FinCoT Structured Reasoning

FinCoT (Financial Chain-of-Thought) is a structured prompting technique that forces the LLM to reason through trading decisions in a fixed 6-step sequence, instead of generating a conclusion directly.

Academic Basis

Paper: FinCoT (arXiv:2506.16123)

Reported results:

• +17.3 percentage points accuracy improvement over vanilla prompting
• −8.9× token consumption compared to unconstrained chain-of-thought

The 6-Step Reasoning Chain

Every trading decision follows this exact sequence:

Step	Name	Question Asked
1	Trend Confirmation	Are multi-timeframe trends aligned?
2	Entry Signal	Which technical indicators triggered? (list exact values)
3	Sentiment Check	Any contrarian signals from funding rate / F&G / debate?
4	Review Match	Which historical lesson matches this setup?
5	Risk Calculation	Stop-loss/take-profit distance, R:R ratio, fee coverage
6	Final Decision	Synthesize steps 1–5 → LONG / SHORT / HOLD + confidence

Why This Works

• Reduces hallucination — the LLM cannot skip to a conclusion without completing each step
• Forces evidence citation — step 2 requires exact indicator values, not vague impressions
• Contrarian gate — step 3 prevents ignoring signals that conflict with the desired direction
• Explicit risk math — step 5 requires calculating fee coverage before committing to a trade

How to Enable

FinCoT is built into the prompt templates. Select a FinCoT-enabled prompt set:

# config.yaml
prompt:
  set: nof1-improved    # recommended — full FinCoT + enhanced data

All 8 prompt sets (default, aggressive, conservative, nof1, nof1-improved, realtime, realtime-eng) have FinCoT integrated.

No Extra LLM Calls

FinCoT is a prompt engineering technique — it does not add extra API calls. The 6 steps are completed within a single LLM response.

Prompt Template Location

Templates are Jinja2 files under prompts/. The FinCoT structure is embedded in the system prompt. Key template variables:

Variable	Source	Content
{{ debate_summary }}	Debate agent	Bull/bear arguments
{{ volatility_alert }}	Market monitor	Price spike context
{{ regime_hint }}	Regime adapter	Market state hint
{{ cex_funding_signal }}	Data enricher	CEX vs DEX funding divergence
{{ onchain_summary }}	Data enricher	MVRV, SOPR, Fear & Greed

Comparing Prompt Sets

Set	Risk Tolerance	R:R Minimum	Best For
default	Moderate	1.5	General use
conservative	Low	2.0	Bear markets, uncertain conditions
aggressive	High	1.2	Strong trending markets
nof1-improved	Moderate	1.5	Recommended: full feature integration
realtime	Moderate	1.5	Fast-moving markets (uses price action over indicators)

Example LLM Output (Abridged)

Step 1 - Trend Confirmation:
  - 1h: Bullish (price above MA20, MA50, MA200)
  - 4h: Bullish (price above MA20, slight divergence at MA200)
  - Alignment: 2/2 confirmed bullish

Step 2 - Entry Signal:
  - RSI(14): 58.3 — momentum without overbought
  - MACD: bullish crossover 3 candles ago, histogram positive
  - Bollinger: price at mid-band, room to upper band

Step 3 - Sentiment Check:
  - Binance funding: +0.01% (slightly positive, no contrarian signal)
  - Fear & Greed: 62 (Greed) — mild caution
  - Debate: Bulls 72% vs Bears 28% — no strong contrarian signal

Step 4 - Review Match:
  - Matches "RSI momentum entry after consolidation" lesson (confidence: 0.78)
  - Lesson outcome: positive in 7/9 similar setups

Step 5 - Risk Calculation:
  - Entry: $43,250 | Stop: $42,100 (2.66%) | Target: $45,800 (5.90%)
  - R:R ratio: 2.21 ✓ (minimum 1.5)
  - Fee coverage: spread 0.05% × 2 = 0.10% vs 5.90% target = 1.7% of profit

Step 6 - Final Decision:
  Action: LONG | Confidence: 0.72 | Leverage: 5x

Next Steps

• Bull/Bear Debate — the debate feeds into Step 3
• Review System — historical lessons feed into Step 4
• config.yaml Reference — prompt set selection