I have a specific but interesting problem at hand:

I know that the price of a stock falls after a negative shock (in t-1). Now it is the day after the shock (t) and I want to know whether the price will likely rebound to its former level or stay at the lower level in the next x periods (t+1...t+x).

This event has happened in several different companies at several times in the past -> These are my Inputs (the event date is known).

This is not a problem of estimating the impact of the shock itself, it is about the effect after the impact on the price.

Alternatively this may be thought of as a problem of estimating the persistance/absorption of the time series or as the lasting effect of a shock or even as the an estimation of the parameters of a mean-reverting process.

Thoughts on what kind/class/category of problem I have here would be much appreciated. Because I really would like to be able to at least articulate my problem


You can detect this kind of bounce with a moving average (MA) model. A moving average model is based on predicting $Y_{t}$ by using the past errors in estimation:


Where $\epsilon_{t-1}$ was the error in estimating $Y_{t-1}$. If there is a shock in period t-1 and that causes an unexpected fall in the value of a companies stock then you'll have a positive error $\epsilon$ in time period t-1. If you have rebound (or otherwise called a dead cat bounce) in period t you'd expect $\theta_{1}$ to have a positive coefficient, i.e. overestimating $Y_{t-1}$ would lead to a correction in estimating $Y_{t}$.

A moving average model of order one, MA(1), looks at just the error in the previous period. A moving average model of order MA(q) would look at the errors in predictions of the last q periods.

For more details have a look at ARIMA modelling (the MA standing for moving average), here's a good place to start:https://www.otexts.org/fpp/8

Technical Disclaimer: the above formulae is actually more than just a moving average as it in effect includes a first order difference term, $Y_{t-1}+\epsilon_{t}$, assuming your data is stationary after being differenced. The following model would be actually be considered a ARIMA(0,1,1) model:


I only mention this as knowing when to difference your time-series is a really important part of ARIMA analysis.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.