ASRI: Aggregated Systemic Risk Index

Interpretable, channel-decomposed monitoring of crypto-native systemic stress

Working Paper DAI-2509 | Dissensus | Dashboard

Abstract

The Aggregated Systemic Risk Index (ASRI) is a composite measure for retrospective, interpretable monitoring of systemic stress arising from DeFi-TradFi interconnection. It aggregates four sub-indices -- Stablecoin Concentration Risk, DeFi Liquidity Risk, Contagion Risk, and Regulatory Opacity Risk -- into a daily composite, decomposable back to its channels. It is not a validated real-time early-warning system.

Against four major crypto crises (Terra/Luna, Celsius/3AC, FTX, SVB), an HAC-robust event study is inconclusive and is not counted toward detection: the apparent "3/4 significant" reading holds only under the invalid asymptotic-normal reference, and on the valid fixed-b reference only SVB nominally clears the 5% threshold (1/4) -- which a placebo/specificity test shows is not crisis-specific (HAC t = 1.72/3.49/3.28/3.86 for Terra/Celsius/FTX/SVB). Fixed-threshold operational detection (ASRI >= 50) separately identifies three of four, with Terra/Luna's pre-window peak (46.0) below the threshold. A Hidden Markov Model labels three regimes (Low/Moderate/Crisis) with >97% persistence, restating the index's own serial correlation rather than statistically distinct generating processes.

On a fair-baseline comparison (identical labels and protocol), ASRI's discrimination (AUROC 0.866) is not statistically distinguishable from its single strongest sub-index (Contagion, 0.851) or from PC1 of its sub-indices (0.858) -- with only four crisis episodes, those gaps are within overlapping bootstrap intervals. ASRI's measured advantage is confined to the Diebold-Yilmaz connectedness benchmark (0.670), which is itself constructed from ASRI's sub-indices and is outperformed by an off-the-shelf Fear & Greed index (0.789). The value of aggregation here is interpretive -- channel decomposition, a single auditable composite, regime structure -- not discriminative. ASRI proxies DeFi-TradFi transmission via TradFi-stress indicators (Treasury/VIX/yield), rather than measuring a realised exposure directly. An open-source implementation with live dashboard is provided.

Paper links:

• arXiv: https://arxiv.org/abs/2602.03874
• DOI: https://doi.org/10.48550/arXiv.2602.03874

Key Findings

Finding	Result
Event study (HAC-robust)	Inconclusive, not counted toward detection: the "3/4 significant" reading holds only under the invalid asymptotic-normal reference; on the valid fixed-b reference only SVB nominally clears (1/4), and a placebo test shows it is not crisis-specific (HAC t = 1.72/3.49/3.28/3.86 for Terra/Celsius/FTX/SVB)
Threshold detection	Fixed ASRI >= 50 identifies 3/4 at ~19-day first-crossing lead; Terra/Luna pre-window peak 46.0 < 50. Walk-forward (training-calibrated thresholds) recovers 4/4 out-of-sample, at a 47--59% false-alarm cost
Regime structure	HMM labels 3 regimes (Low/Moderate/Crisis), persistence >97% (0.997 / 0.992 / 0.980) -- interpretive structure over an autocorrelated trend
Fair-baseline discrimination	ASRI AUROC 0.866 not distinguishable from Contagion channel alone (0.851) or PC1 (0.858); aggregation adds ~+0.008 AUROC over the best single channel
Benchmark caveat	The Diebold-Yilmaz benchmark (0.670) is built from ASRI's own sub-indices (circular) and is beaten by an off-the-shelf Fear & Greed index (0.789) -- outperforming it is not evidence of value
Out-of-sample behaviour	No sustained false alarms on 2024--2025 holdout (single non-systemic Aug-2024 elevation, peak 58.8)

Full numbers: results/baseline_comparison.json, results/event_study_hac.json, results/moving_block_bootstrap_roc.json, results/regenerated_tables_jun2026.json.

Keywords

systemic risk, cryptocurrency, decentralised finance, stablecoin stability, contagion risk, DeFi-TradFi interconnection, interpretable monitoring, channel decomposition

Architecture

ASRI comprises four weighted sub-indices aggregated into a daily composite score:

Sub-Index	Weight	Coverage
Stablecoin Concentration Risk	30%	Peg deviation, dominance, reserve opacity
DeFi Liquidity Risk	25%	TVL drawdowns, protocol concentration, composability
Contagion Risk	25%	Cross-market correlation, exchange flow, cascade metrics
Regulatory Opacity Risk	20%	Classification uncertainty, enforcement patterns

Note: in discrimination terms the Contagion channel carries most of the signal (single-channel AUROC 0.851); the "stablecoin dominance" of some weight specifications is about loading, not predictive power.

Project Structure

asri/
├── src/asri/
│   ├── api/          # FastAPI endpoints
│   ├── ingestion/    # Data source connectors
│   ├── signals/      # Sub-index calculations
│   └── models/       # Database models
├── tests/            # Test suite
├── scripts/          # Analysis + verification scripts (event study, bootstrap, baselines)
├── config/           # Configuration files
└── docs/             # Documentation

API Endpoints

GET /asri/current          # Current ASRI value + sub-indices
GET /asri/timeseries       # Historical data
GET /asri/subindex/{name}  # Individual sub-index
GET /asri/regime           # Regime classification
GET /asri/validation       # Validation summary
GET /asri/methodology      # Documentation

Production Routing

• Canonical dashboard: https://asri.dissensus.ai/
• Canonical API base: https://asri.dissensus.ai/api
• Canonical docs: https://asri.dissensus.ai/docs
• Legacy API compatibility: https://api.dissensus.ai/*

Data Sources

Source	Type	Connector
CoinGecko	Prices, market data (cross-market correlations)	src/asri/ingestion/coingecko.py
DeFi Llama	TVL, stablecoin supply and peg data	src/asri/ingestion/defillama.py
FRED	Macro indicators (Treasury, VIX, yields)	src/asri/ingestion/fred.py
News (CryptoPanic / NewsAPI)	Regulatory news sentiment (VADER)	src/asri/ingestion/news.py

Getting Started

# Clone
git clone https://github.com/studiofarzulla/asri.git
cd asri

# Setup environment
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

# Configure
cp .env.example .env
# Edit .env with your API keys

# Run
uvicorn asri.api.main:app --reload

Reproducing the paper results

All reported numbers reproduce deterministically from the frozen canonical ASRI series results/data/asri_history.parquet (max 84.70 on 2022-11-08, mean 39.20, 1841 rows; sha256 f0fc1502…628a827d; Zenodo concept 10.5281/zenodo.17918238, parquet deposited from v2.0.0 onwards). This series is a released, frozen dataset of record, not a re-derived artefact: the generation pipeline's original live inputs and protocols/bridges universe were never archived, so the daily series is not bit-reproducible from the live ingestion pipeline. The analysis scripts read the frozen parquet directly.

Use Python 3.11–3.13 (3.14 segfaults pandas). See REPRODUCIBILITY.md for the exact env, the script→number map (verified actuals), and provenance; see DATA_PROVENANCE.md for the hash freeze and the documented generation-code bugs.

# verify the frozen series first
sha256sum results/data/asri_history.parquet   # f0fc1502…628a827d

python scripts/event_study_hac.py            # HAC event study: t 1.72/3.49/3.28/3.86; inconclusive (fixed-b: only SVB clears, 1/4)
python scripts/moving_block_bootstrap_roc.py # ASRI 0.866 vs Diebold-Yilmaz 0.670, block-bootstrap CIs
python scripts/baseline_comparison.py        # AUROC ASRI 0.866 / Contagion 0.851 / PC1 0.858 / F&G 0.789
python scripts/generate_detection_table.py   # peaks (FTX 84.7), τ=50 3/4, walk-forward OOS 4/4
python scripts/extract_hmm_diagnostics.py    # 3 regimes, persistence 0.997/0.992/0.980

The generation pipeline (src/asri/backtest/, scripts/generate_asri_series.py) is provided and corrected (D2 coin-ids, real peg, D5 rolling-365 TVL, frozen-universe-snapshot determinism). With a frozen snapshot (scripts/dump_universe_snapshot.py) it runs deterministically, but it produces a code-consistent series — it does not, and is hard-refused from, overwriting the frozen canonical parquet.

Citation

@article{farzulla2025asri,
  author    = {Farzulla, Murad and Maksakov, Andrew},
  title     = {ASRI: An Aggregated Systemic Risk Index for Cryptocurrency Markets},
  year      = {2025},
  eprint    = {2602.03874},
  archivePrefix = {arXiv},
  primaryClass  = {q-fin.RM},
  doi       = {10.5281/zenodo.17918238}
}

Authors

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

License

Paper content: CC-BY-4.0 | Code: MIT