Mercury 2 — Signal Engine

How Mercury 2's Machine Learning Works (Plain English)

What is XGBoost? — XGBoost is like a team of hundreds of tiny decision-makers. Each one looks at market data (price momentum, RSI, volume, etc.) and asks simple yes/no questions like "Is RSI below 30?" or "Is volume higher than average?". Individually, each decision-maker is weak — but combined, they vote together to make surprisingly accurate predictions. Think of it like asking 200 amateur analysts and going with the majority vote.

Why 3 models? — Mercury 2 runs 3 separate XGBoost models with different "personalities": Conservative (careful, fewer false alarms), Aggressive (more trades, higher risk), Balanced (middle ground). Their predictions are averaged — a coin only becomes a pick when all three agree it's likely to go up.

What is LightGBM? — Similar concept to XGBoost, but optimized for speed. Mercury 2 uses a LightGBM regressor (not classifier) to predict how much each coin will gain in the next 24 hours. This powers the "Top 5 Predicted Gainers" section — it's a watchlist, not a trading signal.

What do "16 models trained" mean? — Each of the 3 XGBoost models is trained on 2 years of 1-hour candles across 20 crypto pairs. The models learn patterns like "when RSI drops below 25 and volume spikes 3x, price usually rebounds within 4 hours". Plus the LightGBM gainer model = 4 total model files. Each model was trained on ~350,000 data rows.

How do models become picks? — Every 30 minutes, Mercury 2 feeds the latest price data through all 3 XGBoost models. If the average confidence (probability) is above 55%, and the risk engine approves (checks RSI, funding rate, trend, volatility), a pick is generated with specific TP (take profit) and SL (stop loss) targets.

What is DSR/PSR validation? — Before trusting a model, we run a statistical test called the Deflated Sharpe Ratio. It asks: "Could this model's performance just be luck from testing so many combinations?" If the answer is "probably yes" (DSR < 0.6), the model fails validation and Mercury 2 stops opening new trades until the next weekly retrain produces a better model. Current status: checking...

Abbreviation Legend: conf = ensemble confidence (avg probability from 3 XGBoost classifiers) | R:R = Risk-to-Reward ratio (TP distance / SL distance) | TP = Take Profit target | SL = Stop Loss level | ATR = Average True Range (14-period volatility measure) | RSI = Relative Strength Index (14-period momentum, 0-100) | F&G = Fear & Greed Index (0=extreme fear, 100=extreme greed) | SMA = Simple Moving Average | funding_z = Funding rate z-score (std deviations from 48h mean) | P&L = Profit and Loss | WR = Win Rate | TP1 = Take Profit 1 (1.5R, close 50%) | TP2 = Take Profit 2 (3R, close 25%, rest = runner) | Pos% = Remaining position size (100% → 50% after TP1 → 25% after TP2)