Unified Forward Test Dashboard

The Strategy

Uses on-chain and macro fundamentals — MVRV ratio (Market Value vs Realized Value), M2 money supply lag correlation, and monthly seasonality — to identify crypto that is fundamentally undervalued relative to on-chain activity.

Why It Works

• MVRV SMA Proxy: When market value dips below realized value, coins are historically undervalued
• M2 Liquidity Lag: Global money supply expansion leads crypto rallies by 2-3 months
• Monthly Seasonality: Crypto exhibits strong calendar effects (Jan/Oct historically bullish)

Risk Management

• Stop: Based on ATR volatility bands
• Target: Mean reversion to on-chain fair value
• Hold: 1-14 days typical

Forward Test Validation

All signals logged with full audit trail. No backtests — only forward performance counts.

The Strategy

Forex markets exhibit momentum during active sessions. When a major USD pair moves >0.5% in 3 hours with low volatility, the move typically continues for 2-5 hours.

Statistical Proof

• 3 independent sessions (Feb 17, 2026)
• 30 total trades
• 70% win rate (21W/9L)
• Binomial p-value: 0.021 (statistically significant)
• Total P&L: +$15.11 on $100K portfolio

Entry Criteria

• 3-period rate of change >0.5%
• ATR (volatility) <2%< /li>
• London session (08:00-17:00 UTC) preferred
• Major pairs: EURUSD, GBPUSD, USDJPY, USDCHF, AUDUSD, USDCAD

Risk Management

• Stop: 1.5% from entry
• Target: 2.5%
• Max hold: 8 hours

Why Institutions Ignore It

Forex momentum requires holding through sessions. Big firms can't tolerate overnight risk. Capacity limited to $5-10M per pair.

The Strategy

Multiple stock-picking strategies compete head-to-head in real-time forward testing. Includes Connors RSI-2, sector momentum, fundamental value, and technical breakout approaches.

Key Features

• Multi-strategy: 5+ strategies competing simultaneously
• Real picks: All entries logged at time of signal
• Full tracking: TP/SL monitored with automatic resolution
• Transparent: Wins AND losses tracked — no cherry-picking

Risk Management

• Position size: 2-5% per trade
• Stop loss: Strategy-dependent (2-5%)
• Take profit: Based on ATR multiples

Forward Test Status

Actively tracking 40+ open positions with real-time P&L. Worst strategies will be eliminated; best will be scaled.

The Strategy

Identifies Smart Money footprints using ICT (Inner Circle Trader) concepts: Fair Value Gaps, Order Blocks, and Liquidity Sweeps. Applied to meme tokens and high-volatility crypto where institutional flow creates exploitable patterns.

Why It Works

• Smart Money FVG: Institutional orders leave gaps that price revisits
• ICT Selective: Only trades high-probability setups with volume confirmation
• Meme edge: Retail-dominated tokens amplify smart money signals

Risk Management

• Stop: Below order block / FVG low
• Target: Opposing liquidity pool
• Hold: Hours to 3 days

Forward Test Validation

All signals from the Alpha Engine logged with full audit trail. Meme picks carry extra risk but also outsized reward potential.

The Strategy

Uses advanced statistical methods to identify crypto trading opportunities: Hurst exponent for regime classification (trending vs mean-reverting), variance ratio tests for momentum persistence, and classical pattern detection.

Key Components

• Hurst Regime Adaptive: H>0.5 = trending (momentum), H<0.5 = mean-reverting (fade)
• Variance Ratio Momentum: Detects persistent directional moves vs random walk
• Double Top/Bottom: Classical reversal patterns with volume confirmation
• MACD Divergence: Price/momentum divergence signals trend exhaustion

Risk Management

• Stop: ATR-based dynamic stops
• Target: Regime-dependent (trend-follow vs mean-revert)
• Hold: Hours to 7 days

Academic Basis

Hurst exponent (Mandelbrot, 1968), Variance Ratio test (Lo & MacKinlay, 1988). These are well-established statistical tools repurposed for crypto markets where retail dominance amplifies predictable patterns.

Architecture

• Model: 45% Random Forest (300 trees) + 55% XGBoost (300 rounds)
• Features: 20 engineered features from OHLCV + market data
• Training: 15,000 samples, 139 positive (0.9% pump rate)
• Threshold: 50% probability for pick generation

Key ML Discoveries

• Consolidation Range (17.5% importance): Tight price compression precedes breakouts
• Momentum Ignition (7.9x lift): 3+ green candles with rising volume
• Vol/MCap Ratio: Volume exceeding market cap = extreme pump signal
• Capitulation V-Bottom: ATL + volume spike = highest magnitude pumps

Risk Management

• TP1: +10% | TP2: +20% | SL: -7%
• Time exit: 48 bars (4H candles)
• TokenSniffer integration (score <30 = blocked)

Self-Improvement

Weekly retrain on accumulated outcomes. Drift detection via rolling 20-pick Z-score window. Model version tracking with improvement history.

Files

• claude_gainer_ml/live_scanner.py — Live predictor
• claude_gainer_ml/train_model.py — Training pipeline
• claude_gainer_ml/tp_sl_tracker.py — TP/SL tracker
• claude_gainer_ml/token_sniffer.py — Scam filter
• claude_gainer_ml/self_improver.py — Auto-retrain

Scoring System

• HIGH_VOLUME_RATIO: Volume/MCap significantly elevated
• NEAR_BREAKOUT: Price within 2% of 24h high
• MOMENTUM_BUILDING: Positive 1h + 24h price change
• TIGHT_RANGE_SQUEEZE: Low/high range <5% (compression)
• SMALL_CAP_MOVER: MCap <$100M with volume surge

Pick Criteria

• Gainer Score ≥ 40 for pick generation
• TP: +18% from entry | SL: -7%
• Tracked via GitHub Actions every 4 hours

Files

• crypto_gainer_ml/live_predictor.py — Live scanner
• .github/workflows/crypto-ml-tracker.yml — Automation

Architecture

• Models: XGBoost + LightGBM + Random Forest + Neural Net
• Data: CoinGecko top 200 coins, 5-year historical patterns
• Features: Volume ratios, momentum, compression, breakout proximity

Current Performance

• Tracking since Feb 20, 2026
• Picks tracked with TP/SL outcomes
• Full P&L accounting with profit factor

Files

• crypto_gainer_ml/live_predictor.py — Shared pipeline
• .github/workflows/crypto-ml-tracker.yml — Automation

Architecture

• Scanner: production_scanner.py — full cycle every 15 min (validate + generate + tweak)
• Validator: forward_validator.py — checks TP/SL against real Binance/Yahoo prices
• Auto-tuner: auto_tuner.py — dynamic TP/SL adjustment based on win/loss patterns
• ML Ranker: ml_ranker.py — trains Random Forest at 50+ closed picks
• Database: SQLite (alpha_engine/data/alpha.db)

Strategy Categories (100 Total)

• 33 Core Crypto: RSI-2, MACD, EMA Cross, Bollinger, Ichimoku, VWAP, Supertrend, etc.
• 10 On-Chain: MVRV, Hash Ribbon, NVT, Fear&Greed, SOPR, Funding Arbitrage
• 8 Event-Driven: Token unlock, liquidation cascade, narrative rotation, new pair
• 8 Advanced: Vol risk premium, dynamic momentum, GoPlus sniper, DVOL extreme
• 6 Community: ICT FVG, Swing Failure Pattern, Break of Structure
• 6 Spike: Volume spike, momentum ignition, breakout
• 4 Quant: Cross-sectional momentum, ATR breakout, whale detector
• 11 Forex: London breakout, carry trade, session momentum
• 14 Equity: RSI-2 SPY/QQQ, VIX spike reversal, sector rotation

Learning Cycle

• Every 15 min: Validate open picks, generate new signals, apply auto-tweaks
• At 50+ closed picks: ML ranker auto-trains on historical outcomes
• Continuous: Strategy tweaks (widen/narrow SL, adjust confidence thresholds)

Current Issues

• performance_snapshot.json shows 0 picks despite 29 active — data sync bug
• All forward_wr = 0.0 (insufficient data, need 30+ trades per strategy)
• 5 strategies at 100% failure rate: community_ict_fvg_selective, smart_money_fvg, altcoin_season_rotation

Architecture

• Scanner: live_scanner.py — 81 algorithms, Tier 1 + Scout mode
• Signal Tracker: signal_tracker.py — validates TP/SL against Binance prices
• Elimination Engine: elimination_engine.py — danger zone → probation → elimination
• ML Ranker: ml_signal_ranker.py — heuristic mode (<50 picks), RF auto-trains at 50+
• Database: SQLite (data/kimi_trading.db)

Signal Injection Pattern

Each scan cycle injects real-time data: order book depth, liquidation events, CoinGecko trending, forex rates, exchange netflow, social calls (Telegram/Twitter).

Learning Cycle

• Every 15 min: All 81 algorithms generate signals, ranked by ML weights
• Daily 6 AM UTC: Full refresh — re-evaluate all algorithm performance
• At 50+ closed: RF model auto-trains on signal outcomes
• Continuous: Elimination engine demotes/promotes algorithms

Current Issues

• All confidence scores fixed at 50% (uniform — suspicious)
• performance_stats.json 6 days stale (last update Feb 16)
• All forward_wr = 0.0 across all 24 live signals
• Duplicate signals on same symbols (MANA, ENJ, ALGO appear 2-3x)

Methodology

• Feature Engineering: 60+ features from OHLCV data — RSI, MACD, Bollinger, ATR, volume ratios, momentum, volatility regime indicators
• Model Training: Walk-forward validation with 7 folds per pair/timeframe. Each fold trains on expanding window, tests on unseen future data
• A/B Testing: 4 model variants compete per cycle. C_random_forest currently winning (27.5% avg score vs 25.6% XGBoost)
• Regime Detection: HMM (Hidden Markov Model) classifies market into Bull/Bear/Range/Crash states

Learning Cycle

• Daily 2 AM UTC: Full model retrain on latest data (793 models, ~44 min)
• Every 4 hours: Generate predictions using current best models
• Weekly: A/B test evaluation — promote winning variant, demote losers
• Sundays 2 PM UTC: Meta-labeler retraining (XGBoost on regime accuracy)

Current Issues

• Winner score only 27.5% — low predictive power
• Ensemble variant (stacked) actually performed worst (23.68%)
• Walk-forward on ALGO shows -64.5% PnL, 99% max drawdown
• Only 4 of 14 pairs profitable in backtest (TRX, BTC, ETH, DOGE)

Discord Integration

Posts hourly to Discord (channel 1469431505439948920) with top picks, confidence levels, regime state, and backtest validation metrics.

Methodology

• HMM Engine: Gaussian HMM with 4 hidden states trained on return distribution, volatility, and volume
• Position Sizer: Kelly criterion adjusted by regime confidence (Bull=2.5x, Bear=-0.5x, Range=0.75x, Crash=0x)
• Meta-labeler: XGBoost model trained weekly to predict regime accuracy — improves classification over time

Learning Cycle

• Weekdays 3:30 PM EST: Run regime detection + position sizing
• Sundays 9 AM EST: Meta-labeler retraining

Current Signals

5 active: AUDUSD (Strong Bull, 2.5x), AMZN (Mild Bear, -0.5x), USDJPY (Mild Bull, 1.5x), QQQ (Mild Bear, -0.5x), ATOM (Accumulation, 0.75x). Overall: 10 bulls, 18 bears, 15 neutral.

How It Works

• Battle Test (hourly): All strategies scored against real market data. Survivors graded A+ to F.
• Signal Tracker (every 2h): Tracks crypto/forex signals against outcomes. Auto-tweaks TP/SL/confidence until beating market.
• Elimination: Strategies below C grade get demoted. Below D get eliminated. Fresh challengers rotate in.

Outputs

• battle_test_results.json — survivors/eliminated with grades
• BATTLE_REPORT.md — human-readable markdown report
• OPTIMIZATION_REPORT.md — progress toward market-beating
• validation_results.json — per-signal outcome tracking

Architecture

• Bot: live_trading_bot_canada.py — paper trading simulation
• Capital: $10,000 virtual (DRY_RUN=true)
• Data: CoinGecko (free) + CryptoCompare (free)
• Schedule: Every 4 hours via GitHub Actions

Outputs

• trading_results.json — portfolio state, equity, return %
• PERFORMANCE_REPORT.md — summary with equity curve