How to predict a company's revenues without knowing anything about its business (Part 2 of 2)

In yesterday's post, I suggested one mechanical way of estimating the next quarter's revenues of a firm, based on the straight-line extrapolation of its four-quarter moving average. Today, I'll take it one stage further, investigating whether looking at second-order extrapolation buys us much. And then I'll take a different approach by leaving it entirely to the computer to find the set of coefficients which create the best fit for a linear formula for next quarter's revenue in terms of the previous six quarters.

Second-order Predictor

Here we assume that the rate of change represented by the next quarter's moving average is the straight line extension of the rate of change from the two previous quarters:



The definition of D_n is the gradient from A_n-1 to A_n:


The assumption of straight-line growth of the rate of change of A_n gives us:


Rearranging the equation:


So, by substitution:


And by further substitution:

Assessing the Accuracy of the Predictor

When tested out on the same set of vnedor financial data as before (i.e. from 2003 to 2007), the average error generated by this formula improved a little on the previous formula: 6.6%, across 616 readings.

And creating a new predictor which is the simple average of our two previous predictors merely results in an error rate which is the average of the other two: 6.7%.

Using Excel Solver to derive the optimised linear formula

We can use the Solver add-on within Excel to calculate the linear coefficients which minimise the average error.

This has enabled us to get the average error down from 6.8% (first-order approximation) to 6.6% (second-order) and now down to 4.1% (''Solver''-optimised).

Potential for Improvement

One obvious improvement that could be made, particularly in this era of IT industry consolidation, is to allow for acquisitions. Our current model does not use any information about, for example, the size of the companies that Oracle is acquiring, to generate the forecast on a pure time series basis.

But when one firm acquires another, it makes sense for the analyst to add the revenues of the two firms together. (No company destroys that much value within a quarter!) Our revised time-series model should do the same.