Deep Learning for Stock Pricing: Data Quality, Features, and Compute
Summary
This account summarizes a 2021 conference talk on using machine learning to price stocks. A simple example feeds market-wide daily open, high, low, close, and average prices into an LSTM time-series model. The speaker stresses preparing noisy financial data, handling new listings and limit moves, normalizing inputs, controlling overfitting, and avoiding look-ahead bias. A more complex production model is described as combining processed market data with financial statements, announcements, news, supply-chain information, and stock relationships in a multilayer perceptron.
The talk also emphasizes the computing and data infrastructure needed to train large models, including GPU clusters, distributed storage, and workload scheduling. It reports a long training time for one complex model on a single server and gives infrastructure specifications and performance claims for the firm’s supercomputer. These are speaker-reported examples, not independent evaluations of predictive performance. The account provides no model architecture details, out-of-sample results, transaction-cost analysis, or evidence isolating the contribution of deep learning to investment returns.
Key ideas
- A basic stock-pricing example uses daily price features as inputs to an LSTM time-series model.
- The speaker identifies data cleaning, normalization, overfitting controls, and look-ahead prevention as essential modeling tasks.
- A larger model combines market data with financial, news, announcement, supply-chain, and relationship features.
- Training large models can require substantial compute, storage throughput, and cluster scheduling.
- The talk’s reported infrastructure and business outcomes are not accompanied by independent performance validation.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.