What Quant Models Can Teach Us About Trading Psychology
Above we can see SPY (blue line) plotted against a six-variable trading model that I developed using ensemble modeling. When we have a positive score, the model is deemed to be bullish over a next 10-day horizon. When we have a negative score, the model is deemed to be bearish. The model is flat as of yesterday's close. The model includes such variables as market volatility, breadth, buying/selling participation, and market cycle status.
When the model has been at a score of +2 or higher, the next 10 days in SPY have averaged a gain of +2.08%. When the model has been at a score of -3 or lower, the next 10 days in SPY have averaged a loss of -.83%. Between scores of +1 and -2, the next 10 days in SPY have averaged a small loss of -.08%.
The model has a couple of important implications for trading psychology:
1) Out of the 573 days of my in-sample and out-of-sample periods, nearly half are scores less than +2 and greater than -3: in other words, days with essentially no edge 10 days out. That doesn't mean sources of edge can't be found on different time frames with different models, but this finding is important. Even with a solidly researched source of edge, there are plenty of occasions when not trading is the best trade.
The model signals have been good, but even with their edge, there is plenty of noise. Note, for example, that the model was bearish during much of mid-2015, when prices chopped around quite a bit. We were also bullish during fall, 2015 during a volatile bottoming period. A trader could have an edge with a model but be unable to survive the noise around signal, especially if sized quite large.
The act of developing models itself gives one a feel for markets. The model inputs are there for a reason: the model simply captures when those reasons line up. It is interesting that the most rational of analyses can feed the deepest intuitions.