Experiment: Generative RecSys + HyDE for hypothesis generation

[OnePunchMonk]

Summary

Explore using generative recommendation systems + HyDE (Hypothetical Document Embeddings) to propose novel trading hypotheses. Instead of purely literature-based discovery, generate synthetic "ideal strategies" and search for empirical evidence of them.

This could discover hypothesis combinations not yet written about in papers.

Motivation

Current Research Agent (Phase 1-3) is literature-dependent: only finds strategies already published. A generative + retrieval approach could:

• Generate hypothetical strategies from first principles ("what if we combined X + Y?")
• Score them by embedding similarity to regime contexts
• Retrieve supporting evidence from academic + market data
• Find novel combinations faster than literature search alone

Approach (HyDE-inspired)

1. Generate hypotheticals: Use Claude to generate 5-10 synthetic strategy descriptions given current regime
   • Example: "Short-lived mean reversion in high-VIX regimes with 2-hour windows"
2. Embed & retrieve:
   • Embed each hypothesis using strategy semantics
   • Retrieve nearest empirical patterns from historical backtests
   • Retrieve supporting research papers/blogs
3. Score & validate: Rank by evidence density and backtest Sharpe
4. Propose parameters: Map winning hypothetical to actionable parameter set

Task

1. Design hypothesis generation prompt (what makes a good synthetic strategy description?)
2. Implement embedding-based retrieval of:
   • Similar backtest results from memory (SQLite)
   • Related research findings (from Literature Agent)
3. Create scoring function: evidence density + backtest quality
4. Prototype on 1-2 market regimes (e.g., high-vol, low-vol)
5. Compare results: generated hypotheses vs. literature-only hypotheses

Files to Create/Modify

• src/agent/hypothesis_generator.py — Generative component (NEW)
• src/agent/hypothesis_scorer.py — Embedding + retrieval scoring (NEW)
• src/research/hypothesis_memory.py — SQLite schema for hypothesis archival (NEW)
• tests/test_hypothesis_generation.py — Evaluation (NEW)
• docs/HYPOTHESIS_GENERATION.md — Design doc (NEW)

Acceptance Criteria

• Hypothesis generation prompt is tested on multiple regimes
• Embedding-based retrieval finds similar historical patterns
• Scoring function ranks hypotheses by evidence density
• Prototype on 2+ market regimes shows 1-2 novel hypotheses per regime
• Backtest results of generated hypotheses vs. literature baselines
• Design doc written explaining approach and results

Success Metrics

• Does it find strategies NOT in the literature/grid?
• Do generated hypotheses have >0.5 Sharpe on validation?
• How much faster than literature search?
• False positive rate (hypotheses with <0.4 Sharpe)?

Related

• Depends on: Implement Research Agent Phase 1: Literature Discovery Tool #22 (Literature Agent; need retrieval foundation)
• Related: docs/RESEARCH_AGENT_DESIGN.md (complementary to existing agent)
• Inspiration: HyDE paper (Gao et al. 2022), Generative IR literature

Priority

🔵 EXPERIMENTAL - Novel approach worth prototyping; decide after Phase 1-3 if it adds value

Notes

This is speculative. Start only after Research Agent Phase 1-3 are solid. Could complement or replace literature-only discovery depending on results.