Why a neural net? The logistic-regression model assumes
linear feature relationships. The real edge in trading often comes from
interactions — e.g., "high RVOL + low IV + bullish regime" is more than the sum of those features. A 2-layer MLP w/ ReLU hidden units captures these interactions automatically.
Inputs: same 22 features. Hidden layer: 16 ReLU units. Output: sigmoid → P(win). Trained via backpropagation w/ SGD. Saved separately from the logistic model — use Ensemble to compare.