Sequence Deep Learning in Quant Investing

Article BigQuant

Summary

The document outlines how sequence learning can fit into a conventional quantitative investment process: feature extraction, individual asset return prediction, portfolio construction, and trade execution. It describes using price and volume measures, fundamental and sentiment data, and less conventional sources as model inputs. Sequence models can support both medium-term return forecasts and short-horizon forecasts of prices or volume used to manage execution costs. Portfolio construction is framed as a constrained optimization problem that may also use machine learning, though it is not itself presented as sequence prediction.

The discussion introduces recurrent neural networks and LSTMs, then points to attention, Transformers, and CNNs as sequence modeling approaches. It also notes that reinforcement learning has been more established in games and robotics than in investing, and requires substantial computing resources and simulation. The evidence is a summary of a Two Sigma webinar, not a reported trading test: no datasets, performance measures, or implementation details are supplied. The central practical lesson is to use deep learning within a familiar investment workflow, while recognizing that data and infrastructure matter alongside model choice.

Key ideas

Sequence models can be applied to feature processing and individual asset return forecasts.
Short-horizon forecasts of price and volume can inform execution decisions and transaction cost control.
Portfolio construction remains a constrained optimization problem, even when machine learning helps define objectives or constraints.
The overview covers recurrent networks, LSTMs, attention, Transformers, and CNNs as sequence modeling approaches.
The document reports webinar concepts rather than empirical strategy results, and gives no performance evidence.

Applying Sequence Deep Learning Across the Quantitative Investment Pipeline

Summary

Key ideas

Tags