Choosing Nested Regressions to Study International Stock-Index Links

Summary

The document considers how to model relationships among international stock-index returns using a general-to-specific sequence of regressions. It starts with a lagged return from one market as the sole predictor, adds the dependent market’s own lag to form a pairwise Granger-causality specification, and then includes multiple candidate markets together.

The central question is whether the simplest regression has a useful role when the later models better account for own-market dynamics and other predictors. The document provides no empirical results or answer, and it leaves the model-selection rationale unresolved. Its setup is useful for distinguishing a simple lagged association from predictive tests that control for the target series’ history and from a broader multivariate specification; conclusions would depend on the research objective and appropriate statistical assumptions.

Key ideas

• A single lagged market return tests a basic predictive association with another market’s return.
• Adding the dependent market’s own lag produces a pairwise Granger-causality specification.
• Including several lagged markets jointly examines their conditional relationships with the target market.
• The document raises but does not resolve when the simplest regression is justified.

Tags

# General to specific approach to modelling


# General to specific approach to modelling












I am trying to find the relationship of stock indices across the world. This has been done by the literature, however, I am wondering about the methods chosen.

I have decided to go with what I think is a general to specific approach.

I start with the most basic regression model,

$$Y_t = \alpha +\lambda X_{t-1}+\epsilon_t$$ (1)

I then move to another regression model, the pairwise granger causality test,

$$Y_t = \alpha +\lambda X_{t-1}+\theta Y_{t-1}+\epsilon_t$$ (2)

Finally, I include all of the potential $X's$ in the one regression to consider the links between the variables,

$$Y_t = \alpha +\lambda_1 X_{1,t-1}+...+\lambda_n X_{n,t-1}+\theta Y_{t-1}+\epsilon_t$$ (3)

I can justify why I use equations (2) and (3), but I am having a little bit more difficulty of why I would use equation 1.

My idea was that I should start with the most general framework and see if the relationships hold.

I am not sure if I should just exclude equation (1), is their any ever justification for starting at the most simple regression model as above?

$Y$ and $X$ are returns of different stock markets, i.e. the S&P 500 and FTSE 100.

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.