Choosing Multivariate Models for Long-Horizon Index Return Forecasts

Summary

The document describes a forecasting problem involving several public equity indices observed quarterly, with a requested horizon of at least fifteen to twenty samples. It reports that a VAR or VARMA specification fitted the historical data well but performed poorly out of sample, suggesting that the chosen models may be overfit. The author asks for alternatives rather than presenting a tested replacement.

Candidate approaches raised include copula-based simulation, Markov-chain Monte Carlo resampling, and recurrent neural networks. These are proposals, not demonstrated recommendations: the document gives no dataset, forecast evaluation, benchmark comparison, or evidence that any candidate improves performance. Its practical lesson is mainly the need to take out-of-sample behavior seriously when fitting multivariate return models. Because the series are quarterly and the horizon is long relative to typical financial samples, forecast uncertainty and model complexity are material concerns; the source does not discuss how to address them or provide a validated forecasting procedure.

Key ideas

• The task is to forecast multiple public index return series at a long quarterly horizon.
• The author reports strong in-sample but weak out-of-sample results from VAR and VARMA models.
• Copula simulation, Markov-chain Monte Carlo resampling, and recurrent networks are suggested as possibilities, not validated solutions.
• The document provides no empirical comparison or evidence that an alternative method improves forecasts.

Tags

# Suggestion on the models to estimate public indeces future returns


# Suggestion on the models to estimate public indeces future returns












I would like to to estimate the future returns of some public indeces. I have several of them so it is a multivariate problem. The series are quarterly and the estimation should be of at least 15-20 samples ahead.

I have tried so far with a VAR-VARMA model, but although i see great performance in-sample it is very bad out-of samples, showing it is prone to overfitting.

Now I know this is a 1 million dollar question, but in reality I would just like to have an advice on other methods I could try.

I was thinking at copula + MC resampling, or copula + markov chain MC resampling, or maybe adopting recurrent neural networks or others.

Can anybody suggest me on this?

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.