HFT Trading Simulator
Educational high-frequency trading simulator — real low-latency engineering, zero real money.
- Period
- Jun–Sep 2026
- Role
- Design & implementation
- Commits
- 1,289
- Key metric
- 75+ quant models · 191+ live panels · FIX 4.4
- Links
- Repo ↗
Problem
HFT systems are where software engineering meets physics: nanoseconds matter, allocation is a bug, and a cache miss is a design flaw. I wanted to learn that discipline for real — not by reading about it, but by building a simulator that behaves like a production trading stack, without risking a cent.
Constraints
Educational by definition: zero real money, zero real risk. But the engineering had to be honest — lock-free where a real desk would be lock-free, IPC where real systems use shared memory, and models that a quant would actually recognize.
Architecture
A C++20 signal engine sits at the core, exchanging data with a Python research layer over shared-memory IPC — the same pattern real trading firms use between strategy and execution. A Rust order executor handles the hot path, CUDA accelerates batch computation, and 191+ React dashboard panels render the whole system live.
Key decisions
- Lock-free SPSC ring buffers with cache-line alignment and branchless programming — no mutexes on the hot path.
- Signal Engine V3: HMM regime detection with online Baum-Welch adaptation and Viterbi decoding — sub-millisecond regime switching.
- 75+ quantitative models: GARCH, Kalman filters, Almgren–Chriss optimal execution, and more — each validated, not stubbed.
- ONNX ML inference in C++ — trained models run inside the engine, not in a Python sidecar.
- FIX 4.4 protocol support — the wire format real markets speak.
- AI-assisted hot-path tuning — model-suggested memory-layout and cache-line tweaks on hand-written C++ were benchmarked and adopted where they measured faster; the rest were rejected.
What it proves
Multi-language systems engineering (C++20, Rust, Python, CUDA, TypeScript) with a full CI/CD pipeline across all of them and Kubernetes/Helm deployment. The hard parts — shared memory, lock-free queues, numeric kernels — are real, not simulated.