🧠 What Is This System Doing?
Antigravity ML Crypto Trainer — Explained
This system uses machine learning (Gradient Boosting + Random Forest classifiers) to
predict which crypto pairs will hit their Take Profit (TP) target before their
Stop Loss (SL) within a set time horizon. Here's exactly what it does:
- Data Collection: Pulls 720+ hourly candles for 14 crypto pairs from Kraken exchange
(BTC, ETH, SOL, XRP, DOGE, DOT, SHIB, ALGO, ARB, APE, INJ, FET, DYDX, TRX)
- Feature Engineering: Computes 40+ technical indicators per pair — RSI (3 periods),
EMAs (4 periods), SMAs (3 periods), MACD, Bollinger Bands, ATR, volume ratios, returns (4 periods),
volatility, rate of change
- Target Definition: Binary classification — does price reach +2% (TP) before hitting
-1% (SL) within 12 candles? This is a realistic trading target, not just "price go
up"
- Training: Trains 2 models per pair (Gradient Boosting + Random Forest) with 80/20
time-series split. No data leakage — future data never seen during training
- Scalping Models: Separate models with tight TP/SL (0.5%/-0.3%) and fast indicators
for short-term trades
- Swing Models: Separate models with wide TP/SL (4%/-2%) and trend indicators for
multi-day holds
- TP/SL Optimization: Grid search across 35 TP/SL combinations per pair to find
optimal risk-reward
- Backtesting: Walk-forward out-of-sample testing on unseen data — no curve-fitting
📊 Current Model Performance
Pairs Trained
—
14 crypto pairs from Kraken
Total Models
—
Base + Scalping + Swing
Avg Win Rate
—
Out-of-sample accuracy
Best F1 Score
—
Precision × Recall balance
Training Cycles
—
Daily retraining at midnight UTC
Last Trained
—
Next: midnight UTC
⏱️ Reliability Timeline — When Can You Trust It?
How Long Until Reliable?
ML models improve with more data and more training cycles. Here's the roadmap:
Week 1-2: EARLY STAGE (Current)
Models are training on 720 hourly
candles (~30 days of data). F1 scores range 0.05–0.43. Some pairs show edge (SOL, SHIB,
XRP), but models need more diverse market conditions. Do NOT trade on these signals
yet.
Week 3-6: DEVELOPING
With 2000+ candles and 20+ training
cycles, models will have seen bull/bear/sideways markets. F1 scores should reach 0.30+
consistently. Paper trading begins — forward picks tracked but no real money.
Week 7-12: ADVANCED
Models validated across multiple market
regimes. Pairs with F1 > 0.40 and WR > 55% graduate to "reliable" status. Self-improvement
loop ensures failed predictions feed back into training.
Month 4+: PRODUCTION
Battle-tested models with 100+ training
cycles, 10,000+ candles, and proven edge on forward picks. Ready for small-position live
trading with strict risk management.
🏆 Current Top Picks (Forward Predictions)
⚠️ FORWARD PICKS STATUS
These picks are generated by the ML models ranking pairs by prediction confidence. They are NOT
live trades — they are tracked predictions. When a pick hits its TP or SL, the result feeds
back into the next training cycle. Over time, this self-improvement loop makes the models smarter.
📋 Pair-by-Pair Model Results
| Pair |
Best Model |
Accuracy |
Precision |
F1 Score |
Optimal TP |
Optimal SL |
Signal |
🔄 Self-Improvement Loop
How the System Gets Smarter Every Day
- Daily Retraining: Every midnight UTC, the system fetches fresh market data and
retrains all models. Each cycle adds ~24 new candles per pair to the training set.
- Forward Pick Tracking: Active picks are monitored — when TP/SL is hit, the outcome
is recorded. Winners and losers inform the next training cycle.
- Feature Evolution: 40+ technical indicators are recomputed with latest data. Models
learn which indicators matter most for each pair.
- TP/SL Optimization: Optimal take-profit and stop-loss levels are re-optimized daily
for each pair based on recent price action.
- Walk-Forward Validation: Backtests always use out-of-sample data to prevent
overfitting. If a model's edge disappears, it gets flagged.
Model Improvement Progress
Early (Current)
Developing
Advanced
Production