State-Space Models for Short-Horizon Equity Return Forecasting

Summary

This document raises the question of whether state-space models can improve forecasts of equity returns from five-minute price data. The author describes working with several years of observations across multiple equities, comparing linear regression with an AR(2) model estimated by maximum likelihood and a random-walk-plus-noise model. Parameters are re-evaluated as new data arrives, but the reported in-sample model bias is generally small, leaving uncertainty about whether these models offer useful forecasting gains.

The discussion also asks whether holding a fitted model for a period and forecasting multiple steps ahead might improve robustness. It provides no empirical comparison, out-of-sample results, or answer to that question, so it should be read as a research problem rather than evidence that state-space methods outperform simpler models. Any assessment would need clear forecast horizons and genuinely out-of-sample evaluation, especially for noisy intraday returns.

Key ideas

• The author explores state-space methods for forecasting equity returns from five-minute price observations.
• The compared approaches include an AR(2) model and a random-walk-plus-noise model.
• Model parameters are re-estimated as new observations arrive, but the reported in-sample bias is limited.
• The document does not report out-of-sample evidence that state-space models improve forecasts.

Tags

# Forecasting Equity returns using state-space models


# Forecasting Equity returns using state-space models












I have data for 3 yrs of 5 min prices of various equities. I can construct a linear regression model and try to see the in-sample model performance. But I was wondering whether fitting a state-space model might improve my model bias (performance). I have been using R and constructed an AR(2) process (where the coefficients are estimated by MLE) using the package DLM and a Random-walk-plus-noise model but both of these do not give significant bias (model bias is typically < 50 %). Also I am rolling the model as each new data comes in and re-evaluating the model parameters at each step. I do not know if holding the model for some time and forecasting multi-step ahead is gonna make my model more robust. I tried to search online but among the various papers I got, nothing pinpoints to using a state space model for forecasting equity returns. Does someone have any knowledge of this or can point me to a paper which deals with this? Thanks in Advance !!

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.