SYDE 532 Final Project: Prediction Markets as Complex Adaptive Systems

Team: Het Patel, Aayush Soni, Mehak Dhaliwal, Aarya Penneru

Quickstart

python -m venv venv && source venv/bin/activate 
pip install -r requirements.txt

1. Discover and select triad markets

Run notebooks/01_market_selection.ipynb to browse categories and pick the base / similar / dissimilar triad. Selected IDs live in src/config.py under TRIADS / ALL_MARKETS.

2. Collect data

python -m src.collect.run_pipeline          # full run
python -m src.collect.run_pipeline --force  # re-download everything

3. Transform

python -m src.transform

This builds per-market feature Parquet files (logit prices, returns, activity, order-book fields where available).

4. Stage 1 – Triad analysis

Open notebooks/03_triad_analysis.ipynb for return distributions, volatility clustering, reversal analysis, liquidity regressions, and cross-triad comparison.

5. Politics panel and RQ3 (market conditions)

Resolved US political markets are listed in src/config.py as POLITICS_MARKETS. The default collector (python -m src.collect.run_pipeline) and python -m src.transform iterate ALL_MARKETS (triad markets only). To rebuild politics features from APIs, collect hourly prices, activity, and order-book inputs for each politics slug (same artifacts as the triad pipeline), run build_features from src/transform.py per market, then concatenate the per-market features_<slug>.parquet files into data/processed/features_politics.parquet.

Run notebooks/04_rq3_market_conditions.ipynb for how signatures vary with activity proxies and time-to-resolution. Figures are written under plots/rq3_*.png.

6. RQ4 – Agent-based model

src/abm/ implements an hourly logit-space ABM (informed, noise, herding; latent anchor; reduced-form impact; optional endogenous participation and state-dependent depth). notebooks/05_rq4_abm.ipynb loads empirical targets from features_politics.parquet, runs calibration, compares simulated panels to the empirical pipeline, and saves data/processed/sim_stage1.parquet, data/processed/sim_stage2.parquet, and data/processed/abm_calibration.json. Figures: plots/rq4_*.png.

Project layout

src/
  config.py              – API URLs, triad IDs, POLITICS_MARKETS, settings
  client.py              – rate-limited HTTP helpers
  collect/
    discovery.py         – market discovery helpers
    prices.py            – historical price fetcher
    orderbook.py         – order-book snapshot collector
    trades.py            – paginated trade-history fetcher
    run_pipeline.py      – orchestrator script
  transform.py           – logit transform, returns, feature merge
  analyze/
    return_distribution.py
    volatility.py
    reversal.py
    liquidity_regression.py
    cross_triad.py
  abm/
    model.py             – simulator + feature-schema export
    calibration.py       – targets, distance, random-search calibration
notebooks/
  01_market_selection.ipynb
  02_eda.ipynb
  03_triad_analysis.ipynb
  04_rq3_market_conditions.ipynb
  05_rq4_abm.ipynb
data/
  raw/                   – archived JSON responses
  processed/             – Parquet features, trades, synthetic ABM panels
plots/                   – triad and RQ3/RQ4 figures