Experimental Stock Pricing Model
Premiums and Discounts. The model fitting process makes average premium or discount very close to 0%. An important characteristic of the model is the distribution of these premiums or discounts (technically the model views these as errors or components of price it can't explain with the fundamental variables.) By examining the percentile of a company's premium or discount, one can find out how many stocks in the database had comparable, or higher, levels of premium.
Statisticians Note. Using the parameters from the September 30, 1996 data, the model has about a 63.5% R-Square (in continuously compounded premium/discount space). This month, I reduced the number of input variables and changed the architecture slightly to compensate. The number of weights in the network has been reduced by about 8% (about 90 total). Why the change? I was using a somewhat awkward mechanism for including 5 year sales growth in the model and compensating for those companies missing that data. Now, I replace the missing 5 year sales growth number with the ValueLine industry average for the company (only for stocks in industries with more than 5 companies). The new model has comparable performance with a more parsimonious specification.
As a bonus this month, I ran Michael Murphy's TWIT$ portfolio from the November issue of Wired. I was able to price all but two of the stocks (Euphonix and Diamond Multimedia). Overall, as of September 30, 1996 the weighted average premium (shares times 9/30/96 prices) for this portfolio was a modest 10.57%. Here are the predicted prices and premiums
Ticker: NightScope Price; Premium/Discount as of 9/30/1996