ETF prices: GLD, SLV, USO, UNG",1002 "description": "High-frequency prices for the gold, silver, crude oil, and natural gas ETFs, matched to their underlying assets at 1-, 5-, and 30-minute frequencies."1003 }1004 ],1005 "methodology": [1006 {1007 "title": "HAR-X and HAR-CJ-X models",1008 "description": "Heterogeneous autoregressive realized-volatility regressions with cross-market terms, estimated at 1-, 5-, and 30-minute sampling frequencies."1009 },1010 {1011 "title": "Jump decomposition (Barndorff-Nielsen–Shephard)",1012 "description": "Realized variance is decomposed into continuous and jump components to identify which channel carries volatility between ETFs and underlyings."1013 },1014 {1015 "title": "Minnesota-prior Bayesian VAR",1016 "description": "A Bayesian VAR with Minnesota prior traces impulse responses and the direction and asymmetry of volatility transmission between markets."1017 }1018 ],1019 "reproducibility": [1020 "Full LaTeX source in phd_chap2_20260731/ is organized as main.tex plus a sections/ folder (introduction, data, methods, results, conclusion), with 13 tables, 8 figures, and master.bib.",1021 "The folder is a frozen July 31, 2026 submission snapshot for the Journal of Futures Markets and includes the compiled main.pdf.",1022 "The thesis-adapted version lives in these-ulaval/chapitre2/, with labels prefixed ch2: and references merged into the consolidated thesis bibliography."1023 ],1024 "keywords": [1025 "commodity ETFs",1026 "indicative NAV",1027 "volatility transmission",1028 "realized volatility",1029 "jumps",1030 "HAR models",1031 "Bayesian VAR",1032 "high-frequency data"1033 ]1034 },1035 {1036 "slug": "chapter-3",1037 "repo": "https://git.spboucher.ai/phd_thesis",1038 "hero": {1039 "headline": "Fed Tone Moves Prices, Novelty Moves Volatility",1040 "subheadline": "When the Fed speaks, what moves markets — what it says or how new it is? Tone drives directional returns; novelty drives volatility."1041 },1042 "abstract": [1043 "This essay asks what moves markets when the Federal Reserve speaks: the policy tone of the statement or its informational novelty. FOMC statements (217 releases, 2000–2025) are decomposed into hawkish/dovish tone and novelty using a dual-model NLP ensemble (MiniLM and BERT with TSDAE+MNRL fine-tuning and PCA-based reference selection), then linked to 1-minute futures data across seven contracts around 148 FOMC events from 2008 to 2025.",1044 "The two dimensions play distinct roles. Tone predicts directional returns: a one-standard-deviation dovish shift generates equity gains building to about +12 basis points within two hours. Novelty predicts volatility: the stance-by-novelty interaction on the VIX persists from 5 to 120 minutes (t = -5.06), and stance moves realized volatility in six of seven contracts at the 1% level, while pre-announcement placebo tests are null."1045 ],1046 "findings": [1047 {1048 "value": "+12 bps",1049 "label": "Equity gain within two hours after a 1σ dovish tone shift"1050 },1051 {1052 "value": "t = −5.06",1053 "label": "Stance × novelty interaction on VIX, persisting 5–120 minutes"1054 },1055 {1056 "value": "6 of 7",1057 "label": "Contracts where stance moves realized volatility (p < 0.01)"1058 },1059 {1060 "value": "148",1061 "label": "FOMC events with 1-minute futures data, 2008–2025"1062 },1063 {1064 "value": "217",1065 "label": "FOMC statement releases scored for tone and novelty, 2000–2025"1066 },1067 {1068 "value": "7",1069 "label": "Futures contracts studied: ES, VX, ZN, ZF, DX, CL, GC"1070 }1071 ],1072 "contributions": [1073 {1074 "title": "Separating policy tone from novelty",1075 "description": "Decomposes FOMC statements into hawkish/dovish tone and informational novelty, showing the two dimensions have distinct, complementary effects on returns and volatility."1076 },1077 {1078 "title": "Dual-model NLP ensemble for Fed text",1079 "description": "Combines MiniLM and BERT with TSDAE+MNRL fine-tuning and PCA-based reference selection to score 217 FOMC statements from 2000 to 2025."1080 },1081 {1082 "title": "Tone predicts directional returns",1083 "description": "A one-standard-deviation dovish shift generates equity gains that build to roughly +12 basis points within two hours of the announcement."1084 },1085 {1086 "title": "Novelty predicts volatility dynamics",1087 "description": "The stance-by-novelty interaction on the VIX persists from 5 to 120 minutes after the release, with a t-statistic of -5.06."1088 },1089 {1090 "title": "Broad cross-market volatility evidence",1091 "description": "Policy stance moves realized volatility in six of seven futures contracts at the 1% significance level, spanning equities, rates, currencies, and commodities."1092 },1093 {1094 "title": "Placebo-validated announcement effects",1095 "description": "Pre-announcement placebo tests are null, confirming the estimated tone and novelty effects are driven by the FOMC announcements themselves."1096 }1097 ],1098 "data": [1099 {1100 "title": "FOMC statements, 2000–2025",1101 "description": "217 statement releases scored for hawkish/dovish policy tone and informational novelty via the dual-model NLP ensemble."1102 },1103 {1104 "title": "1-minute futures data around 148 FOMC events",1105 "description": "Minute-level prices from 2008 to 2025 for seven contracts: ES, VX, ZN, ZF, DX, CL, and GC."1106 }1107 ],1108 "methodology": [1109 {1110 "title": "NLP ensemble (MiniLM + BERT)",1111 "description": "A dual-model text ensemble with TSDAE+MNRL fine-tuning and PCA-based reference selection extracts tone and novelty from FOMC statements."1112 },1113 {1114 "title": "Event regressions and minute-level panels",1115 "description": "High-frequency event-window regressions and minute-level panel specifications link tone and novelty to returns and realized volatility across contracts."1116 },1117 {1118 "title": "Jordà local projections",1119 "description": "Local projections trace the dynamic response of returns and volatility to tone and novelty shocks over the post-announcement window."1120 },1121 {1122 "title": "Placebo tests and five inference methods",1123 "description": "Pre-announcement placebo windows and five alternative inference methods validate that the estimated effects are announcement-driven and statistically robust."1124 }1125 ],1126 "reproducibility": [1127 "Full LaTeX source in PHD_chapitre3_theses_20260731/ is organized as chapitre3.tex plus sections/, tables/ (24 tables), 19 figures, a mathematical proofs appendix, and master.bib, with the compiled chapitre3.pdf included.",1128 "The folder is a frozen July 31, 2026 snapshot of the March 6, 2026 manuscript; thesis adaptations live separately in these-ulaval/chapitre3/.",1129 "In the assembled thesis, the chapter's proofs and extras appear as appendices A–B and its labels are prefixed ch3:, with references merged into the 271-key consolidated bibliography."1130 ],1131 "keywords": [1132 "FOMC announcements",1133 "monetary policy tone",1134 "informational novelty",1135 "NLP",1136 "high-frequency finance",1137 "local projections",1138 "realized volatility",1139 "event study"1140 ]1141 }1142];1143