Skip to main content

Architecture

The backend of TradeSim is built to stream real-time market data, simulate trades, and provide machine learning predictions without maintaining persistent state internally. It acts as a pipeline for real-time data flow from exchange → processing → frontend.

Core Technologies​

  • Python – Core backend and data handling logic
  • FastAPI – Provides REST and WebSocket endpoints
  • scikit-learn – Used for training and serving the ML predictor
  • Uvicorn – ASGI server for async support
  • orjson – Fast JSON serialization/deserialization
  • OKX API – External data source for orderbook snapshots

Architectural Overview​

1. Real-Time Data Routing​

  • The backend connects to the OKX WebSocket API to receive live orderbook data.
  • This data is then:
    • Published to a WebSocket endpoint: /ws/okx (for frontend visualization)
    • Forwarded to the prediction module via internal WebSocket or async message queue

2. Prediction Pipeline​

  • The prediction service listens for incoming orderbook snapshots via WebSocket.
  • It runs a scikit-learn ML model on each snapshot to forecast the midpoint 60 seconds into pricethe future.
  • Predictions are streamed to the frontend via the /ws/prediction WebSocket endpoint.

3. Trade Simulation Engine​

  • Stateless simulation using the current orderbook data supplied with each request.
  • Computes:
    • Simulated market order fill
    • Execution metrics like slippage, latency, etc.
  • Available at the REST endpoint: POST /simulate/trade

4. High-Performance JSON Handling​

To ensure minimal latency in message encoding/decoding:

  • orjson is used for fast serialization and deserialization of JSON data in both WebSocket and REST pipelines.
  • This significantly reduces the overhead of handling high-frequency market data and responses.

5. Data Flow Overview​

              [OKX WebSocket API]
                       ↓
          [Backend Data Relay Service]
               ↙              ↘
     [/ws/okx]          [Prediction Module]
                                ↓
                        [/ws/prediction]

Design Characteristics​

  • Stateless: No long-term state is stored on the backend — all logic is driven by live or request-time data.
  • Modular: Prediction and simulation are separated logically and can scale independently.
  • High-Performance: Fully async stack, fast JSON parsing (orjson), and WebSocket-first design allow real-time responsiveness.