DeFi risk analysis: a static stop-loss stress test on one day of hourly ETH data

Live: summer-fi-risk-analysis.vercel.app

This is an independent portfolio case study. A small Python model replays 24 hourly ETH price and gas observations for 5 August 2024 against static stop-loss buffers. It uses illustrative Aave V3 and Morpho parameters. A Next.js page lets a reader change the market, entry LTV and buffer and see the outcome. The repository is not affiliated with Summer.fi, Aave or Morpho.

Results

All numbers come from data/aug_5_crash_data.csv and reports/sweep_results.csv.

• Scenario. ETH goes from 2,910 at 00:00 to a 2,120 low at 06:00 and closes at 2,450 (-27.1% to the low, -15.8% at the close). Peak gas is 681.4 Gwei at 08:00.
• Sweep. 50 static runs: 2 protocols x 5 entry LTVs (0.60 to 0.80) x 5 buffers (2% to 12%).
• Aave V3 parameters. 21 runs end in a stop-loss exit and 4 in liquidation.
• Morpho parameters. 12 runs end in a stop-loss exit, 6 in liquidation and 7 stay active.
• Reading the result. A wider buffer turns a liquidation into a stop-loss exit at the same entry LTV. The sweep shows this only for this one day of data.

How it works

scripts/risk_model.py holds a DeFiRiskModel. At each hourly row it computes LTV as debt / (collateral x price). It checks the liquidation threshold first, then the stop-loss trigger (threshold minus buffer). Gas does not change LTV. It adds slippage (0.1% + gas/100 x 0.5%, capped at 5%) and a fixed 300,000-gas cost to the modeled exit value. Only a static strategy exists. web-showcase/src/app/riskModel.ts mirrors the model for the interactive page, and tests/test_ts_parity.py checks that both give the same status, trigger hour and exit values. notebooks/aug5_analysis.ipynb plots the data and the sweep, with its executed output committed.

Run it

Needs Python 3.14 (the version tested; pandas==3.1.0 needs 3.11 or newer) and Node 22.18 or newer (the parity test runs the TypeScript model with Node's built-in type stripping). No network or keys are needed after install.

python3 -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
python -m pytest tests                      # 13 tests passed on 2026-10-24
python scripts/risk_model.py --combined-sweep   # rewrites reports/sweep_results.csv (identical to the committed file)
python scripts/risk_model.py --protocol AaveV3 --ltv 0.70 --buffer 0.05
jupyter nbconvert --to notebook --execute notebooks/aug5_analysis.ipynb --stdout > /dev/null

cd web-showcase
npm ci
npm run lint && npm run build
npm run dev

Repository layout

• data/: the 24-row hourly price and gas CSV.
• scripts/: the Python risk model and command-line entry point, plus the Node helper the parity test calls.
• tests/: 13 pytest tests: the Python model, and its agreement with the TypeScript model.
• notebooks/: plots of the data and the sweep.
• reports/: the sweep CSV and a written backtest report.
• web-showcase/: the Next.js case-study page and docs viewer.

Documentation

• reports/backtest_report_aug5.md: assumptions, sweep counts and limits.
• CONSULTING_REPORT.md: the case-study write-up and recommendation.
• PROJECT_OVERSIGHT.md: scope notes for the workstream.

Sources

The CSV is described as a "supplied scenario". Its source is not recorded in this repository, and I have not checked it against on-chain or exchange data. Treat the prices and gas values as illustrative inputs, not as a market record.

Scope and limits

• Illustrative heuristic model. It is not a forecast, a valuation, or a liquidation simulator for live markets.
• One day, 24 hourly rows, one collateral asset. Hourly data cannot show whether a stop-loss would have executed before a breach within the hour, so liquidation is checked first.
• Protocol thresholds (0.825 for Aave V3, 0.945 for Morpho) and penalties (5% and 1%) are fixed parameters, not read from the live protocols. Check them before relying on any result.
• The web model is a separate TypeScript port. tests/test_ts_parity.py compares it with the Python model on the 50-run sweep grid and four off-grid inputs. It does not prove the two agree on every input.
• The notebook's committed output (three charts) was executed on 2026-10-24 with the pinned requirements. It is not re-run in CI.
• Not verified: the data source, the parameter values against current protocol configuration, and the pages at /docs.

License

MIT, see LICENSE. It covers this project's code, not the supplied scenario data, whose source is uncited.

Part of Rahil Bhavan's projects.