A Two-Stage LLM-Driven Mean-Reversion Trading Strategy (Arena v1)

This Arena strategy is a systematic short‑term mean‑reversion framework. It uses large language models (LLMs) as constrained decision engines, not as price predictors.

All quantitative features are computed outside the model and passed in structured form. The LLM is only allowed to rank candidates and manage a portfolio within strict rules.

Core Design Principles

1) Mean reversion over prediction

Liquid equities frequently overshoot medium‑ and long‑term averages due to short‑term order flow, sentiment shocks, or volatility clustering. The strategy targets these statistical stretches and aims to capture the normalization window over ~2–10 trading days.

2) Strict separation of signals and decisions

The system precomputes and supplies the LLM with a fixed set of features, for example:

• Distance from MA50 and MA200
• Short‑term volatility and drawdown metrics
• Recent price changes and liquidity measures
• Earnings proximity and event‑risk penalties
• Compressed sentiment / headline features

The LLM is not asked to calculate indicators or invent new variables.

3) Constrained role of the LLM

Each model acts like a rules‑aware portfolio operator:

• Fixed universe (liquid equities)
• Fixed max positions
• Fixed position sizing
• Sector concentration limits
• Earnings risk penalties
• JSON‑only outputs (auditable)

Portfolio Structure, Capital Allocation & Leverage

Each LLM controls an independent portfolio under identical constraints:

• Initial capital: €10,000 (or equivalent)
• Max positions: 10 (long‑only)
• Position size: €3,000 per new position (fixed)
• Leverage: up to 3× gross exposure

Leverage increases the maximum gross exposure while keeping position sizing discipline constant.

Two-Stage Architecture

Stage 1 — Universe scan

Objective: reduce the universe to a shortlist of ~5 candidates with the strongest mean‑reversion setup.

Typical selection bias:

• Large negative deviation from MA50/MA200
• Elevated short‑term volatility (overshoot)
• High liquidity / market cap
• No imminent earnings

Stage 2 — Portfolio management

Objective: rotate capital into the best remaining reversion opportunities.

For each holding and candidate, the LLM outputs actions:

• BUY (fixed €3,000)
• HOLD
• SELL (when reversion is “done” or risk rises)

Timing

The system runs twice per day (example schedule):

• EU run: 10:30 CET
• US run: 19:00 CET

Risk Management

Risk is controlled structurally rather than via forecasting:

• Earnings/event risk explicitly penalized
• Liquidity thresholds prevent forced execution
• Sector limits reduce concentration
• Fixed sizing reduces implicit conviction bets
• Leverage capped at portfolio level

What this strategy is (and isn’t)

It is:

• systematic mean‑reversion
• long‑only, liquidity‑aware
• deterministic and auditable

It is not:

• a discretionary narrative strategy
• a price‑prediction model
• a news‑only trading system

Sources

• https://en.wikipedia.org/wiki/Mean_reversion_(mathematics)
• https://www.investopedia.com/terms/m/meanreversion.asp