Using an LLM as a Guardrailed AAPL Risk Manager for Position Sizing

Summary

This strategy uses a large language model to set long-only exposure for AAPL according to market states, rather than asking it to predict price direction. Historical price features are discretized into readable states, and monthly statistics for each state are passed to the model to produce an exposure policy. Position sizes are mapped to partial or full investment and then adjusted through volatility targeting. Volatility and drawdown stops add hard limits, with a re-entry mechanism intended to avoid remaining sidelined indefinitely. The policy is refreshed in a monthly walk-forward process and evaluated out of sample.

The article reports a lower maximum drawdown than buy-and-hold while maintaining participation in the stock’s gains. It presents the approach as a framework, not a proven source of alpha: the state space is coarse, the prompt and guardrails remain tunable, and the result is specific to AAPL and the tested period. The evidence does not establish that the LLM generalizes to other assets or future markets, and no claim of live performance follows from the backtest.

Key ideas

• The LLM assigns exposure by market state instead of forecasting whether the next return will be positive or negative.
• Monthly walk-forward updates use historical state statistics to create a policy table.
• Volatility targeting translates policy categories into position sizes, while stop rules and forced re-entry constrain risk.
• The reported drawdown improvement is a backtest result for AAPL and does not demonstrate generalization or live profitability.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.