hftbacktest-cpp

Project lineage:hftbacktest-cpp is a C++20 port of HftBacktest, originally created by nkaz001 and developed with its contributors. This independent port is not affiliated with or endorsed by the upstream project.

A compact, installable C++20 event-driven backtesting core for limit-order-book research. It models local and exchange clocks, order entry and response latency, queue position, exchange fills, fees, inventory, and sharded NumPy event input.

This is an early 0.1.x port. Its architecture and algorithms originate in the upstream HftBacktest project; this repository supplies an independent C++ implementation and packaging.

Quick start

Requirements: CMake 3.20+, a C++20 compiler, zlib, and Git. GCC and Clang are tested on Linux; AppleClang is tested on macOS.

git clone https://github.com/shenron0101/hftbacktest-cpp.git
cmake -S hftbacktest-cpp -B hftbacktest-cpp/build -DCMAKE_BUILD_TYPE=Release
cmake --build hftbacktest-cpp/build --parallel
ctest --test-dir hftbacktest-cpp/build --output-on-failure
./hftbacktest-cpp/build/hbt_market_making

Deterministic example output:

book ts_ns=1050000 bid=99.900 ask=100.100 mid=100.000
order id=1 side=buy price=99.900 ack_ts_ns=1200000
fill id=1 side=buy price=99.900 qty=1.000 maker=true
order id=2 side=sell price=100.100 ack_ts_ns=2350000
fill id=2 side=sell price=100.100 qty=1.000 maker=true
summary fills=2 position=0.000 balance=0.200 fee=0.020 volume=2.000 final_ts_ns=4050000
latency entry_ns=100000 response_ns=50000

What is included

Capability	Status	Notes
L2 market depth	Included	Hash-map and bounded ROI-vector books
Event-driven replay	Included	Separate local/exchange timestamps and multi-asset scheduling
Order lifecycle	Included	Submit, modify, cancel, acknowledgements, fills
Queue models	Included	Risk-adverse and probabilistic queue position
Exchange models	Included	No-partial-fill and partial-fill variants
Latency and fees	Included	Constant/interpolated latency; linear/inverse assets; fee models
NumPy input	Included	.npy and stored/DEFLATE .npz readers via zlib
Package consumption	Included	find_package(hbt CONFIG REQUIRED), hbt::hbt, hbt::io
Live trading and venue connectors	Excluded	Use the original project where those capabilities are required
L3/order-by-order book	Not yet ported	Planned after API stabilization

API sketch

#include<hbt/hbt.hpp>
hbt::HashMapMarketDepth depth{0.01, 0.001};
depth.update_bid_depth(99.99, 2.0, 1'000);
depth.update_ask_depth(100.01, 3.0, 1'000);
constdouble bid = depth.best_bid();
constdouble ask = depth.best_ask();

For an installed package:

find_package(hbtCONFIGREQUIRED)
target_link_libraries(my_backtestPRIVATEhbt::hbthbt::io)

Architecture

flowchart LR
D[NPY / NPZ event shards] --> R[EventReader]
R --> S[EventSet scheduler]
S --> L[Local processor]
S --> X[Exchange processor]
L <-->|latency-aware order buses| X
L --> B[Strategy / Backtest API]
X --> Q[Queue + fill model]
L --> A[Position, balance, fees]
X --> A

Loading

The hot-path model implementations are templates. A small processor interface provides type erasure at the multi-asset scheduler boundary.

Build options

Option	Default	Purpose
HBT_BUILD_TESTS	top-level ON	Build 54 retained core tests plus port regressions
HBT_BUILD_EXAMPLES	top-level ON	Build and test the deterministic synthetic example
HBT_BUILD_BENCHMARKS	OFF	Build the fixed-workload replay benchmark
HBT_NATIVE_ARCH	OFF	Enable -march=native; unsuitable for portable release binaries
HBT_WARNINGS_AS_ERRORS	OFF	Promote project warnings to errors

Install and consume:

cmake -S . -B build -DHBT_BUILD_TESTS=OFF -DHBT_BUILD_EXAMPLES=OFF
cmake --build build --parallel
cmake --install build --prefix "$PWD/stage"
cmake -S consumer -B consumer-build -DCMAKE_PREFIX_PATH="$PWD/stage"
cmake --build consumer-build --parallel

Benchmark

The benchmark performs exactly 5,000,000 deterministic depth updates and reports a checksum. Throughput is machine-dependent; compare builds on the same idle host.

cmake -S . -B build-bench -DCMAKE_BUILD_TYPE=Release \
-DHBT_BUILD_TESTS=OFF -DHBT_BUILD_EXAMPLES=OFF -DHBT_BUILD_BENCHMARKS=ON
cmake --build build-bench --parallel
./build-bench/hbt_replay_benchmark

Workload	Toolchain / host	Native optimization	Median of 5 runs
5M L2 depth updates	GCC 13.3 / Ryzen 9 5900X / Linux x86-64	Off	149.0M events/s

Measured 2026-06-19 with the commands above. This is a transparent baseline, not a cross-project claim. CI records its own raw output as an artifact.

Limitations

• This port has not reached API stability; minor releases may change interfaces.
• The current NPZ reader intentionally supports classic ZIP32 archives only, with stored or DEFLATE members.
• Replay correctness depends on normalized, chronologically valid event data and realistic latency/queue assumptions.
• No venue connectors, bundled strategies, alpha models, private data, or generated coefficient artifacts are included.
• This software is research infrastructure, not investment advice or a guarantee of live execution behavior.

Roadmap

• Differential fixtures against the pinned upstream implementation.
• Stable data-schema documentation and stronger malformed-input diagnostics.
• L3 market depth and additional exchange semantics where upstream parity is testable.
• Profiling-led performance work without weakening deterministic behavior.
• API stabilization toward 1.0 after production feedback.

Original HftBacktest Project

The original implementation is HftBacktest, created by nkaz001. Its architecture, event model, queue/exchange abstractions, and algorithms are the basis of this port. Read the upstream documentation and recognize the upstream contributors.

The pinned reference and adapted component inventory are recorded in NOTICE. Upstream remains the authoritative project for its own roadmap, Python/Rust APIs, live trading, examples, and venue integrations.

Contributing and security

See CONTRIBUTING.md for build and review expectations, SECURITY.md for private vulnerability reporting, and CODE_OF_CONDUCT.md for community standards.

License

MIT. See LICENSE and NOTICE. The upstream and port copyright notices are both retained.