Research direction: HPO parameter maps and adaptive parameter switching (GPU and RL notes)

[luisleo526]

Status: research direction, not planned work. Nothing here is measured yet, and there are no performance claims.

Idea

For a given strategy, use hyperparameter optimisation (HPO) to map how parameters behave under a chosen objective, then switch between a few robust parameter sets during forward execution.

1. Walk-forward HPO maps. Run HPO per time window, not only over the whole history. Use the parameter heatmaps to find broad plateaus of good results rather than single sharp peaks.
2. A small candidate set. Pick 3-5 robust parameter sets from the plateau centres. Fewer choices means less overfitting.
3. Adaptive switching in forward execution. Choose among the candidates from market-state features (volatility, trend strength). Start with a simple regime classifier or a contextual bandit; consider full reinforcement learning only if the simple version is not enough.
4. Validation. Out of sample, compare against the strategy with its single best fixed parameter set. The approach is only worth keeping if it wins there.

GPU (CUDA) notes

• A single backtest does not parallelise well on a GPU: each bar depends on the previous bar's position and orders.
• What fits a GPU: computing indicators for many parameter sets at once, and screening large parameter sweeps of a simplified strategy model with all parameter states stepping forward together. Exact results still come from the CPU engine.
• Training a reinforcement-learning or switching model on a GPU is natural; forward execution only runs inference, which a CPU handles.

Boundary

The engine stays computation only: the host pushes data and the engine computes results. HPO, the switching model and any GPU work live on the host side as consumers of the engine's results. No engine changes are part of this idea.

Next step, if pursued

A small measurement: GPU screening speed for a simple strategy over roughly 10,000 parameter sets, and an out-of-sample comparison of the switching approach against the best fixed parameter set.