ANTIGRAVITY_ML_TRAINER

Live Picks Tracker — Real Forward Performance (Not Backtest)

        🕐 Dashboard updated: — |
        📡 Next scan: midnight UTC (7:00 PM EST) |
        🔄 Discord: hourly updates
      

📊 ML Dashboard 💻 Source Code ⬅️ Back

Total Picks

—

Wins

—

Losses

—

Win Rate

—

Total P&L

—

Active Now

—

🔬Forensic Analysis — Why Each Pick Failed

Loading pick data...

📡Active Picks (Live) 0

Pair	TF	Dir	Picked (EST)	Entry	TP	SL	Prob	Conf	P&L	Why Picked
Loading...

📋Closed Picks (Results) 0

Pair	TF	Dir	Entry	Close	P&L	Outcome	Duration	Model
Loading...

🧪Model Quality Assessment (Base Models)

🛠️What's Being Fixed — Improvement Roadmap

✅ Fix #1: Raise Confidence Threshold (0.45 → 0.60) ✅ DEPLOYED

The model was taking picks with probability 0.48-0.52 (coin-flip territory). Now requires 0.60+ probability to generate a pick. This alone would have filtered out 3 of the 5 losses.

Status: Code committed & pushed — activates on next scan at midnight UTC

✅ Fix #2: Widen SL from 0.75× ATR to 1.5× ATR + 0.5% min ✅ DEPLOYED

The 15m SL was ~0.2% — normal 15-minute noise wipes this out instantly. BNB lost 0.22% (within normal volatility). Wider SL gives trades room to breathe while maintaining R:R ratio. Also enforces 0.5% minimum SL distance.

Status: Code committed & pushed — activates on next scan at midnight UTC

✅ Fix #3: BTC Regime Filter — Block BUY during bearish moves ✅ DEPLOYED

All 5 losses were BUY signals during a market-wide dip. Now checks if BTC dropped >0.5% in 4h AND >0.2% in 1h — if so, all BUY signals are blocked. Would have prevented all 5 losses.

Status: Code committed & pushed — activates on next scan at midnight UTC

✅ Fix #4: Max 3 Concurrent Picks per Direction ✅ DEPLOYED

All 5 picks were BUY, all same direction = correlated risk. Now limits to max 3 BUY + 3 SELL at any time to prevent correlation blow-up.

Status: Code committed & pushed — activates on next scan at midnight UTC

✅ Fix #5: Prefer 1h/4h Timeframes Over 15m ✅ DEPLOYED

15m models have the lowest F1 scores and highest noise. Timeframe priority reordered to: 1h → 4h → 1d → 15m (last).

Status: Code committed & pushed — activates on next scan at midnight UTC

✅ Fix #6: v1.5 Research Overhaul — 9 Evidence-Based Fixes ✅ DEPLOYED

Complete retraining of 36 pair/TF combos with: isotonic calibration (CalibratedClassifierCV), purged walk-forward CV (75/25 split + 20-bar gap), early stopping (50 rounds), meta-labeling M2 trade/no-trade filter, SMOTE disabled, reduced complexity (300 trees, depth 4-5). 197 model files updated.

Status: All models retrained and pushed to production — Feb 23, 2026

🔄 Fix #7: Feed Losses Back Into Training (Self-Improvement) ⏳ ONGOING

Every closed pick (win or loss) gets added to training data. After 30+ cycles, the model learns which market conditions lead to losses and adjusts weights accordingly.

ETA: Ongoing — accumulating data with every scan cycle. Meaningful improvement expected after 30+ cycles.

🎯When Will We See Wins?

Honest Timeline Based on Data

5 losses is too small a sample to conclude the model is broken. That said, the losses revealed real bugs (SL too tight, no regime filter) that need fixing.
With fixes applied, expect first TP hit within 24-48h. The 1h/4h models with wider SL have 2:1+ R:R ratio, meaning even 40% win rate is profitable.
Meaningful forward stats require 30+ closed picks (~1-2 weeks). Until then, any stat is noise.
The model architecture IS sound — Gradient Boosting + Random Forest with 40+ features is proven for tabular financial data. The issue was pick selection, not prediction.
Model CAN learn — each daily retraining adds 24 new candles. After 30 days, models will have 720+ new data points to learn from. The self-improvement loop is real.
Compare fairly: Simpleton Signals v0.07 has 51.3% win rate over 200+ trades. We need 200+ trades for a fair comparison. 5 trades proves nothing statistically.