| # Chapter 4 Baseline Codebase |
|
|
| This repository is a cleaned baseline-only snapshot for the ten runs used in |
| Chapter 4 of the thesis. |
|
|
| It intentionally excludes the later improvement experiments: discrete target |
| states, Top-K shortlist, frequency experiments, simulated improvement results, |
| and majority voting. |
|
|
| ## Contents |
|
|
| - Baseline source code exported from commit |
| `3c05c1afb50afa4734296fce18cc77c776f79c1d`. |
| - The ten historical Chapter 4 W&B ids and setup metadata in |
| `scripts/baseline_runs.json`. |
| - A single-run entrypoint: `run.sh`. |
| - A ten-run launcher with the same baseline setup as the selected original |
| runs: `launch.sh`. |
| - Local reusable market/news input data packaged as |
| `assets/ch4_baseline_local_data.zip`. |
|
|
| The raw local data directories are ignored by git after extraction: |
|
|
| ```text |
| stockbench/storage/parquet/ |
| stockbench/storage/cache/news/ |
| stockbench/storage/cache/news_by_day/ |
| stockbench/storage/cache/financials/ |
| stockbench/storage/cache/corporate_actions/ |
| stockbench/storage/cache/stock_indicators/ |
| ``` |
|
|
| Old LLM response caches, generated backtest reports, and local W&B outputs are |
| not included. |
|
|
| ## Setup |
|
|
| Clone the Hugging Face dataset repository: |
|
|
| ```bash |
| git clone https://huggingface.co/datasets/stock-agent/chapter4-baseline-clean |
| cd chapter4-baseline-clean |
| ``` |
|
|
| Restore local price/news data: |
|
|
| ```bash |
| unzip -q assets/ch4_baseline_local_data.zip -d . |
| ``` |
|
|
| Data archive checksum: |
|
|
| ```text |
| sha256 3fb0ccdf05aacfeef15c032aa62b956aa6a22a247a0b56145e54b30ea7ce6dbf |
| ``` |
|
|
| Create the conda environment: |
|
|
| ```bash |
| conda env create -f environment.yml |
| conda activate stockagent |
| ``` |
|
|
| Configure the LLM API key. The original runs used the `qingyuntop` |
| OpenAI-compatible profile. No API key is included in this repository. |
|
|
| ```bash |
| export QINGYUNTOP_API_KEY="your key" |
| ``` |
|
|
| Optional W&B configuration: |
|
|
| ```bash |
| export WANDB_MODE=offline # default |
| # export WANDB_MODE=online # upload new runs to W&B |
| ``` |
|
|
| ## Run |
|
|
| Run one baseline: |
|
|
| ```bash |
| MODEL=deepseek-v3.1 RUN_PREFIX=DEEPSEEK_V3_1 ./run.sh |
| ``` |
|
|
| Run the ten Chapter 4 baseline setups: |
|
|
| ```bash |
| ./launch.sh |
| ``` |
|
|
| By default, `run.sh` uses: |
|
|
| ```text |
| START=2025-03-01 |
| END=2026-02-28 |
| PROFILE=qingyuntop |
| NEWS_ENABLED=true |
| DATA_MODE=offline_only |
| CACHE_MODE=llm_write_only |
| WANDB_MODE=offline |
| ``` |
|
|
| ## Reproduction Notes |
|
|
| The manifest records the original W&B ids and original code commit for each |
| selected run. `launch.sh` does not force those old W&B ids; it runs the same |
| baseline setup and lets the current run generate fresh timestamped ids. |
|
|
| The packaged local input data contains: |
|
|
| ```text |
| parquet: 13440 files |
| cache/news: 5585 files |
| cache/news_by_day: 6010 files |
| cache/financials: 63 files |
| cache/corporate_actions: 42 files |
| cache/stock_indicators: 6783 files |
| ``` |
|
|
