Skip to content

AI Model Framework

AIM’s trading system predominantly leverages Large Language Models (LLMs) augmented by a continuous feedback mechanism:

LLM Ensemble Approach

  • Multiple LLMs: AIM blends insights from top-tier models such as GPT, Claude, and others, pooling their strengths to form comprehensive market perspectives.
  • Dynamic Model Selection: Depending on realtime market conditions and data availability, AIM can weight certain models more heavily to optimize decision-making.

Investment Opinion Aggregation

  • Diverse Model Outputs: Each LLM provides an independent set of predictions and trading signals based on its own training and prompts.
  • Consensus Mechanism: AIM integrates these signals via a consensus algorithm. For instance, if a majority of models indicate a bullish trend, the system adjusts its stance accordingly.

Once a unified investment opinion is formed, AIM distributes the decision across distinct AI-managed portfolios, each aligned with a different time horizon. This enables more granular risk control and strategic diversification:

  • 📘 LONG TERM (1+ year)
    Focused on macroeconomic indicators and multi-year market cycles. These positions are low-frequency, high-conviction bets.

  • 📙 MID TERM (1–4 months)
    Targets swing trades using technical analysis, trend confirmations, and mid-range sentiment patterns.

  • 📕 SHORT TERM (days to weeks)
    Executes high-frequency trades based on volatility, real-time sentiment analysis, and on-chain signals such as whale alerts and fund movements.

Each portfolio is managed by a dedicated AI agent fine-tuned for its strategy type, ensuring optimized execution under different market conditions.

Self-Adaptive Learning with Chart SEAL

To enhance its adaptability and performance, AIM incorporates techniques inspired by the SEAL (Self-Adapting Language Models) framework. Rather than relying solely on human-curated finetuning data, AIM's architecture enables its LLM agents to evolve autonomously by learning from their own successes and failures in real-world trading.

This self-adaptive mechanism allows AIM to:

  • Adjust to new market regimes and trends in near real-time.
  • Refine its strategy logic based on observed outcomes—without manual retraining.
  • Develop more resilient and flexible decision-making pathways over time.

By integrating this continuous, self-improving structure, AIM transitions from a static prompt-based system to a dynamic intelligence engine—capable of improving through experience.

Continuous Self-Feedback Loop

  • Performance Tracking: The system monitors how each LLM-driven insight performs under different market conditions.
  • Iterative Fine-Tuning: Models are retrained or reweighted based on historical performance, ensuring adaptability and long-term improvement.

Execution Layer

  • Order Optimization: Trades are intelligently split and scheduled to minimize slippage and market impact.
  • High Frequency Execution: AIM’s execution interface operates at sub-second speeds, capitalizing on opportunities as they emerge.