No. of Recommendations: 24
If you'll share the methodology and datasets, I can make this do exactly what you do.
I described it in moderate detail before, though not a couple of probably meaningless tweaks that make me feel important and mysterious.
Broad strokes for the major bottom detector:
It's based mostly on new highs and new lows, each expressed as a percentage of total issues, looking for an extreme in both of those at the same time. Low #highs and high #lows. Nothing surprising. Those numbers are smoothed a bit.
Those two numbers are combined with a relative weighting into a single metric, as they aren't equally important. That combo metric is ranked among a lookback history of itself, and the percentile tested against a cutoff. Different levels of cutoff give different certainties of a bottom, and have different average forward returns, which is why I sometimes mention whether the signal was weak or strong on a given day. I do have a single "canonical" cutoff though.
My short term bottom and major bottom models are constructed differently, not just two different cutoffs of the same metric.
The main flaw is that it will give a whole bunch of signals all the way down a "waterfall" market crash like late 2008. Despite being bad signals, one year forward returns are still fine from the earlier signals, which is what it's tuned to accomplish, but the first month or two can be (ahem) arbitrarily low, just as a market bottom can be arbitrarily low. Most of the time you only get a short cluster or two, so it's not a big deal, but it gets confused in a big lasting meltdown and signals too early.
Jim
