Using ARIMA to model and forecast stock prices using user-friendly stats program

Question

Hi people,

Can anyone please offer some insight into this for me?

I'm coming from a functional magnetic resonance imaging research background where I analyzed a lot of time series data, and I'd like to analyze the time series of stock prices (or returns) by: 1) modeling a successful stock in a particular market sector and then cross-correlating the time series of this historically successful stock with that of other newer stocks to look for significant relationships; 2) model a stock's price time series and use forecasting (e.g., exponential smoothing) to predict future values of it. I'd like to use non-linear modeling methods (ARIMA and ARCH) to do this.

Several questions: How often do ARIMA and ARCH modeling methods (given that the individual who implements them does so accurately) actually fit the stock time series data they target, and what is the optimal fit I can expect? Is the extent to which this model fits the data commensurate with the extent to which it predicts this stock time series' future values?

Rather than randomly selecting stocks to compare or model, if profit is my goal, what is an efficient approach, if any, to selecting the stocks I'm going to analyze?

Which stats program is the most user-friendly for this?

Any thoughts on this would be great and would go a long way for me.

Thanks, Brian

Answer 1

if you want free then R is where you need to start for a stats package.

Most of the economists I know either use Stata or Matlab for their modeling. Given that you wanted something with a smaller learning curve I would reccomend Stata.

Answer 2

Expectably, almost all forecasting (AI powered or not) techniques have been utilized to attempt an “efficient” stock market prediction. Results posted are not that encouraging, how could they after all; if someone comes up with the holy grail model, he’s not expected to publish it, but use it to make some money, right? In general, a fit gets “optimal” if the trades it suggest have an accuracy bigger than 50% + transaction costs, so that it is expected to return profits to anyone using it. If this sounds feasible, don’t get disappointed with your results; the sure thing is that you’ll get to learn a lot during the process, plus much fun is guaranteed :)

---

with the generous contribution of the MineKnowledge team

Answer 3

If you want powerful statistics, go for R. It has a non-trivial learning curve, but you will reap on the benefits of investing some time in it. R is free, and that is a huge asset.

As for the approach, wouldn't you agree that the stock price already includes all the information which can be predicted? What is keeping the markets from incorporating this information? Also, wouldn't it be like comparing apples and oranges to use one stock's past behaviour as a predictor of another stock's future behaviour?

Just a tought...

Answer 4

Thank you guys for your useful thoughts and recommendations!

It looks like R and STATA are a good stats programs to use starting off.

gd047, good point--I was thinking exactly that just after posting...It makes sense that Econometrics (unlike other domains of research) must be an area of research in which the content of current peer reviewed journal papers is unlikely to reflect the most innovative/successful techniques out there, especially for topics where the goal of using these techniques is to make money. I at least take some comfort in this since the absense of reported successful techniques does not necessarily reflect that they haven't been been figured out yet and then that a person can't figure one out using proper statistical techiniques (though I also realize that it doesn't mean that they have been figured out or that one can figure them out either). In any case, as you stated, what it will be a fun process!

Arrieta, Yes, I think that a stock's price already includes all the information that can be predicted, and modeling it to forecast future values of it is useful--But what that information is that is influencing the price values is the challenging part. But, I also think looking for unlikely relationships between a historically successful stock and a newer stock (based on a hypothesis) may have predictive value, since a significant finding may highligt additional, previously unkown, factors that influence a stock's success. Yes, I'm also trying to figure out how to characterize a stock's behavior as distinguished from other stocks to determine which ones are comparable and which ones aren't--I imagine reading more about this topic and with more experience I'll get a better idea about categories of stocks, specifically with respect to their performance/behavior...

Brian

Answer 5

"an accuracy bigger than 50% + transaction costs"

Thanks for putting it into perspective---now I have a goal. In your experience, is it the exception or the rule to achieve an accuracey of more than 50% + transaction costs with a model? Best case scenario, how often can I expect to do this assuming I employ available techniques accurately? Given that successful models are seldom published, how good are the available methods (ARIMA, ARCHE, etc.) at predicting values? With respect to behavior of the values, do time series data of stocks from one market sector behave more similarly with each other than with the stocks of other market sectors, or are there better ways of grouping them?

Thanks.

Answer 6

I second the motion on STATA as being easy to learn. I've done some work in the area of vector autoregressive time series models. This models the interaction between two time series and is a fairly recent (1980s) approach to multiple time series - especially those with lagged interactions. STATA does handle this kind of analysis.