Each month starting in August, I will post 6-month performance forecasts for several dozen Exchange Traded Funds (ETFs). The forecasts will be on a separate page at SixMonthStockMarketForecast.com, not part of the monthly blog email. This will be a true test in public view. Do not use any Alpha Test forecasts for any investing decision. They are still unproven.

Simultaneously, a much broader test begins, documenting performance results for approximately 650 market indexes, ETFs, and major common stocks including nearly all stocks in the S&P 500 and the NASDAQ 100. These forecasts are proprietary and will not be published here.

Actual Forecasts to Date My first stock market forecast in September 2007 used economic and other data going back to 1984 to guess what would happen to the Value Line Arithmetic Average (VALUA) six months later in early 2008. The expected accuracy was an R.square score of 0.64. (R.square=1.0 is perfect) In nearly 200 forecasts since then, the actual average score achieved is about 0.54, fairly close. The graph below shows how those forecasts lined up with reality. Not too bad considering that Technical Analysis methods that many gurus peddle have an abysmal tested R.square of 0.0.

Back Test Results The back-test of my expanded forecasting set has wrapped up. It followed the same approach as my original forecasts: use data from as far back as 1984 to make estimates “as if” the actual results were unknown. That testing led to over 100,000 “as-if” forecasts. For 2/3 of the stocks tested the approach was largely successful with R.square ranks from about 0.3 to 0.9. For the other third of the stocks tested the modeling approach was essentially a failure. Large company stocks that mirrored the overall economy did best. Young highly volatile technology companies without much history were least predictable. Many stock estimates grew better over time as more and more data became available.

Expected R.Square Matters The graph below shows the results of the back-test of 100,000 “as-if” forecasts, plotting forecasted gain/loss versus what actually occurred. Points are color-coded based on the expected R.square – given existing data, how accurate the resulting forecasts were most likely to be.

Points with an expected R-square of around 0.3 landed like a shotgun splatter: they are generally in the right direction, but they are not accurate. Points with higher and higher expected R.square get increasingly near to the desired one-to-one match between forecast and actual.

The only real test of a forecasting methodology is what happens in real-time in the future with new conditions. But, the back-test was pretty encouraging.