my journey into algorithmic trading and machine learning
c++
Archived a month ago
S
No1
Script Kiddie!
Gemini Writeup:
__
Here is the problem I faced and the solution I built:
📉 The Problem: Visualizing the "Black Box"In algo trading and ML, the data isn't just text—it is millions of floating-point numbers (market ticks, tensor weights, probability distributions) generated at high frequency. I needed to see this data in real-time to debug my models, but the backend requirements were heavy:
• Server-Side Dependency: My trading engines and ML models relied on direct file system access and C++ DLLs that simply cannot run in a browser.• Data Volume: A single training epoch or backtest could generate gigabytes of
🛑 Why Standard Web Tech Failed (JSON)I quickly realized I couldn't use standard REST or JSON over WebSockets.
• The Serialization Bottleneck: Converting a 1GB array of market ticks into JSON (text) explodes the size to ~3GB and forces the CPU to burn cycles parsing strings back into numbers.• Latency: For real-time trading visualization, the "garbage collection pause" caused by parsing massive JSON objects was unacceptable.
⚡ The Solution: A Custom Binary BridgeTo bridge this gap, I built a custom binary RPC protocol that treats the browser like a direct extension of my server's memory.
• Zero-Copy Math: I implemented a system in
__
Here is the problem I faced and the solution I built:
📉 The Problem: Visualizing the "Black Box"In algo trading and ML, the data isn't just text—it is millions of floating-point numbers (market ticks, tensor weights, probability distributions) generated at high frequency. I needed to see this data in real-time to debug my models, but the backend requirements were heavy:
• Server-Side Dependency: My trading engines and ML models relied on direct file system access and C++ DLLs that simply cannot run in a browser.• Data Volume: A single training epoch or backtest could generate gigabytes of
Float64 arrays.🛑 Why Standard Web Tech Failed (JSON)I quickly realized I couldn't use standard REST or JSON over WebSockets.
• The Serialization Bottleneck: Converting a 1GB array of market ticks into JSON (text) explodes the size to ~3GB and forces the CPU to burn cycles parsing strings back into numbers.• Latency: For real-time trading visualization, the "garbage collection pause" caused by parsing massive JSON objects was unacceptable.
⚡ The Solution: A Custom Binary BridgeTo bridge this gap, I built a custom binary RPC protocol that treats the browser like a direct extension of my server's memory.
• Zero-Copy Math: I implemented a system in
rpc-args.ts and encoding.ts that serializes TypedArrays (like Float64Array) directly into binary. This allows my trading algorithms to dump raw numerical buffers into a WebSocket, which the browser reads immediately without parsing overhead.• Remote Execution: I built a transport layer (sender.ts) that lets the client invoke server-only functions (like executing a trade via a DLL) by sending compact binary commands.