Newsletter: Crossing the Threshold - Classifying the Market State
Turning time series into classifiers using ROC analysis.
People have a natural tendency to categorize things and ideas. Often there is a tendency to create these classifications in terms of opposites. They can be as simple as "good" or "bad." Sometimes it's "high" or "low." Sometimes we do this with a list of information and then add them up to make a decision.
You can decide, for instance, whether this is a good day to play golf by taking a number of criteria you think are important for a good game. How similar is today to your ideal day for playing golf? How's the weather? Are the links crowded? How much time do you have? How's your back feeling? By classifying these criteria as "good or bad" and adding them up based on their importance, you can make a decision to play or not.
Generally dividing criteria up in a binary way helps us with classification, but what if the concepts are complex? Can we still use a binary view to make decisions?
In the Market Climatology model I use binary factors based upon fuzzy set ideas of similarity to classify the current market risk state. But market and economic data are generally continuous time series. Turning continuous data into binary factors requires a cutoff level, or threshold in order to be considered in one state or another. Most of us do this already in market analysis though the threshold can be somewhat arbitrary.
For instance, the MARKIT Purchasing Managers Index (PMI) is interpreted as meaning growth when it is above 50 and a recession when it's below 50. Why? Because the PMI is a survey of purchasing managers about current conditions. If the number is above 50, it means that more than 50% of the respondents thought conditions were improving. If it's below 50, then more than 50% think conditions are getting worse. It's basically a consensus measure. So, for most people, 50 is the default threshold for classifying the current state of manufacturing as growing or contracting. But is 50 really optimal? What empirical evidence tells us that 50 is, indeed the threshold for classifying recessions and recoveries?
There is a methodology for determining the optimal threshold called ROC. It's little known in economics and investments though it is used in areas like medicine and AI to determine thresholds for classification models. I use it in setting many of my thresholds and I referred to it briefly in the hypertext book. For subscribers of this newsletter, I'm going to show you how using the PMI.