Learning to allocate risk

01 / THE PRACTICAL QUESTION

A decision, before a model.

An allocation policy learns to trade across changing markets, but improved training reward may not translate into lower losses on unseen paths.

02 / DATA & COMPARISON

The idea in plain language.

PPO learns allocations from a reward that includes a tail-risk penalty. A penalty changes incentives; it does not enforce a guaranteed loss ceiling.

Inverse-volatility allocation

Assign larger weights to assets with lower recently observed volatility.

CVaR

Another name for Expected Shortfall here: average loss in the selected tail.

Data. Chronological or independently seeded training and evaluation returns.

Baseline. Equal weight and trailing-information portfolio weights.

03 / THE EXPERIMENT

What the saved experiment shows.

The saved PPO policy has higher daily 95% ES than inverse-volatility allocation in all three held-out seeds. This fixed-budget synthetic benchmark supports the simpler baseline for that metric.

Evidence record: learning-benchmarks.json#projects/13

04 / RESULTS & LIMITATIONS

Evidence with its boundaries attached.

The related lab is a cross-project demonstration. Read this investigation’s evidence and limits before transferring its conclusions.

05 / REPRODUCE

Reproduce and challenge the result.