What I continue to find in my data explorations is that many market indicators are highly correlated with price change, but that this is because of high correlations at either end of the distribution. In other words, when price change is either very strong or very weak, the indicators also tend to be extreme, so that they don't provide additional information. When, however, price change is small (up or down only a little), the correlations with indicators is quite a bit less and the indicators have more unique predictive value. What this suggests to me is a strategy of modeling large price-change days/weeks with price based predictors and modeling small price-change days/weeks with alternate predictors (sentiment, breadth, etc.). I will be trying this strategy out when I return on Tuesday.
On the heels of my last analysis, I decided to look at strong versus weak closes in daily SPY data going back to January, 2003 (N = 744). Over the entire sample, the correlation between where the market closes within its daily range and its price change for the day is a very high .77. Among days when price change was small, however, (up or down less than .20%; N = 170), the correlation between where the market closes within its range and price change is only .35. Performing a median split on the small price-change days, I found that SPY averaged a next-day gain of .18% (55 up, 30 down) when it closed near the lows for the day, but only a gain of .04% (45 up, 40 down) when it closed near its daily highs.
This modest reversal effect dissipates after one day. It will be interesting to track in other markets, including individual stocks. At this juncture, however, the basic strategy of modeling strong up days, neutral days, and strong down days separately is what is most important. Those who assert that technical indicators are valuable and those who assert they are worthless may both be right: they seem to be helpful within particular ranges of price-change distributions.