OmegaLoss

Defined in fynance.models.loss

class OmegaLoss(threshold=0., **kwargs)

Bases: BaseLoss

Negative Omega ratio as a differentiable loss.

\(\Omega = \frac{E[\max(r - L, 0)]}{E[\max(L - r, 0)] + \varepsilon}\), the ratio of expected gains to expected losses relative to a threshold L. Fully differentiable through torch.relu. Minimizing the loss maximizes the Omega ratio.

Parameters:

thresholdfloat, optional

Return threshold L separating gains from losses. Default 0.

**kwargs

Forwarded to BaseLoss (rf, period, eps).

Notes

Both gains and losses are \(O(|r - L|)\), so a fixed absolute eps is dimensionally wrong: on an all-gains batch (zero losses) the ratio would explode (e.g. -1e6) and dominate gradients. The denominator is therefore floored with a returns-scaled value |r - L|.mean() / MAX_RATIO (plus a bare eps backstop for the degenerate all-zero-diff batch), capping the ratio at roughly MAX_RATIO in the low-loss regime regardless of the return scale.

The ratio is then passed through a smooth saturating map, MAX_RATIO * tanh(ratio / MAX_RATIO), instead of a hard clamp. A hard clamp pins the loss to a constant on an all-gains batch and so zeroes the gradient in exactly the regime training still wants to push on; tanh is near-linear for normal-regime ratios (leaving their numerics unchanged) yet keeps a residual, non-zero gradient when the ratio is large. This keeps the loss finite and bounded while preserving the sign convention (minimizing it maximizes the ratio).

forward(y_pred, y_true=None)

Compute the negative Omega ratio (scalar).