Skip to main content

Review & Reflection System

The review system gives Quant Flow a memory that persists across trading sessions and improves over time. It consists of 6 independently toggleable enhancements, each backed by academic research.

Overview​

Trade closes
    ↓
6a: InstantReflector      — rule-based, no LLM, confidence update
    ↓
ReviewAgent               — stores lesson in experience database
    ↓
6b: Regime tag            — lesson tagged with market regime
6c: Lesson type tag       — positive / negative / unknown
6d: Source type tag       — factual / subjective / mixed
    ↓
Next decision
    ↓
ReviewAgent.get_similar_lessons()
    → filters by regime match (6b)
    → applies bias correction (6c)
    → applies fact/subjective weighting (6d)
    ↓
VFT section injected into prompt (Step 4: Review Match)

Weekly trigger:
6a: WeeklyReflector       — LLM-based strategy analysis
6e: PromptMetaReflector   — suggests prompt improvements

Enhancement 6a: Dual-Granularity Reflection​

Paper: Adaptive Multi-Agent Bitcoin Trading (arXiv:2510.08068)

Instant Reflection (after every trade close)​

  • Pure rule-based, no LLM call
  • If the trade was profitable: matched lesson confidence × 1.05
  • If the trade was unprofitable: matched lesson confidence × 0.95
  • Ultra-fast — completes in milliseconds

Weekly Reflection (every Monday at 8:00 by default)​

  • Full LLM call — analyzes the past week's trades
  • Detects systematic biases (e.g., "consistently held too long in ranging markets")
  • Generates strategy-level adjustment suggestions
  • Produces a summary report saved to logs/
review_agent:
  instant_reflection_enabled: true
  weekly_reflection_enabled: true
  weekly_reflection_day: 0     # 0=Monday
  weekly_reflection_hour: 8

Enhancement 6b: Regime-Aware Memory​

Paper: Adaptive Memory for Bitcoin Regime Detection (engrXiv 2025)

Every lesson is stored with a source_regime field (trending / ranging / volatile / unknown). When retrieving similar lessons for a new decision:

  • Regime matches: normal similarity score
  • Regime mismatches: similarity score × regime_mismatch_factor (default 0.4)

The Verbal Fine-Tuning (VFT) section in the prompt labels lessons:

[趋势市经验] / [Trending Market Lesson]
[震荡市经验] / [Ranging Market Lesson]
review_agent:
  regime_aware_enabled: true
  regime_mismatch_factor: 0.4

Enhancement 6c: Confirmation Bias Protection​

Paper: FinCon (arXiv:2407.06567) + Selective Memory Equilibrium

Without protection, the experience database naturally fills with positive lessons (trades that worked) while negative lessons (mistakes) get evicted. This creates a false picture of past performance.

Protection mechanisms:

  • Maximum ratio of positive lessons enforced (max_positive_ratio: 0.7 → at most 70% positive)
  • When evicting old lessons, negative lessons are protected from disproportionate removal
  • Negative lessons get a confidence boost (negative_confidence_boost: 1.15) so they rank higher in retrieval

In the VFT prompt section, negative lessons are prefixed with [Avoid]:

[Avoid] Do not enter long on RSI > 72 during ranging markets — 
historically results in mean-reversion losses.
review_agent:
  bias_protection_enabled: true
  max_positive_ratio: 0.7
  negative_confidence_boost: 1.15

Enhancement 6d: Fact-Subjective Split​

Paper: FS-ReasoningAgent (arXiv:2410.12464, ICLR 2025)

Lessons are tagged by source type:

  • Factual: based on hard indicator data (RSI value, price level, funding rate)
  • Subjective: based on interpretation or qualitative analysis (sentiment, news impact)
  • Mixed: combination of both

Dynamic weighting:

  • Trending markets: subjective lessons boosted (trending_subjective_boost: 1.3)
  • Ranging/volatile markets: factual lessons boosted (ranging_factual_boost: 1.3)

In the VFT section, lessons are labeled [Factual] or [Subjective] so the LLM can appropriately weight them.

review_agent:
  fact_subjective_split_enabled: true
  trending_subjective_boost: 1.3
  ranging_factual_boost: 1.3

Enhancement 6e: Prompt Meta-Reflection​

Paper: ATLAS Adaptive-OPRO (arXiv:2510.15949)

After each weekly reflection, PromptMetaReflector evaluates prompt quality on 4 dimensions:

DimensionWhat It Measures
FinCoT completenessDid the LLM complete all 6 steps?
Lesson citation rateHow often did it reference past lessons?
Decision consistencyDid confidence scores align with outcomes?
Confidence calibrationWere high-confidence decisions actually more accurate?

Based on the evaluation, it generates specific prompt improvement suggestions. These are saved to logs/prompt_optimization/ for human review — they are not automatically applied.

review_agent:
  prompt_meta_reflection_enabled: false
  prompt_optimization_dir: "logs/prompt_optimization"
Human-in-the-Loop

Prompt meta-reflection generates suggestions but does not modify prompts automatically. A human must review and apply changes to files in prompts/.

All Configuration Options​

review_agent:
  # 6a
  instant_reflection_enabled: false
  weekly_reflection_enabled: false
  weekly_reflection_day: 0
  weekly_reflection_hour: 8

  # 6b
  regime_aware_enabled: false
  regime_mismatch_factor: 0.4

  # 6c
  bias_protection_enabled: false
  max_positive_ratio: 0.7
  negative_confidence_boost: 1.15

  # 6d
  fact_subjective_split_enabled: false
  trending_subjective_boost: 1.3
  ranging_factual_boost: 1.3

  # 6e
  prompt_meta_reflection_enabled: false
  prompt_optimization_dir: "logs/prompt_optimization"

Start with 6a and 6c — they have the most immediate impact and lowest cost:

review_agent:
  instant_reflection_enabled: true
  bias_protection_enabled: true

Add 6b and 6d after you have at least 50 closed trades in the database (regime-aware memory needs sufficient history to be meaningful).

Next Steps​