Comparing Custom Loss Functions for Deep Learning Stock Selection

Summary

This study compares loss functions for neural models that forecast five-day cumulative returns for selected Chinese A-shares. It uses a DeepAlpha DNN as the main model and compares mean squared error with absolute error, Pseudo-Huber, negative information coefficient, and ordinal losses; it also applies weighted mean squared error to an LSTM. The workflow filters stocks using fundamental conditions, builds a transformed factor set, trains on roughly three years of data, and evaluates the following year with five-day rebalancing.

The reported 2023 comparison favors absolute error overall, citing a positive return and lower drawdown than the alternatives. Pseudo-Huber is described as having higher return but less stable performance, while negative IC optimization is challenged by the loss function’s non-convexity. The evidence is a single historical evaluation, and the supplied text leaves ordinal and weighted-loss results largely unspecified. It does not establish that the ranking generalizes across periods, universes, or implementations.

Key ideas

• The study tests whether loss-function choice changes neural stock-selection results for five-day return forecasts.
• Absolute error is reported as the strongest overall approach in the evaluated period.
• Pseudo-Huber is described as less stable despite having the highest return in the comparison.
• The negative IC objective is harder to optimize because the correlation-based function is not strictly convex.
• The reported evidence comes from a limited historical backtest and does not demonstrate out-of-sample robustness.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.