The Extremity Premium — analysis code

Working Paper DAI-2510 | Dissensus AI

Code and data-analysis pipeline for:

The Extremity Premium: Sentiment Regimes and Adverse Selection in Cryptocurrency Markets Preprint: arXiv:2602.07018 · Status: under review at Computational Economics.

The paper documents an "extremity premium" — extreme Fear & Greed sentiment regimes exhibit higher spread-setting uncertainty than neutral periods, beyond what realized volatility predicts — and stress-tests it against volatility, momentum, functional-form, and multiple-testing controls. An agent-based model is included as an illustrative device only; because its spread–uncertainty link is coded rather than emergent it does no inferential work, and the paper's inferential weight rests entirely on the empirical analysis.

Reproduction

See REPRODUCE.md for the ordered pipeline (script → paper table/figure), the run convention, and data provenance. In short: install requirements.txt (Python 3.10–3.13; not 3.14 — it segfaults pandas 2.3.1), and run the analysis scripts from this code/ directory.

python -m venv venv && source venv/bin/activate
pip install -r requirements.txt          # pandas, numpy, scipy, statsmodels, arch, mesa, matplotlib, seaborn
python analysis/real_spread_validation.py   # builds results/real_spread_data.csv
python analysis/extremity_premium_analysis.py
# ... see REPRODUCE.md for the full ordered list

What the pipeline actually does

1. Spread estimation — daily Corwin–Schultz spreads from Binance BTC/USDT OHLCV (Abdi–Ranaldo and LOB validation as robustness).
2. Uncertainty construction — a heuristic aleatoric/epistemic decomposition over market observables (signals/); the epistemic proxy is cross-exchange dispersion, not Bayesian model uncertainty.
3. Regime / extremity-premium tests — extreme vs neutral sentiment regimes, volatility-quintile stratification, momentum control, extended-sample validation, placebo and permutation inference, ETH cross-asset replication (analysis/).
4. Agent-based model — a Mesa ABM (simulation/, agents/) used for a consistency check and an SMM moment-matching validation on a reduced-form chartist–fundamentalist model (not the full agent specification).

Data provenance (public; no API keys committed)

• Binance BTC/USDT and ETH/USDT daily OHLCV (spreads, returns, volatility).
• Alternative.me Crypto Fear & Greed Index (sentiment regimes) — a proprietary composite; raw component series are not published (see the paper's limitations). Note: sentiment is the F&G index, not Reddit/social text.
• Deribit DVOL implied-volatility index (robustness).
• Bybit / Binance L2 order-book snapshots (spread validation).

The large raw trees (data/lob/, data/data2/, data/binance/, ~16 GB) are gitignored; the committed analysis inputs live in results/ (real_spread_data.csv, full_sample_btc_data.csv).

Repository layout

code/
├── analysis/        # the paper's analysis scripts (spreads, extremity premium, robustness, SMM)
├── signals/         # uncertainty decomposer
├── simulation/      # Mesa market environment, order book, matching engine (ABM)
├── agents/          # market maker / informed / noise trader agents
├── data_ingestion/  # OHLCV / F&G / DVOL fetchers
├── config/          # configuration
├── results/         # committed analysis inputs + outputs
├── tests/           # unit / validation tests
└── REPRODUCE.md     # ordered reproduction pipeline

(feature_engineering/ and monitoring/ contain scaffolding from an earlier real-time design that the paper's offline daily-frequency pipeline does not use.)

Citation

@article{farzulla_extremity_premium,
  title  = {The Extremity Premium: Sentiment Regimes and Adverse Selection in Cryptocurrency Markets},
  author = {Farzulla, Murad},
  year   = {2026},
  note   = {Preprint arXiv:2602.07018; under review, Computational Economics}}

Authors

• Murad Farzulla -- Dissensus AI & King's College London
  • ORCID: 0009-0002-7164-8704
  • Email: murad@dissensus.ai

License

Paper content: CC-BY-4.0. Code: MIT --- see LICENSE.