weight_by_softmax

API documentation for tradeexecutor.strategy.weighting.weight_by_softmax Python function.

weight_by_softmax(alpha_signals, temperature=2.0)

Softmax-temperature weighting that smoothly interpolates between equal-weight and winner-take-all allocation.

Applies the softmax function with a temperature parameter to transform raw signal values into portfolio weights:

w_i = exp(signal_i / T) / sum_j(exp(signal_j / T))

The temperature T controls concentration:

• T → ∞: converges to equal weights (1/N)

• T → 0: converges to winner-take-all (100% in top signal)

• T ≈ 1-2: moderate tilt toward higher signals

Weights always sum to 1.0, eliminating the cash drag problem inherent in equal weighting under greedy allocation loops.

Pros:

• Single tunable parameter with intuitive behaviour

• Smooth, differentiable — no hard cutoffs or discontinuities

• Numerically stable (uses max-subtraction trick)

• Naturally handles any number of assets

Cons:

• Sensitive to signal scale — signals should be comparable magnitude

• Low temperatures can over-concentrate into noisy top signals

• Does not account for correlation between assets

References:

• Zakamulin (2025), “Entropy-Regularized Portfolio Selection via Softmax Sharpe Allocation”, SSRN 5539560

• Britten-Jones (1999), “The Sampling Error in Estimates of Mean-Variance Efficient Portfolio Weights”, Journal of Finance

• Goodfellow, Bengio & Courville (2016), “Deep Learning”, Ch. 6.2.2 — softmax as generalisation of logistic function

Parameters:

• alpha_signals (Dict[int, float]) – Signal objects keyed by pair ID.

• temperature (float) – Controls equal↔concentrated tradeoff. Higher = more equal, lower = more concentrated.

Returns:

Weight map summing to 1.0.

Return type:

Dict[int, float]