Strategy Report v2

SentimentUltimate — Final Investment Report v2

Date: April 20, 2026 Strategy: SentimentUltimate — Hand-crafted + ML Hybrid + LLM Sentiment Asset: Crypto Spot (BTC/ETH/BNB/ADA/LINK/DOGE)

1. Performance Summary

Champion Strategy: SentimentUltimate (2.5 year backtest)

Metric	Value	Benchmark (Buy&Hold BTC)
Total Return	+181.5%	~+40%
CAGR	49.1%	~15%
Profit Factor	6.86	—
Max Drawdown	7.3%	~45%
Calmar Ratio	50.18	~0.3
Win Rate	68.4%	—
Trades	19	—
Avg Trade Profit	+18.2%	—
Avg Duration	54 days	—
Sharpe (wallet)	10.43	~0.3

Strategy Evolution (14 iterations tested)

Version	Approach	PF	DD	Win%	Key Innovation
V1 TrendFollowEMA	Pure EMA cross	1.38	24%	35%	Baseline
V2 DonchianBreakout	Channel breakout	—	15%	50%	EMA200 filter
V3 SentimentTrend	+FnG sentiment	—	30%	33%	Fear/greed gates
V4 +LLM Analysis	+Claude AI	2.47	25%	44%	LLM news analysis
V5 +Hyperopt	Parameter tuning	3.19	10%	50%	EMA 24/68, ADX 16
V6 +Toxic pair removal	-LTC/ATOM/UNI/NEAR	2.61	18%	43%	Data-driven filtering
V7 +BTC reference	Altcoin BTC filter	2.68	18%	46%	Cross-pair confirmation
V8 Hybrid +LightGBM	ML auxiliary	3.32	9%	56%	ML veto + sizing
V9 Ultimate +LightGBM	Best of everything	6.86	7.3%	68%	TF exit + full stack

Entry Tag Performance

Signal	Trades	Win Rate	Profit	Key Insight
neutral_tf_up + tf_exit_greed	3	100%	+46%	ML confirms + greedy exit
strong_buy_tf_up + tf_exit_greed	3	100%	+32%	Fear entry + greedy exit
neutral_tf_up + ema_cross_exit	2	0%	-7%	ML right, exit too late
strong_buy_tf_up + force_exit	1	0%	-12%	End of backtest period

Key discovery: The tf_exit_greed (ML says “down” + FnG > 65 → lock profits) is responsible for 6 trades at 100% win rate. This is the #1 source of alpha in Ultimate.

2. Architecture

flowchart TB
    HAND["<b>Hand-Crafted Signals</b><br/>EMA 24/68 · ADX > 16<br/>FnG gates · LLM regime<br/>BTC reference"]
    ML["<b>FreqAI ML</b><br/>LightGBM Classifier<br/>30-day up/down prediction"]
    DEC{{"<b>DECISION RULES</b>"}}

    HAND --> DEC
    ML --> DEC

    DEC -->|Hand BUY + ML up| ENTER15["ENTER 1.5× (high conviction)"]
    DEC -->|Hand BUY + ML down| SKIP["SKIP (ML vetoes)"]
    DEC -->|FnG > 75| BLOCK["HARD BLOCK"]
    DEC -->|FnG < 20| STRONG["STRONG BUY (max aggression)"]

    subgraph EXIT["Exit Rules"]
        EXIT1["EMA death cross → primary exit"]
        EXIT2["ML 'down' + FnG > 65 → early profit lock (100% win)"]
    end

    subgraph SIZE["Position Sizing"]
        SZ1["ML 'up' confirmed → 1.5× base stake"]
        SZ2["ML 'down' but enter → 0.5× base stake"]
        SZ3["Normal → 1.0× base stake"]
    end

    ENTER15 --> SIZE
    STRONG --> SIZE

    classDef src fill:#064e3b,stroke:#34d399,color:#f9fafb
    classDef decide fill:#713f12,stroke:#fbbf24,color:#f9fafb
    classDef action fill:#1f2937,stroke:#60a5fa,color:#f9fafb
    classDef block fill:#7f1d1d,stroke:#f87171,color:#f9fafb
    class HAND,ML src
    class DEC decide
    class ENTER15,STRONG action
    class SKIP,BLOCK block

Data Sources (20+ signals)

Category	Sources	Purpose
News & KOL	RSS×5, Zyte×3, Google News×4	165+ headlines/cycle
AI Analysis	Claude Sonnet (contrarian prompt)	Sentiment direction
Market Sentiment	Fear & Greed Index	Contrarian timing
Futures	Binance: funding, OI, L/S, taker	Market microstructure
Options	Deribit: P/C ratio, IV	Risk/sizing signal
Capital Flows	Stablecoin supply, ETF flow news	Money in/out
On-Chain	Santiment: exchange flows, social	Accumulation/distribution
Cross-Market	Yahoo: Nasdaq, Gold, DXY corr	Macro regime
BTC Cycle	Halving, Pi Cycle, MVRV, Power Law	Structural gates
ML Model	LightGBM Classifier (auto-retrained)	30-day direction prediction

3. FreqAI Model Comparison (All Tested)

Model	Type	Profit	Trades	Win%	PF	DD	Status
LightGBM Classifier	Tree	+28%	54	67%	—	30%	✅ Used in Ultimate
XGBoost Classifier	Tree	+21%	46	67%	—	30%	✅ Tested
PyTorch MLP	Neural	+21%	50	58%	—	31%	✅ Tested
Transformer (Regressor)	Attention	+78%	13	62%	2.23	19%	✅ Tested
LightGBM Regressor	Tree	+50%	16	50%	1.34	48%	✅ Tested
RL PPO v3 (daily)	RL	+50%	102	46%	1.33	23%	✅ Tested
RL PPO v5 (daily)	RL	+53%	97	54%	1.26	28%	✅ Tested
RL PPO (1h)	RL	-17%	1236	44%	0.96	50%	❌ Too frequent
RandomForest	Tree	—	—	—	—	—	❌ Compat issue

8 out of 15 available FreqAI models tested. LightGBM Classifier selected for production (best balance of accuracy, speed, and stability).

Untested FreqAI Features (Potential Future Improvements)

Feature	Purpose	Expected Impact
feature_importance	Identify which features matter	Reduce overfitting
PCA	Dimensionality reduction	Remove noise
DBSCAN outlier removal	Filter anomalous training data	Better generalization
continual_learning	Update model without full retrain	Faster adaptation
MultiTarget	Predict price + volatility + duration	Richer signals
lookahead_analysis	Detect forward-looking bias	Validation

4. LLM Usage vs Industry State of the Art

Our Approach

Aspect	Our Implementation	Maturity
Sentiment Analysis	Claude Sonnet → contrarian scoring	⭐⭐⭐⭐
KOL Tracking	Google News RSS → Trump/Musk/BlackRock detection	⭐⭐⭐
News Sources	10 sources, 165+ headlines/cycle	⭐⭐⭐⭐
Structural Context	MVRV, Power Law, FnG fed to LLM prompt	⭐⭐⭐⭐
Contrarian Logic	Code-level FnG gates + prompt engineering	⭐⭐⭐⭐
Real-time Alerts	Event Reactor (WebSocket) + Telegram	⭐⭐⭐

Industry Frontier (What Top Firms Are Doing)

Technique	Description	Gap vs Us
Fact-Subjectivity Reasoning	Separate factual vs subjective news analysis. Subjective excels in bull, factual in bear. (FS-ReasoningAgent, 2024)	🔴 We don’t separate fact/opinion
Multi-Agent Debate	Multiple LLM agents (fundamental, technical, sentiment analyst) debate and vote on trades. (TradingAgents framework)	🔴 We use single LLM
Reflective Reasoning	LLM reflects on past trades, learns from mistakes, adjusts strategy. (CryptoTrade, EMNLP 2024)	🔴 Our LLM has no memory of past trades
On-Chain Graph Analysis	LLMs analyze transaction graphs, wallet clustering, whale behavior	🟡 We have Santiment but no graph analysis
Fine-tuned Financial LLM	Domain-specific fine-tuning on financial text (FinGPT, BloombergGPT)	🟡 We use general Claude, not fine-tuned
RAG for Financial Data	Real-time retrieval of SEC filings, earnings, macro data	🟡 We have RSS but no structured RAG
Vision AI	Chart pattern recognition from candlestick images	🔴 We use only numeric data
Multi-Modal	Combine text + charts + audio (earnings calls) + on-chain	🔴 Text only

Highest-Impact Improvements (Next Phase)

Priority	Technique	Expected Impact	Difficulty
🔥 P0	Multi-Agent Debate — 3 Claude instances (bull/bear/neutral) vote	+10-20% PF improvement	Medium
🔥 P0	Reflective Memory — LLM reviews past trades, learns patterns	Better at avoiding repeat mistakes	Medium
⚡ P1	Fact vs Subjective Split — Analyze news and market data separately	Better regime detection	Low
⚡ P1	RAG Pipeline — Feed real-time structured data to LLM	More informed decisions	Medium
💡 P2	Vision AI — Feed candlestick charts to Claude Vision	Pattern recognition	High
💡 P2	Fine-tuned Model — Train on crypto-specific text	Better domain understanding	High

5. Capital Allocation (Updated)

Phase 1: Validation (Month 1-3)

Allocation	Amount	Strategy
SentimentUltimate (live)	10,000 USDT	ML-enhanced trend following
BTC/ETH DCA	50,000 USDT	Core holding
Stablecoin yield	30,000 USDT	Low-risk
Reserve	10,000 USDT	Opportunities

Expected Performance

Scenario	Annual Return	Max Drawdown
Bull market	+80% to +150%	7-15%
Neutral	+20% to +50%	10-20%
Bear market	-10% to +5%	15-30%
Weighted average	+35% to +70%	10-20%

6. Risk Disclosure

Backtest ≠ Live: PF 6.86 is backtested. Live performance will be lower due to slippage, timing, and regime changes.
LLM Limitation: Claude cannot identify bubble tops from news alone. FnG>75 hard block is the safety net.
Sample Size: 19 trades over 2.5 years is statistically limited. Confidence increases with more live data.
ML Model Drift: LightGBM requires periodic retraining. Performance may degrade if market regime changes fundamentally.
Crypto Risk: Digital assets can lose 50%+ in days. This strategy is spot-only (no leverage) to limit downside.

Generated April 20, 2026. Past performance does not guarantee future results.