MetaPS: Adaptive Strategy Selection for Financial Markets Using Simulated Supervision

Background – In financial markets no single strategy consistently wins; momentum, mean‑reversion, risk‑control, and event‑driven rules succeed or fail depending on market conditions, driven by heavy‑tailed returns, volatility clustering, and regime dynamics.

Problem definition – At each time step t the market provides observation o_t, recent history h_t, and action space A. MetaPS assumes a library S={s_1,…,s_N} where each s_i is an executable module mapping the current market state to a structured action (direction, asset, confidence, risk, size). A router ranks a candidate subset C_t and predicts which program to invoke, producing the final action a_t.

Method

Programmatic strategy library – The library defines executable primitives for multi‑stock trading (trend‑following, reversal, news‑driven, mean‑reversion decay, risk reset, macro‑rotation, volatility breakout) and for a controlled commodity‑exchange sandbox (producer, processor, contractor, merchant, hoarder, balancer). Each program follows a shared interface and returns structured fields; a deterministic execution layer enforces feasibility, size, cost, and domain constraints.

Simulation‑guided supervision – Because state‑strategy labels are not directly observable, MetaPS runs a market simulator or historical back‑tester M to generate counterfactual rollouts. For a given state o_t, M executes each candidate s_i over a horizon H with discount γ, yielding rewards r_{t+τ}^i. These rollouts define training views V={V1,V2,V3} where V1 favors short‑term winners, V2 medium‑term advantage, and V3 balances return, risk, and behavior regularization. Scores q_i^v(o_t) are derived from the rollouts, and the target policy for view v is constructed accordingly.

MetaPS learning and inference – The simulated labels are turned into instruction‑following examples. For state t and view v, the model input is (o_t, h_t, C_t) and the target text includes the chosen strategy identifier, structured execution fields d_{t,v}, and a concise rationale m_{t,v}. Training uses a weighted language‑model loss. During inference, no simulator is available; the router receives only observable inputs, predicts a strategy identifier, and calls the selected executable program.

Experiments

Setup – Two environments with identical high‑level strategy‑selection interfaces: a multi‑stock benchmark (training 2022‑2024, testing 2025, $1M initial capital, with transaction costs, slippage, and position limits) and a controlled commodity sandbox (cash, inventory, price, production capacity, contracts, and institutional notices).

Main stock results – MetaPS‑9B achieved the highest stock return of 50.29 % in 2025, while MetaPS‑4B attained the best terminal equity of 31 514 in the sandbox. Both models outperformed all baselines, demonstrating that supervised meta‑strategy learning improves strategy switching.

Extension to sandbox – In the sandbox, MetaPS‑4B raised terminal equity from 23 879 to 31 514 and return from 175.92 % to 264.14 %, increasing win‑rate from 41.0 % to 59.0 %; similar trends were observed for the 9B model.

Strategy behavior analysis – Under the 2025 regime, MetaPS‑9B‑V3 prioritized momentum‑following while retaining news‑impulse, mean‑reversion, risk‑reset, and return‑drift as secondary programs, matching the expected effect of simulation‑guided supervision.

Scaling and target analysis – The optimal training target varies with model size: V3 is best for 0.8B and 9B models, V2 for 2B, and a serial V1→V2→V3 curriculum for 4B. Across all Qwen scales, the best MetaPS variant outperforms the corresponding base model.

Rolling‑window validation – Using the 4B router, different targets performed variably across training windows and test periods; V3’s value depended on market regime and its position in the curriculum.

The results confirm that converting simulated rollouts into supervised signals enables large language models to learn effective strategy‑level routing, yielding consistent gains over fixed‑strategy baselines and direct decision agents in both realistic stock markets and controlled economic sandboxes.