Choosing Forecasting Features and Targets for Daily Stock Index Models

Summary

The document asks whether normalized prices, returns, or nominal price changes are preferable inputs and targets for a deep neural network forecasting daily stocks or indices. The response does not recommend one representation or provide a literature review. Instead, it argues that feature choice should follow from a clear definition of the prediction task and an understanding of the forces that drive the asset being forecast.

For an equity index, possible explanatory information includes its constituent stocks, valuation measures, interest rates, volatility, momentum, market regime, and investor confidence. The answer encourages matching candidate data to those mechanisms and considering whether daily observations capture them adequately or whether intraday information is needed. It offers examples of possible inputs, such as volume and put-call ratios, but no tests, citations, or evidence comparing them. The practical takeaway is a problem driven feature design process; model architecture and the choice between prices, returns, and changes remain unresolved and should be evaluated against the intended forecast and data.

Key ideas

• Define the forecasting objective before choosing input and target variables.
• Feature engineering should reflect plausible drivers of the stock or index being forecast.
• Potential index features include constituent information, valuation, rates, volatility, momentum, and market regime.
• Consider whether daily data captures the relevant dynamics or whether intraday observations are needed.
• The response supplies no empirical comparison of price, return, and change targets or specific literature references.

Tags

# Optimal Input and Target Variables for Forecasting Using a Deep Neural Network on Daily Stock/Index Data


# Optimal Input and Target Variables for Forecasting Using a Deep Neural Network on Daily Stock/Index Data












What is the optimal input and target variables for forecasting with a deep neural network on daily stock/index data? More specifically I’m training a temporal convolutional network, but a more general answer is also appreciated.

Are normalized closing prices, daily returns, or nominal daily changes better inputs? For target variables, which is more interesting: nominal price, returns, or daily changes?

Are there any literature references on these topics?

## Answer by John (score 2)

https://quant.stackexchange.com/a/74700

There are hundreds of articles online about how to forecast daily close price using some variant of machine learning, most of the are total crap.

What is it exactly you want to predict? do you understand the dynamics of the problem at hand? machine learning is plain useless without understanding of the fundamentals, feature engineering is about knowing the problem and being able to dissect it in small piece of information for the ML algo to digest.

So ask yourself what is fundamentally driving your equity index? its stock components? their P/E ratios? the interest rates? the VIX index? the current momentum? the market regime? the investors levels of confidence?

Then ask yourself which data captures that? maybe P/C ratios would be a starting point? volumes? would daily points be sufficient is intraday needed?

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.