# The Corporate Project Economy — Research Project

An end-to-end empirical research project built on the **SPID** database
(19,227 projects / 39,930 mentions extracted from 153,894 SEC filing sections,
500 firms, 2010–2026). It runs a full battery of econometric / statistical
analyses and produces a 30-page LaTeX paper with figures and tables.

## Layout
```
research/
  code/      numbered analysis scripts (01..10) + common.py
  data/      derived datasets (projects/mentions/panel/kg/embeddings .parquet)
  figures/   publication figures (.pdf + .png)
  tables/    LaTeX table fragments (\input-ed by the paper)
  results/   machine-readable results (.json) per analysis
  paper/     main.tex -> main.pdf (30 pages)
```

## Reproduce
```bash
PY=../../.venv/bin/python          # from research/code/
$PY 01_build_datasets.py           # build datasets from spid.duckdb
$PY 02_descriptives.py             # figs + summary/type/sector tables
$PY 03_count_models.py             # Poisson / NegBin FE count models
$PY 04_amount_models.py            # hedonic log-capital regressions
$PY 05_diffusion.py                # AI logit + logistic S-curve
$PY 06_did_event_study.py          # COVID-19 difference-in-differences
$PY 07_survival.py                 # Kaplan-Meier + Cox PH
$PY 08_network.py                  # KG degree dist + tech communities
$PY 09_embeddings.py               # KMeans + t-SNE on 384-d embeddings
$PY 10_ml_prediction.py            # gradient-boosted mega-project classifier
cd ../paper && latexmk -pdf main.tex
```

## Methods used
PPML (Poisson) & negative-binomial fixed-effects count models; hedonic OLS with
firm/year FE and clustered SE; logistic regression + nonlinear (logistic
S-curve) diffusion fitting; dynamic difference-in-differences event study with
parallel-trends test; Kaplan–Meier and Cox proportional-hazards survival models;
network community detection (greedy modularity) and eigenvector centrality;
KMeans clustering with silhouette selection and t-SNE; gradient-boosted trees
with cross-validated ROC-AUC and permutation importance.

## Headline findings
- Strong within-firm persistence of project activity; significant post-2020 level shift.
- Large capital premia for physical project families; AI projects are frequent but **not** large.
- AI-adoption S-curve inflects in **2023**; AI share of new projects 0.7% (2015) → 4.7% (2025).
- Precisely estimated **null** for differential pandemic-driven digitalization (parallel trends hold).
- Project lifecycles: M&A ~49 months → digital/AI ~144 months; larger projects complete slower.
- Energy-transition technologies are most central in the innovation network (modularity 0.52).
- Text-only mega-project classifier: CV AUC **0.743**.