Task
The trader wanted to stop watching charts by hand. He needed a system that finds entry points on its own using a trained model. Rigid rules built from a set of indicators did not suit him.
Solution
I trained a PyTorch neural network on historical quotes and built it into a bot that collects data, calculates features, makes predictions and generates signals. Inference is separate from the strategy, so entry and exit logic can change without retraining the model.
What was done
- Trained a PyTorch neural network on historical quotes
- Automated data collection, feature calculation and signal generation
- Separated inference from the trading strategy to change logic without retraining
- Added retraining of the model on fresh history
- Packaged it in Docker and started trading
Result
Signals are calculated on the data stream automatically, with no manual chart watching. The model is retrained on fresh history and plugged in without rewriting the bot.
Technologies
PythonPyTorchWebSocket
