Testing Lagged and Transformed Predictors for Stock Returns

Summary

The document asks how to search for predictive relationships between a target series and many candidate variables. Its example is an inverted, three-month-delayed variable that appears to lead stock prices, and the proposed application is predicting future S&P 500 price changes from assets such as gold, oil, and currencies. The author seeks traditional statistical techniques as well as a comparison with machine-learning approaches.

The text poses the research problem but supplies no proposed tests, results, or recommendation about which method works best. It leaves open practical questions such as how to select transformations and lags, assess whether apparent lead relationships persist out of sample, and account for testing many candidate predictors. Those issues matter because an observed lagged association does not by itself demonstrate reliable forecasting power or causality.

Key ideas

• The question concerns finding predictive relationships across many candidate time series and a target stock series.
• The example predictor is inverted and delayed before comparison with the target.
• The proposed application uses other asset classes to forecast future S&P 500 price changes.
• The document asks about both traditional statistical methods and machine-learning approaches but provides no method or results.
• Testing many variables, transformations, and lags can produce apparent predictive relationships that require validation.

Tags

# Which technique determines if var x1 leads var y? Assuming var x1 may need to be transformed


# Which technique determines if var x1 leads var y? Assuming var x1 may need to be transformed












Suppose I want to predict future changes in variable y (stock price over time). I notice that variable x1, inverted and delayed three months, tends to lead y. Which technique can I use to find other predictive relationships like this? Suppose I have variables x2 through x100 that I want to analyze to see if they also lead y in some way.

What are the non-machine learning (ie traditional statistical) options? Do the machine learning ones work better?

Context: want to predict future changes in S&P 500 price based on other assets like gold, oil, currencies, etc.

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.