Reinforcement Learning and Meta-Labeling¶
Under development
This lesson is part of the course scaffold and is being actively written. The learning objectives and outline below define its final scope.
Learning objectives¶
By the end of this lesson you will be able to:
- Formulate trading as a reinforcement learning problem — state, action, reward — and identify exactly where the formulation strains against market reality
- Demonstrate, with your own experiments, why naive RL fails in low signal-to-noise, non-stationary markets
- Implement meta-labeling to size and filter an existing primary signal and measure the improvement over the raw signal
- Combine an ensemble of weak signals and compare it against the best single signal under identical costs
Outline¶
- Trading as an RL problem — states, actions, rewards, and episodes
- Why naive RL fails — noise, non-stationarity, and the cost of samples
- What survives the critique — where RL-adjacent methods still contribute
- Meta-labeling — splitting direction from sizing with a secondary model
- Sizing and filtering with meta-labels — measuring the lift
- Ensembles of weak signals — combination methods and correlation discipline