Objective research and reviews to aid investing decisions
 Looking for information on debt consolidation loans? Want to know more about reverse mortgages? |
| Find a mortgage company online |
About nine months ago, reader Wes Williams of Stafford TX inquired:
"The Financial Forecast Center (FFC) provides an S&P 500 forecast that disagrees with yours, projecting a much lower level for the fall. There is no accountability at their site for their past predictions. Does it fit into your web site's charter to track THEIR forecasts?"
Our charter is flexible. FFC's "forecasts are generated in-house using artificial intelligence. The forecast models are 100% quantitative and use a global, long-range economic dataset. Thus, the forecasts are very objective." FFC forecasts the average value by month for the S&P 500 index for the current month and the next five months, including fairly large error ranges for 50% and 90% confidence levels. Based on nine of these forecasts since May 2006, we conclude that...
Past reviews of artificial intelligence/neural net stock market models suggest approaches that tend to be heavy on math and logic and light on economics and investing (see our blog entry of 3/24/06). The models are generally empirical, with little or no theoretical discussion or explanations. Brute force calculation presumably overcomes any lack of nuance in indicator selection and interaction. The only way to assess such models is to see whether the outputs are right or wrong.
The following chart shows nine recent six-month FFC forecasts (by publication date) for the monthly average S&P 500 index, along with the actual average value of the S&P 500 index for 5/06-1/07. We omit forecasts for 9/24 and 11/12 because FFC issued two similar forecasts in those months. When Wes sent his request, the 5/15/06 forecast was current; hence, his comment about the low level predicted for the fall. While the sample is far too small to draw statistically reliable conclusions about the accuracy of the FFC approach, we note that:
This chart is similar to that in which we depict our own inflation forecast chasing actuals during 2000-2006 (see our blog entry of 7/10/06). Because of that similarity, we assess also the FFC Consumer Price Index (CPI) forecast.
The next chart shows nine recent six-month FFC forecasts (again by publication date) for the monthly CPI, along with the actual CPI for 5/06-12/06. We omit the forecast for 9/24 because FFC issued two identical forecasts in that month. The chart shows that reality has substantially surprised the FFC model to the downside in recent months. Again the sample is far too small to draw firm statistical conclusions about the accuracy of the model.
We can, however, check whether our own Inflation Forecast (which projects inflation based on recent actuals and requires just a little intelligence) has chased actuals more effectively than has the FFC model over this relatively short period.
The final chart compares the average absolute errors for the FFC CPI forecast and the CPI forecast embedded in our (CXO) inflation forecast during 5/06-12/06 for forecast horizons of one to six months. The sample sizes for all horizons are far too small to determine reliably whether one model is better than another, but initial data does not indicate that the more complex FFC model is better.
In summary, while the nine-month sample assembled does not support statistically reliable conclusions, the Financial Forecast Center's approaches to forecasting the S&P 500 index and the CPI do not look promising.
We are therefore not going to continue monitoring these forecasts.
For related research, see Blog Synthesis: Some Trading Indicators. See also our blog entry of 7/19/06 on genetic programming (a form of artificial intelligence), with those involved noting that its application "in a stock selection context is as much an art as a science."
