ResNeXt Multi-Task Learning for Financial Forecasting

Summary

The document introduces a multi-task learning design that uses ResNeXt to extract features from financial data and feed several prediction tasks. It explains grouped convolutions, residual connections, bottleneck blocks, and cardinality, then describes a model with a shared feature extractor, task-specific representations, and separate output layers. Classification and regression tasks use different loss functions, combined as a weighted total; training is described as individual-task pretraining followed by joint fine-tuning with Adam and learning-rate adjustment.

The practical section outlines an MQL5 implementation of the bottleneck component, including convolution, normalization, and tensor transformations. The article presents the architecture as a way to reduce feature-engineering demands and computational load while modeling varied financial inputs. It provides no completed trading evaluation or measured forecasting results: implementation of the broader multi-task system and testing on historical data are deferred to a later installment. Claims about accuracy and adaptability should therefore be treated as motivation, not demonstrated evidence in this article.

Key ideas

• ResNeXt uses grouped transformations and residual connections to build deep feature extractors.
• Cardinality controls the number of parallel transformations within a block and offers another scaling dimension alongside depth and width.
• A shared representation can support several related financial prediction tasks through task-specific output layers.
• Classification and regression outputs can use separate loss functions combined in a weighted multi-task objective.
• The article describes an MQL5 bottleneck implementation but does not report historical-data performance.

Tags

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