chapter twelve

12 Augmenting investment decisions: LLM-guided portfolio allocation

 

This chapter covers

  • Integrating quantitative momentum signals
  • Designing system prompts that provide bounded discretion within explicit constraints
  • Building a historical simulation that respects information timing
  • Understanding the gap between simulation and live trading
  • Positioning human judgment as the final decision layer

In Chapter 10, we built a quantitative engine that processes the market's numerical signals—price movements, momentum, and cross-asset correlations. In Chapter 11, we constructed a qualitative engine that extracts signals from text—news sentiment, policy tone, and emerging risks. Each engine captures different information. The quantitative engine detects statistical patterns but cannot explain the drivers behind price movements. The qualitative engine interprets narratives but lacks systematic numerical rigor.

This chapter combines these approaches into a decision-support workflow—but not by simply merging two signal streams. Instead, we elevate the LLM's role. In Chapter 11, the LLM acted as a signal extractor, reading text and outputting scores. Here, the LLM becomes a decision advisor: an agent that receives structured inputs, reasons within explicit constraints, and produces actionable recommendations for human review. This architectural shift—from extraction to interpretation—lays the groundwork for the fully agentic workflows we will explore in Chapter 14.

12.1 The synthesis challenge: from separate engines to unified decisions

12.1.1 The information asymmetry problem

12.1.2 The synthesis workflow in our 4-Layer Framework

12.1.3 The role of macroeconomic indicators

12.1.4 News sentiment vs. structured macro data

12.2 Setting up the simulation infrastructure

12.2.1 Defining the portfolio universe

12.2.2 Timing assumptions

12.3 Designing the system prompt

12.3.1 Bounded discretion under constraints

12.3.2 No date information

12.3.3 Output structure

12.4 Quantitative signal generation

12.5 The LLM portfolio advisor

12.6 Running the historical simulation

12.6.1 What this simulation does and does not show

12.6.2 Path dependence, not learning

12.7 Results and interpretation

12.7.1 Interpreting the trade-off

12.7.2 Visualizing strategy behavior

12.7.3 The decision log as audit trail

12.8 Human oversight and the production gap

12.8.1 Why automation is not the goal

12.8.2 The recommended workflow

12.8.3 What this simulation omits

12.8.4 Extensions and the path forward

12.9 Summary