tjarvis91/vfaix-vpa-options-trader

What the fuck is this model about?

tjarvis91/vfaix-vpa-options-trader, also known as Franken-B, is a 9 billion parameter vision-language model (VLM) built on Qwen3.5-9B-VL. It's designed for US equity options trading, processing chart images (1D/1H/5M) and market data to output trading decisions (BUY/SELL/HOLD/NO_TRADE) with conviction and a risk plan. The model is optimized for Volume-Price Analysis (VPA), pattern recognition, and chart-text fusion, aiming to assist in algorithmic and quantitative finance strategies.

What makes THIS different from all the other models?

This model is a specialized vision-language model specifically engineered for options trading, a niche application not commonly addressed by general-purpose LLMs. Its unique architecture involves stacking three LoRA adapters (V3.7s base → V5.0 LM LoRA → V5.8 last-3 LM LoRA) on Qwen3.5-9B-VL, a process called "composition-lock" where the order of adapters is critical for performance. Franken-B demonstrated significant improvements over its predecessor (V3.7) in simulated backtests, including a +6,813 pp increase in 2-year return and a +106.76% vision-only fusion gain, making it the best in its family for chart-fusion results. It's designed for local-first deployment, including an FP8 vLLM runtime for consumer GPUs.

Should I use this for my use case?

Good for:

• Algorithmic options trading: If you are developing automated trading strategies for US equity options, particularly those relying on Volume-Price Analysis and chart pattern recognition.
• Research in multimodal finance AI: Researchers interested in the application of vision-language models to financial markets, especially the "composition-lock" technique for LoRA merging.
• Local-first inference: Users who require a trading AI that runs locally on consumer GPUs (e.g., 5070 Ti) with sub-millisecond inference, without relying on cloud APIs.
• Backtesting and simulation: Evaluating trading strategies against historical data, as the model provides detailed output including conviction and risk plans.

Not ideal for:

• General-purpose language tasks: This model is highly specialized and not intended for conversational AI, content generation, or other broad LLM applications.
• Non-trading financial analysis: While finance-related, its focus is specifically on options trading decisions, not broader economic forecasting or financial reporting.
• Users seeking guaranteed returns: As explicitly stated, this is a research artifact, and simulated returns are not guaranteed forward returns. Options trading involves substantial risk.