Skip to content
All library documents

Using Cross-Asset Features in Machine Learning for Stock Returns

Article Quant Q&A · Author: Eka
Read the original

Summary

The document asks whether quants build one prediction model per stock or train a model that can generalize across many securities. It highlights two concerns with training on a single stock and applying that model elsewhere: survivorship bias and the narrow patterns learned from one asset. The response emphasizes cross-asset dependence, suggesting that models can use features from multiple assets to capture relationships in returns.

It points to vector autoregression as an econometric example and to machine learning research in empirical asset pricing as a source of methods. It also describes a possible single-asset setup using high-frequency features, with synchronized data from other assets added as inputs. These are illustrative directions rather than a complete training recipe. The document does not specify target construction, validation, model selection, or evidence that a pooled or single-asset design predicts prices reliably.

Key ideas

  • Training on one stock and transferring the model may miss asset-specific behavior and introduce survivorship bias.
  • Cross-asset features can help a model use relationships among securities.
  • A model may predict one asset while including synchronized data from other assets as inputs.
  • The document presents modeling options but gives no validation design or empirical performance results.

Tags

Machine learningEquitiesStatistics
Full text
# How quants use ML models for stock market prediction


# How quants use ML models for stock market prediction












I am learning machine learning to use it for stock market price forecasting. While doing that I got this question. If we take any country with stock exchange they have more than one investment assests for trading and investing such as commodity, stock, futures,option,forex etc.

Lets say a quant wants to make a machine learning model for stock price prediction in US market. There are thousands of companies (about 2800) stocks are listed in NYSE.

How a quant will make a ML model for predicting stock price?

Currently what I am doing is that I teach a ML model using data(OHLC) from a particular stock(eg APPL between 1990-2016) and use this as my standard model. Then I use this standard model to predict AAPLs or other 2800's stock price(eg: IBM,F). But this method has many drawbacks and these are the once which comes into my mind 1) survival ship bias 2) trained only in one stock and only have knowledge about its pattern.

so how a quant trains a ML model? In this US stock market example he/she/it instead of creating a general model for 2800 stocks whether they create 2800 individual models for 2800 stocks and use it for prediction of that stock

example for predicting GM future price, a ML model is created for GM called GMmodel and then it is used for prediction GMs future price. am I right?

- Here I meant quant as individuals and institutions who use machine learning for trading

## Answer by Jacques Joubert (score 5)

https://quant.stackexchange.com/a/44878

The correlation matrix is a very important part of modeling stock returns. It is often better to build a model that takes in multiple assets features so that it can use this correlation to its advantage.

A good example of this is a VAR model from econometric. A great example in the machine learning context is the paper titled Empirical Asset Pricing via Machine Learning. This paper will provide great insights!

The question then becomes one of should I have multiple asset returns inputs and outputs or should I model a single asset at a time using multiple assets?

If we look at Lopez de Prado (another great ML researcher) we will see in his book titled Advances in Financial Machine Learning that because of the volume sampling being used, it would seem that his techniques speak to modeling one asset at a time using HFT features which can be constructed using mainly the price data. However if you sample additional assets using the same timestamps as the primary asset, then I do believe you could use multiple assets price data.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

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