fractal-defi 1.3.0

Open-source Python research library for DeFi strategies. Compose protocol-agnostic entities (lending, perps, DEX and LP) into typed strategies; backtest, simulate, track experiments.

Fractal

Fractal — open-source Python research library for DeFi strategies. Compose protocol-agnostic entities (lending, perps, DEX and LP) into typed strategies; backtest, simulate, track experiments.

Why Fractal

Most DeFi backtesters are product-shaped: pick a protocol, run a strategy, get a P&L curve. Fractal is library-shaped — small primitives with a big composition surface. You write a strategy once against the generic BasePerpEntity / BaseLendingEntity / BasePoolEntity / BaseSpotEntity contracts and swap concrete implementations (Hyperliquid, Aave, Uniswap V3, GMX, your own) without touching the strategy code.

Features

• Protocol-agnostic entities. Concrete implementations for Aave V3 lending, Hyperliquid perps, Uniswap V2/V3 LP and spot, Lido stETH, plus generic Simple* building blocks. Each entity is a typed state machine with update_state for accruals and protocol-correct action_* methods.
• Composable strategies. Register any number of entities under named slots, return ActionToTake from a single predict() hook, get a typed StrategyResult with metrics and a flat DataFrame.
• Live + synthetic data. Loaders for Binance public REST, Hyperliquid info API, Aave V3 GraphQL, GMX, TheGraph (Uniswap V2/V3, Lido), plus log-normal GBM and bootstrap simulators for offline stress tests.
• Experiment tracking. DefaultPipeline runs your strategy across a parameter grid, logs metrics + artifacts per run via MLflow, and supports sliding-window scenarios for stability analysis.
• Type-safe and dev-friendly. BaseStrategy[Params] generic + dataclass coercion, atomic per-action validation, delegate-resolved action arguments, deterministic replay.
• Battle-tested. 1100+ unit / invariant / real-data / integration tests, plus a Docker-based end-to-end MLflow harness in tests/mlflow_tests/.

Structure

fractal/
├── core/
│   ├── base/           # Entity / Strategy / Observation contracts
│   ├── entities/       # Aave, Hyperliquid, Uniswap V2/V3, stETH, simple/*
│   └── pipeline.py     # MLflow grid-search pipelines
├── loaders/            # Binance / Hyperliquid / Aave / GMX / TheGraph / sims
└── strategies/         # BasisTrading, HyperliquidBasis, TauReset
examples/               # quick_start, basis, tau_reset, holder, agentic_trader
docs/                   # Sphinx site (make html)
tests/
├── core/               # offline unit + invariant + e2e synthetic
├── loaders/            # real-API loader tests
└── mlflow_tests/       # Docker MLflow + end-to-end pipeline scripts

Install

pip install fractal-defi

Or from source (latest unreleased dev branch):

git clone https://github.com/Logarithm-Labs/fractal-defi.git
cd fractal-defi
pip install .

Requires Python 3.10–3.13.

Contributors and anyone running the test suite want the editable install with the dev extras (pulls in pytest, pylint, flake8, isort, pre-commit, sphinx). See CONTRIBUTING.md:

pip install -e ".[dev]"
pre-commit install

Quick start

A passive Aave-style lender accruing 5% APY for a year, no network or MLflow needed. After installing, write the file below as quick_start.py and run it:

from dataclasses import dataclass
from datetime import UTC, datetime, timedelta
from typing import List

from fractal.core.base import (Action, ActionToTake, BaseStrategy,
                               BaseStrategyParams, NamedEntity, Observation)
from fractal.core.entities import SimpleLendingEntity, SimpleLendingGlobalState


@dataclass
class LendingParams(BaseStrategyParams):
    INITIAL_BALANCE: float = 10_000.0
    LENDING_APY: float = 0.05


class PassiveLender(BaseStrategy[LendingParams]):
    def set_up(self) -> None:
        self.register_entity(NamedEntity("LENDING", SimpleLendingEntity()))
        self._funded = False

    def predict(self) -> List[ActionToTake]:
        if self._funded:
            return []
        self._funded = True
        return [ActionToTake("LENDING", Action(
            "deposit", {"amount_in_notional": self._params.INITIAL_BALANCE},
        ))]


hours = 365 * 24
start = datetime(2024, 1, 1, tzinfo=UTC)
observations = [
    Observation(timestamp=start + timedelta(hours=i), states={
        "LENDING": SimpleLendingGlobalState(
            collateral_price=1.0, debt_price=1.0,
            lending_rate=0.05 / hours, borrowing_rate=0.0,
        ),
    })
    for i in range(hours + 1)
]

result = PassiveLender(params=LendingParams()).run(observations)
print(result.get_default_metrics())

StrategyMetrics(accumulated_return=0.05127, apy=0.05127, sharpe=0.0, max_drawdown=0.0)

The same script lives at examples/quick_start.py in the repo. The repo also ships heavier examples covering basis trading, LP rebalancing, agentic trading and a toy hodler — see below.

Examples

Path	What it shows
examples/quick_start.py	Hello-Fractal: passive lending with hourly compounding
examples/holder/	Toy spot HODL with simple buy / sell triggers
examples/basis/	Hyperliquid basis trade — perp short hedged against spot long
examples/tau_reset/	Active Uniswap V3 LP with τ-reset rebalancing
examples/agentic_trader/	LLM-driven trading agent over historical klines

After cloning the repo and installing the package, run an example directly:

python examples/quick_start.py
python examples/basis/backtest.py
python examples/tau_reset/backtest.py

The grid-search variants (examples/<name>/grid.py) log results to MLflow. The repo ships a self-contained Docker MLflow stack — bring it up, point your shell at it, run the grid:

bash tests/mlflow_tests/scripts/start_mlflow.sh   # MLflow on :5500
export MLFLOW_URI=http://localhost:5500
python examples/basis/grid.py

Documentation

• ARCHITECTURE.md — entity-as-state-machine semantics, the InternalState / GlobalState split, delegate-resolved actions, notional accounting, pipeline internals.
• CHANGELOG.md — release notes.
• CONTRIBUTING.md — setup, branching, the testing pyramid, the pre-commit / CI pipeline. Required reading before opening a PR.
• Sphinx API reference can be built locally from the repo (make docs); a hosted version is on the project's docs site.

License

BSD 3-Clause. See LICENSE.

Built by Logarithm Labs.