# Stambaugh 1999 Replication Last updated: {sub-ref}`today` ```{toctree} :maxdepth: 1 :caption: Project Notes project_overview ``` ## Table of Contents ```{toctree} :maxdepth: 1 :caption: Notebooks πŸ“– Walkthrough: replicating Stambaugh (1999) ``` ```{toctree} :maxdepth: 1 :caption: Pipeline Charts πŸ“ˆ cb/charts.md ``` ```{postlist} :format: "{title}" ``` ```{toctree} :maxdepth: 1 :caption: Pipeline Dataframes πŸ“Š ``` ```{toctree} :maxdepth: 1 :caption: Appendix πŸ’‘ myst_markdown_demos.md apidocs/index ``` ## Pipeline Specs | Pipeline Name | Stambaugh 1999 Replication | |---------------------------------|--------------------------------------------------------| | Pipeline ID | [ashishkmaheshwari/stambaugh_1999_replication](./index.md) | | Maintainer | Ashish Maheshwari & Omar Anabtawi | | Contributors | Ashish Maheshwari & Omar Anabtawi | | Repository | | | Pipeline Web Page | Pipeline Web Page | | Date of Last Code Update | 2026-08-20 13:04:16 | | OS Compatibility | Windows, Linux, macOS | | Linked Dataframes | | **Build Commands:** ``` pip install -r requirements.txt doit ``` Does the dividend–price ratio predict stock returns? For decades the standard test regressed next month's excess return on this month's dividend yield and usually found a positive, "significant" slope. Stambaugh (1999, *Journal of Financial Economics* 54) showed the test is broken in a quantifiable way: the dividend yield is highly persistent and shares a price with the return, so the OLS slope is **biased upward in finite samples**. A positive slope is what you should expect even when the true slope is zero. This project rebuilds the data from CRSP, replicates the paper's Table 1, Table 2, and Figure 1 within a documented tolerance, and extends every exhibit through 2024 β€” where the problem turns out to be worse than in the paper's own sample: the gap between the naive and the finite-sample p-value has grown from roughly threefold to tenfold. **[Read the project site](https://ashishkmaheshwari.github.io/finm32900_stambaugh_final_project/)** β€” overview, walkthrough notebook, interactive playground, and full report. ## Exhibits | Exhibit | Content | Headline result | |---|---|---| | Table 1 | Finite-sample properties of the OLS slope, by subsample | Our finite-sample p-value 0.177 vs the paper's 0.17 | | Table 2 | Bayesian posteriors under four prior/likelihood specifications | All sixteen cells within ~0.03 of the paper | | Figure 1 | Ξ² vs ρ across methods and subperiods | Reproduces the paper's ρ > 1 overshoot in 1977–96 | | Updates | Tables 1 and 2 on samples through 2024 | Naive p 0.015 vs honest p 0.149 in 1997–2024 | Plus two educational products: a guided walkthrough notebook and an interactive browser playground where two sliders show the bias growing with persistence and shrinking with sample size. ## Repository structure ``` . β”œβ”€β”€ dodo.py # PyDoit build file β€” runs everything β”œβ”€β”€ chartbook.toml # site configuration β”œβ”€β”€ requirements.txt β”œβ”€β”€ .env.example # template for your .env (WRDS_USERNAME) β”œβ”€β”€ src/ β”‚ β”œβ”€β”€ settings.py # configuration: paths, dates, credentials β”‚ β”œβ”€β”€ pull_CRSP_index.py # WRDS pull: CRSP monthly market index β”‚ β”œβ”€β”€ pull_fama_french.py # WRDS pull: risk-free rate β”‚ β”œβ”€β”€ calc_predictor_data.py # DATA CLEANING ONLY -> tidy monthly panel β”‚ β”œβ”€β”€ monte_carlo.py # simulation engine (Table 1 Part A) β”‚ β”œβ”€β”€ stambaugh_bias.py # bias-corrected estimators (Figure 1) β”‚ β”œβ”€β”€ bayesian.py # conjugate posteriors (Table 2 specs A, B) β”‚ β”œβ”€β”€ mcmc.py # Metropolis-Hastings (Table 2 specs C, D) β”‚ β”œβ”€β”€ create_table_01_partC.py β”‚ β”œβ”€β”€ create_table_01.py # Table 1 assembly -> _output/*.tex β”‚ β”œβ”€β”€ create_table_02.py # Table 2 assembly -> _output/*.tex β”‚ β”œβ”€β”€ create_figure_01.py # Figure 1 -> _output/figure_01.png β”‚ β”œβ”€β”€ 01_walkthrough.ipynb.py # guided tour notebook (jupytext source) β”‚ └── test_*.py # unit tests β”œβ”€β”€ reports/report.tex # the write-up; inputs the generated exhibits β”œβ”€β”€ docs_src/ # site sources (edit these) β”œβ”€β”€ docs/ # BUILT site β€” generated, served by Pages β”œβ”€β”€ _data/ # pulled and cleaned data (git-ignored) └── _output/ # generated tables, figures (git-ignored) ``` Data cleaning lives in its own file, separate from all analysis. Raw data never enters the repository. ## Setup ```bash conda create -n stambaugh python=3.12 -y conda activate stambaugh pip install -r requirements.txt cp .env.example .env # then set WRDS_USERNAME=your_login ``` The first WRDS connection prompts for your password and offers to create a `.pgpass` file so later runs are non-interactive. `.env` is git-ignored and must never be committed. ## Running it ```bash doit ``` That pulls from WRDS, builds the tidy panel, regenerates every table and figure, compiles the report, executes the notebook, rebuilds the site, and runs the tests. PyDoit tracks dependencies, so re-running rebuilds only what changed. Individual stages: ```bash doit pull # WRDS pulls doit clean_data # tidy panel doit table_01 figure_01 # paper-sample exhibits doit table_01_updated # extended-sample exhibits doit notebook # execute the walkthrough doit compile_latex_docs # report PDF doit build_chartbook_site # the published site doit run_pytest # test suite ``` Note that `table_02` runs eight Metropolis-Hastings chains and takes several minutes. ## Testing ```bash pytest -q src/ ``` The suite is split deliberately. Simulation-based tests verify the bias mechanism itself β€” that the bias is positive when innovations are negatively correlated, vanishes when they are not, shrinks like 1/T, and matches the Kendall/Stambaugh analytical formula β€” and run anywhere, including CI without credentials. Data-dependent tests check our estimates against the paper's published values within stated tolerances, and skip with an explanatory message when the panel has not been built. ## Data sources - **CRSP Monthly Stock Market Indexes** (`crsp.msi`, WRDS) β€” value-weighted returns with (`vwretd`) and without (`vwretx`) dividends. Their difference gives the dividend series, which is how the dividend–price ratio is reconstructed without a separate dividend file. - **Fama–French monthly factors** (`ff.factors_monthly`, WRDS) β€” the one-month risk-free rate, for continuously compounded excess returns. Stambaugh uses a NYSE-only value-weighted index; our WRDS instance provides no pre-built NYSE-only monthly index carrying both return columns, so we use the CRSP total-market value-weighted index and document the choice in the report. ## A note on `docs/` The built site is committed so GitHub Pages can serve it directly without a build step. Edit `docs_src/`, never `docs/` β€” the latter is regenerated by `doit build_chartbook_site` and hand edits are lost. ## Team - Ashish Maheshwari - Omar Anabtawi FINM 32900, Full-Stack Quantitative Finance, Summer 2026.