FastMM

Documentation: https://ziy.bio/FastMM/

FastMM is a fast market-making engine in C++20.

• Backtests, replay and live trading run the same strategy code.
• Strategies are written in C++ or Python. Python quoting hooks are compiled with Numba and called on the trading thread; model code runs as ordinary Python on another thread.
• Every session is recorded, and replaying a recording sends the same orders again.
• Orders pass pre-trade risk limits before they are sent, and reaching the loss limit cancels all orders.
• The trading thread does not allocate memory or wait on network I/O; in simulation a market-data update becomes an order in about 120 nanoseconds (p50 on one core of a desktop; bench/README.md says what that measurement contains and what it leaves out).

Quick start

pip install "fastmm-engine[hot]"                         # Linux x86-64, CPython 3.9+
fastmm init my-mm && cd my-mm && python backtest.py      # a strategy, a config and a backtest

The C++ programs ship as a release tarball and as ghcr.io/ziyangliu-666/fastmm (Deploy a release). Building them needs Linux, gcc 13+ or clang 16+, CMake 3.25+, Ninja, OpenSSL 3 and zlib:

git clone https://github.com/ziyangliu-666/FastMM && cd FastMM
./scripts/bootstrap.sh                                   # checks toolchain, configures the release preset
cmake --build --preset release -j && ctest --preset release -j"$(nproc)"
./build/release/bin/fastmm-backtest --config configs/backtest-example.toml --data synthetic --out runs/first
python3 tools/report.py runs/first                       # -> runs/first/report.html, open it
./scripts/run-sim.sh --duration 30s                      # sim exchange + live engine, ends with a report

On real data, a day of Binance BTCUSDT perpetual at the venue's own fees:

python3 python/fastmm/data/__main__.py fetch --symbol BTCUSDT --date 2024-03-27
./build/release/bin/fastmm-backtest --config configs/backtest-binance.toml \
    --data binance:BTCUSDT,2024-03-27

What that result means, and what it cannot: Backtest on real BTCUSDT data.

Write a strategy

#include "fastmm/strategy.hpp"

using namespace fastmm;

struct MyParams {
  FASTMM_PARAMS(MyParams)
  FASTMM_PARAM_BPS(half_spread_bps, 0.005_bps, 0_bps, 1000_bps, "half spread around the mid, bps")
  FASTMM_PARAM(Qty, quote_qty, 0.001_qty, 0_qty, 1000_qty, "quantity per side, base units")
};

struct MyMM : StrategyBase<MyParams> {
  static constexpr std::string_view name() noexcept { return "my_mm"; }

  void on_book(auto& ctx, InstrumentId id, const auto& book) noexcept {
    if (!book.is_valid()) return ctx.pull_quotes(id);  // empty or crossed
    const Instrument& inst = ctx.instrument(id);
    const Price half = book.mid() * params().half_spread_bps;
    DesiredQuotes q;
    q.bid(inst.round_price(book.mid() - half, Side::Buy), inst.round_qty(params().quote_qty));
    q.ask(inst.round_price(book.mid() + half, Side::Sell), inst.round_qty(params().quote_qty));
    ctx.set_quotes(id, q);  // diffed against resting orders
  }
};

Next: tutorial, or a strategy in Python.

Limitations

• Exchanges: Binance Spot, Binance USDⓈ-M perpetuals, Bybit spot, Deribit, and Nasdaq ITCH market data.
• Linux on x86-64 only.
• The shipped strategies are reference implementations, not an edge: the example backtest is profitable only because it is configured with a maker rebate (Economics).
• No state survives a restart, monitoring is a status file to pull from (fastmm-top, or its Prometheus endpoint) with no alerting of its own, and venue-side cancel-on-disconnect is armed only on Deribit (Running this in production).

Documentation · Performance · Architecture · MIT license