V2 Model Improvements

V2 models fix critical bugs in the original training code:

Issue	V1 (Broken)	V2 (Fixed)
Double-softmax	Loss stuck at 0.3133, no learning	Loss decreases, models learn
Regularization	None (instant overfit)	BatchNorm + Dropout(0.3)
'both' spread	Used only log features (90)	Uses both features (180)
Probability calibration	0.99-1.0 (overconfident)	0.50-0.90 (realistic)

Key change: V2 probabilities are properly calibrated. A 0.75 probability now means something real, unlike V1's meaningless 0.99.

---

Probability Distribution (V2 vs V1)

V1 Models (BROKEN):
  Probabilities: 0.99-1.0 for everything
  Interpretation: Model not learning, outputting constant predictions

V2 Models (FIXED):
  both_100ep_v2: 0.50-0.87, mean=0.67
  both_50ep_v2:  0.50-0.90, mean=0.66
  log_75ep_v2:   0.50-0.92, mean=0.65
  log_50ep_v2:   0.50-0.88, mean=0.63
  coint_25ep_v2: 0.50-0.90, mean=0.62
  coint_10ep_v2: 0.50-0.81, mean=0.59

V2 probabilities are properly calibrated - use p>0.70 threshold (not p>0.85).

---

Model Performance Summary

Model	Cointegrated Pairs	Prob Range	Best For
both_100ep_v2	9	0.71-0.82	Most signals, balanced
both_50ep_v2	7	0.72-0.83	Highest individual probs
log_50ep_v2	6	0.71-0.78	Traditional pairs
log_75ep_v2	4	0.71-0.79	Conservative
coint_25ep_v2	3	0.70-0.79	Extreme z-scores
coint_10ep_v2	3	0.71-0.74	Quick validation

Recommendation: Use both_100ep_v2 as primary model - most cointegrated pairs detected.