Simple Multivariate Model

[arkhan]

Hi, I am new to quants and want to make a simple multivariate regression model to predict stock market return based on trailing -> volatility (Std.Dev), valuation (P/E multiple), technical (30DMA).

What i plan to do is use monthly observations 5yrs (60 data points) and run regression to obtain factor exposures (Beta i).

Stock Reutrns (i) -> F ( stdev , P/E, M.avg) in period (i-1)

Reutrn stock market (in period i) = B1 x Std dev (in Period i-1) + B2 x P/E (in Period i-1) + B3 x 30 DMA (in Period i-1).

Hence using the trailing values for the independent varaibles i would be able to predict next months stock return.

Please let me know is this simple multi-factor regression model empirically sound and kindly identify problems in this...or should i make some modifications.

Thanks a lot

ARK

 

[Darou]

It is obvious that you will be able to fit this simple regression model. It depends, how you want to use the results. For real trading, this will not work. You have to do some back-testing and you will observe the following:

In sample, these kind of models usually work quite fine. But, with new data (out-of-sample), these kind of approaches do not work. Even if it works somehow, transaction costs will kill you.

Note that for for back-testing, you should divide your available data into at least three sets:
1) One for learning. This in-sample data set (about 1/3 of your total data) will determine your factors B1 to B3.
2) One for testing. This out-of-sample data (another 1/3 of your data) will show if your strategy is successful.

If 2) is successful enough use
3) For verification. This out-of-sample data (last 1/3 of your data) will show if you can really try to go live.

If 3) is still very good: Go to real live data - but, do not trade immediately. Test with real time data and real transaction cost. If it still works: Great, you will be rich!

If 3) is not as good as 2): Stop. Your Strategy does not work and similar will not work either.

Good luck!

Darou